CN115859095A

CN115859095A - Intention recognition method based on intention recognition model, training method and training equipment

Info

Publication number: CN115859095A
Application number: CN202211371788.8A
Authority: CN
Inventors: 李勇; 李银峰; 高宸; 杜小毅; 韦华周; 罗恒亮; 金德鹏
Original assignee: Tsinghua University; Beijing Sankuai Online Technology Co Ltd
Current assignee: Tsinghua University; Beijing Sankuai Online Technology Co Ltd
Priority date: 2022-11-03
Filing date: 2022-11-03
Publication date: 2023-03-28

Abstract

The application provides an intention recognition method, an intention training method and intention recognition equipment based on an intention recognition model. The method comprises the following steps: acquiring a set to be identified; the to-be-identified set comprises a plurality of to-be-identified consumption behavior data; generating a plurality of dual hypergraph information according to the set to be identified, wherein the dual hypergraph information comprises a first hypergraph and a second hypergraph corresponding to each feature information; inputting the dual hypergraph information into an encoder of an intention recognition model, and generating intention characteristics of consumption behavior data to be recognized; and inputting the intention characteristics into a decoder of an intention recognition model to obtain intention information of the consumption behavior data to be recognized. The consumption behavior data can be fully modeled, and unknown intention information in the consumption behavior data is found based on a small amount of marked intention information, so that the accuracy and interpretability of e-commerce recommendation are improved, and the user experience is improved.

Description

Intention recognition method based on intention recognition model, training method and training equipment

Technical Field

The present application relates to big data technology and electronic commerce technology, and in particular, to an intention recognition method, an intention training method, and an intention recognition device based on an intention recognition model.

Background

With the development of internet technology, users can complete the purchase of goods based on the internet. And may recommend items to the user based on the internet. In order to facilitate recommendation of an item to a user, it is necessary to determine user intention information. The intention information characterizes the use and reason of the user after purchasing the item; for example, when a user purchases an item and applies the item to a family party, the intention information at this time is "family party".

In the prior art, when identifying the intention information, the corpus information of the user can be identified based on a clustering manner, and then the intention information is obtained.

However, in the above manner, a large number of articles are provided in the e-market scene, the intention information obtained for the material information according to the existing clustering algorithm cannot be adapted to the scene, and the obtained intention information is inaccurate.

Disclosure of Invention

The application provides an intention recognition method, an intention training method and intention recognition equipment based on an intention recognition model, which are used for solving the problems that no intention recognition method suitable for E-commerce scenes exists and intention recognition is inaccurate.

In a first aspect, the present application provides an intention recognition method based on an intention recognition model, the method including:

acquiring a set to be identified; the set to be identified comprises a plurality of consumption behavior data to be identified, the consumption behavior data to be identified comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information;

generating a plurality of dual hypergraph information according to the set to be identified; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user;

inputting the plurality of dual hypergraph information into an encoder of an intention recognition model, and generating intention characteristics of the consumption behavior data to be recognized; wherein the intention characteristics characterize sub-characteristics of consumption behavior data to be identified on each characteristic information;

inputting the intention characteristics into a decoder of the intention recognition model to obtain intention information of the consumption behavior data to be recognized.

In one possible implementation, inputting the plurality of dual hypergraph information into an encoder of an intention recognition model, generating intention features of the consumption behavior data to be recognized includes:

inputting the information of the plurality of dual hypergraphs into the encoder to generate convolution characteristics corresponding to each hypergraph convolution layer in the encoder;

aggregating each convolution characteristic to obtain an aggregated characteristic; and determining the intention characteristics of the consumption behavior data to be identified according to the aggregation characteristics.

In one possible embodiment, the first hypergraph has a first correlation matrix characteristic, a first node degree matrix characteristic, a first hyper-edge matrix characteristic, and a first node characteristic;

the first incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the first hypergraph belongs; the first node degree matrix characteristic represents the connection relation of user nodes in the first hypergraph; the first hyper-edge matrix characteristic represents a node corresponding relation of edges in a first hyper-graph; the first node feature characterizes node information of a user node in a first hypergraph;

the second hypergraph has a second incidence matrix characteristic, a second node degree matrix characteristic, a second hyper-edge degree matrix characteristic and a second node characteristic; the second incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the second hypergraph belongs; the second node degree matrix characteristic represents the connection relation of characteristic information nodes in a second hypergraph; the second hyperedge matrix characteristic represents the node corresponding relation of edges in a second hypergraph; the second node features characterize node information of the feature information nodes in the second hypergraph.

In one possible embodiment, inputting the plurality of dual hypergraph information into the encoder, generating convolution features corresponding to each hypergraph convolution layer in the encoder, includes:

inputting the information of the plurality of dual hypergraphs into the encoder, and generating a first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and partial feature of the second hypergraph;

generating a second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph;

and processing based on each hypergraph convolutional layer according to the first initial embedding characteristic and the second initial embedding characteristic corresponding to the characteristic information to obtain a convolutional characteristic corresponding to each hypergraph convolutional layer.

In one possible implementation, generating a first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and the partial feature of the second hypergraph includes:

determining the propagation characteristics in the first graph corresponding to the characteristic information according to the first incidence matrix characteristics, the first node degree matrix characteristics, the first super-edge degree matrix characteristics and the first node characteristics of the first hypergraph corresponding to the characteristic information; wherein the propagation characteristics in the first graph represent characteristic expressions of users on characteristic information;

determining the first inter-graph propagation characteristics corresponding to the characteristic information according to the second incidence matrix characteristics, the second super-edge matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; wherein the first inter-graph propagation feature characterizes a feature expression between a user and feature information;

and determining a first initial embedding feature corresponding to the feature information according to the first intra-map propagation feature corresponding to the feature information and the first inter-map propagation feature corresponding to the feature information.

In one possible implementation, generating a second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph includes:

determining the propagation characteristics in the second graph corresponding to the characteristic information according to the second incidence matrix characteristics, the second node degree matrix characteristics, the second super-edge degree matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; wherein the feature representation of the feature information on the user is propagated in the second graph;

determining a second inter-graph propagation characteristic corresponding to the characteristic information according to the first incidence matrix characteristic, the first hyper-edge matrix characteristic and the first node characteristic of the first hyper-graph corresponding to the characteristic information; wherein the second inter-graph propagation feature characterizes a feature expression between the user and the feature information;

and determining a second initial embedding feature corresponding to the feature information according to the second intra-map propagation feature corresponding to the feature information and the second inter-map propagation feature corresponding to the feature information.

In one possible embodiment, determining the intention characteristics of the consumption behavior data to be identified according to the aggregation characteristics includes:

and processing the aggregation characteristics based on a multilayer perception mode to generate the intention characteristics of the consumption behavior data to be identified.

In one possible embodiment, the set to be identified includes a first set and a second set; the consumption behavior data to be identified in the first set has actual intention information, and the consumption behavior data to be identified in the second set does not have actual intention information.

In one possible embodiment, the set to be identified includes a first set and a second set; the consumption behavior data to be identified in the first set have actual intention information, and the consumption behavior data to be identified in the second set do not have actual intention information; inputting the intention characteristics into a decoder of the intention recognition model to obtain intention information of the consumption behavior data to be recognized, wherein the intention information comprises:

clustering the intention characteristics of the consumption behavior data to be identified to obtain a plurality of intention information;

and decoding the intention characteristics corresponding to the consumption behavior data to be recognized according to the decoder of the intention recognition model so as to classify the consumption behavior data to be recognized into the plurality of pieces of intention information to obtain the intention information of the consumption behavior data to be recognized.

In a possible embodiment, according to a decoder of the intention recognition model, performing decoding processing on intention features corresponding to the consumption behavior data to be recognized, so as to classify the consumption behavior data to be recognized into the obtained plurality of intention information, and obtain intention information of the consumption behavior data to be recognized, includes:

according to the decoder of the intention identification model, decoding intention characteristics corresponding to the consumption behavior data to be identified so as to classify the consumption behavior data to be identified into the plurality of intention information and obtain probability distribution information of the consumption behavior data to be identified on each characteristic information; wherein the probability distribution information represents the probability distribution condition that the consumption behavior data to be identified belongs to the intention information in the intention information;

determining the weight information of the consumption behavior data to be identified on each characteristic information according to the sub-characteristics of the consumption behavior data to be identified on each characteristic information;

determining an initial intention of the consumption behavior data to be identified on each characteristic information according to the probability distribution information and the weight information of the consumption behavior data to be identified on each characteristic information;

and determining intention information of the consumption behavior data to be identified according to each initial intention corresponding to the consumption behavior data to be identified.

In one possible embodiment, the method further includes:

and recommending articles to the user according to the intention information of the consumption behavior data to be identified.

In a second aspect, the present application provides a model training method applied to intent recognition, the method comprising:

acquiring a set to be trained; the to-be-trained set comprises a plurality of to-be-trained consumption behavior data, the to-be-trained consumption behavior data comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information;

generating a plurality of dual hypergraph information according to the set to be trained; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user;

inputting the information of the dual hypergraph into an encoder of an initial model for training, and generating the intention characteristics of the consumption behavior data to be trained; the intention features represent sub-features of the consumption behavior data to be trained on each feature information; and updating the encoder according to the intention characteristics;

inputting the intention characteristics into a decoder of the initial model for training so as to update the decoder to obtain an intention recognition model; the intention recognition model is used for recognizing consumption behavior data to be recognized to obtain intention information. Inputting the dual hypergraph information into an encoder of an initial model for training, and generating the intention characteristics of the consumption behavior data to be trained, wherein the intention characteristics comprise:

aggregating the convolution characteristics to obtain aggregate characteristics; and determining the intention characteristics of the consumption behavior data to be trained according to the aggregation characteristics.

the first incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the first hypergraph belongs; the first node degree matrix characteristic represents the connection relation of user nodes in the first hypergraph; the first hyper-edge matrix characteristic represents a node corresponding relation of edges in a first hyper-graph; the first node characteristics characterize node information of user nodes in the first hypergraph;

determining the propagation characteristics in the first graph corresponding to the characteristic information according to the first incidence matrix characteristics, the first node degree matrix characteristics, the first super-edge degree matrix characteristics and the first node characteristics of the first hypergraph corresponding to the characteristic information; wherein the propagation features in the first graph characterize feature expression of a user on feature information;

determining the propagation characteristics in the second graph corresponding to the characteristic information according to the second incidence matrix characteristics, the second node degree matrix characteristics, the second super-edge degree matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; wherein the second graph propagates feature expression of feature characterization feature information on a user;

determining second inter-graph propagation characteristics corresponding to the characteristic information according to the first incidence matrix characteristics, the first super-edge matrix characteristics and the first node characteristics of the first hyper-graph corresponding to the characteristic information; wherein the second inter-graph propagation feature characterizes a feature expression between the user and the feature information;

In one possible embodiment, determining the intention characteristics of the consumption behavior data to be trained according to the aggregated characteristics includes:

and processing the aggregation features based on a multi-layer perception mode to generate the intention features of the consumption behavior data to be trained.

In a possible embodiment, the consumption behavior data to be trained has actual behavior information, and the actual behavior information represents whether the user purchases the item in the consumption behavior data; updating the encoder according to the intent characteristics, including:

determining an independent loss function according to each pair of sub-features corresponding to the consumption behavior data to be trained and the total number of the consumption behavior data to be trained in the set to be trained;

according to the sub-characteristics corresponding to the consumption behavior data to be trained, determining the predicted behavior information of the consumption behavior data to be trained, and according to the predicted behavior information and the actual behavior information of the consumption behavior data to be trained, determining a BPR loss function;

and determining an overall loss function according to the BPR loss function, the independent loss function and a preset hyper-parameter, and updating the decoder according to the overall loss function.

In one possible embodiment, the set to be trained includes a first set and a second set; the consumption behavior data to be trained in the first set has actual intention information, and the consumption behavior data to be trained in the second set does not have actual intention information.

In one possible embodiment, the set to be trained includes a first set and a second set; the consumption behavior data to be trained in the first set have actual intention information, and the consumption behavior data to be trained in the second set do not have actual intention information; inputting the intention characteristics into a decoder of the initial model for training so as to update the decoder to obtain an intention recognition model, wherein the intention recognition model comprises the following steps:

clustering the intention characteristics of the consumption behavior data to be trained to obtain a plurality of intention information;

decoding the intention characteristics corresponding to the consumption behavior data to be trained in the first set according to the decoder of the initial model to obtain a cross entropy loss function;

according to the decoder of the initial model, decoding the intention characteristics corresponding to the consumption behavior data to be trained in the second set so as to classify the consumption behavior data to be trained in the second set into the plurality of intention information to obtain a two-classification cross entropy optimization loss function;

and updating the decoder according to the cross entropy loss function and the two-class cross entropy optimization loss function to obtain the intention identification model.

In a possible embodiment, according to the decoder of the initial model, performing decoding processing on the intention features corresponding to the consumption behavior data to be trained in the second set to classify the consumption behavior data to be trained in the second set into the obtained plurality of intention information, so as to obtain a two-class cross-entropy optimization loss function, including:

according to the decoder of the initial model, decoding the intention characteristics corresponding to the consumption behavior data to be trained in the second set so as to classify the consumption behavior data to be trained in the second set into the plurality of acquired intention information; for each seed feature, performing similarity calculation on the sub-features of each consumption behavior data to be trained in the second set to obtain similarity information of each consumption behavior data to be trained in the second set on each sub-feature;

for each seed feature, determining score information of each consumption behavior data to be trained in the second set on each sub-feature, wherein the score information represents a score of the consumption behavior data to be trained belonging to the same intention information on the dimension of the sub-feature;

and determining the two-classification cross entropy optimization loss function according to the similarity information of each to-be-trained consumption behavior data in the second set on each sub-feature and the fraction information of each to-be-trained consumption behavior data in the second set on each sub-feature.

In a third aspect, the present application provides an identification apparatus for intent recognition, the apparatus comprising:

the acquisition unit is used for acquiring a set to be identified; the set to be identified comprises a plurality of consumption behavior data to be identified, the consumption behavior data to be identified comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information;

the processing unit is used for generating a plurality of dual hypergraph information according to the set to be identified; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user;

a first determining unit, configured to input the plurality of dual hypergraph information into an encoder of an intention recognition model, and generate an intention characteristic of the consumption behavior data to be recognized; wherein the intention characteristics characterize sub-characteristics of consumption behavior data to be identified on each characteristic information;

a second determining unit, configured to input the intention characteristics into a decoder of the intention recognition model, so as to obtain intention information of the consumption behavior data to be recognized.

In a possible implementation manner, the first determining unit is specifically configured to:

inputting the information of the plurality of dual hypergraph into the encoder to generate convolution characteristics corresponding to each hypergraph convolution layer in the encoder;

the first incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the first hypergraph belongs; the first node degree matrix characteristic represents the connection relation of user nodes in the first hypergraph; the first hyper-edge matrix feature characterizes a node correspondence of edges in a first hyper-graph; the first node feature characterizes node information of a user node in a first hypergraph;

the second hypergraph has a second incidence matrix characteristic, a second node degree matrix characteristic, a second hyper-edge degree matrix characteristic and a second node characteristic; the second incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the second hypergraph belongs; the second node degree matrix characteristic represents the connection relation of characteristic information nodes in a second hypergraph; the second hyperedge matrix characteristic represents the node corresponding relation of edges in a second hypergraph; the second node characteristics characterize node information of characteristic information nodes in the second hypergraph.

In a possible implementation manner, when the plurality of dual hypergraph information is input into the encoder, and a convolution feature corresponding to each hypergraph convolution layer in the encoder is generated, the first determining unit is specifically configured to:

inputting the plurality of dual hypergraph information into the encoder, and generating a first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and the partial feature of the second hypergraph;

In a possible embodiment, when generating the first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and the partial feature in the second hypergraph, the first determining unit is specifically configured to:

determining the propagation characteristics between the first graphs corresponding to the characteristic information according to the second incidence matrix characteristics, the second super-edge matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; wherein the first inter-graph propagation feature characterizes a feature expression between a user and feature information;

In a possible embodiment, when generating the second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph, the first determining unit is specifically configured to:

determining a second intra-graph propagation characteristic corresponding to the characteristic information according to a second incidence matrix characteristic, a second node degree matrix characteristic, a second super-edge degree matrix characteristic and a second node characteristic of a second hypergraph corresponding to the characteristic information; wherein the second graph propagates feature expression of feature characterization feature information on a user;

In a possible embodiment, the first determining unit, when determining the intention characteristic of the consumption behavior data to be identified according to the aggregated characteristic, is specifically configured to:

and processing the aggregation characteristics based on a multi-layer perception mode to generate the intention characteristics of the consumption behavior data to be recognized.

In a possible embodiment, the set to be identified acquired by the acquiring unit includes a first set and a second set; the consumption behavior data to be identified in the first set has actual intention information, and the consumption behavior data to be identified in the second set does not have actual intention information.

In one possible implementation manner, the set to be identified acquired by the acquisition unit includes a first set and a second set; the consumption behavior data to be identified in the first set has actual intention information, and the consumption behavior data to be identified in the second set does not have actual intention information; when the intention feature is input into the decoder of the intention recognition model and the intention information of the consumption behavior data to be recognized is obtained, the second determination unit is specifically configured to:

In a possible embodiment, when the decoder according to the intention recognition model decodes intention features corresponding to the consumption behavior data to be recognized, so as to classify the consumption behavior data to be recognized into the obtained plurality of intention information, and obtain the intention information of the consumption behavior data to be recognized, the second determining unit is specifically configured to:

according to the decoder of the intention identification model, decoding intention characteristics corresponding to the consumption behavior data to be identified so as to classify the consumption behavior data to be identified into the plurality of intention information and obtain probability distribution information of the consumption behavior data to be identified on each characteristic information; wherein the probability distribution information characterizes the probability distribution condition that the consumption behavior data to be identified belongs to the intention information in the plurality of intention information;

determining weight information of the consumption behavior data to be identified on each feature information according to the sub-features of the consumption behavior data to be identified on each feature information;

In one possible embodiment, the apparatus further includes:

and the application unit is used for recommending articles to the user according to the intention information of the consumption behavior data to be identified.

In a fourth aspect, the present application provides a model training apparatus for intent recognition, the apparatus comprising:

the device comprises an acquisition unit, a training unit and a training unit, wherein the acquisition unit is used for acquiring a set to be trained; the to-be-trained set comprises a plurality of to-be-trained consumption behavior data, the to-be-trained consumption behavior data comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information;

the processing unit is used for generating a plurality of dual hypergraph information according to the set to be trained; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user;

the first updating unit is used for inputting the dual hypergraph information into an encoder of an initial model for training and generating the intention characteristics of the consumption behavior data to be trained; the intention features represent sub-features of the consumption behavior data to be trained on each feature information; and updating the encoder according to the intention characteristics;

the second updating unit is used for inputting the intention characteristics into a decoder of the initial model for training so as to update the decoder to obtain an intention recognition model; the intention recognition model is used for recognizing consumption behavior data to be recognized to obtain intention information.

In a possible implementation manner, the first updating unit is specifically configured to:

In a possible implementation manner, when the plurality of dual hypergraph information is input into the encoder, and a convolution feature corresponding to each hypergraph convolution layer in the encoder is generated, the first updating unit is specifically configured to:

and processing based on each hyper-graph convolution layer according to the first initial embedding characteristic and the second initial embedding characteristic corresponding to the characteristic information to obtain convolution characteristics corresponding to each hyper-graph convolution layer.

In a possible embodiment, when the first updating unit generates the first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and the partial feature in the second hypergraph, the first updating unit is specifically configured to:

In a possible embodiment, when generating the second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph, the first updating unit is specifically configured to:

In a possible embodiment, the first updating unit, when determining the intention characteristic of the consumption behavior data to be trained according to the aggregated characteristic, is specifically configured to:

and processing the aggregation features based on a multilayer perception mode to generate the intention features of the consumption behavior data to be trained.

In a possible implementation manner, the consumption behavior data to be trained acquired by the acquiring unit has actual behavior information, and the actual behavior information represents whether the user purchases an item in the consumption behavior data; when the encoder is updated according to the intention characteristic, the first updating unit is specifically configured to:

according to the sub-features corresponding to the consumption behavior data to be trained, determining the predicted behavior information of the consumption behavior data to be trained, and according to the predicted behavior information and the actual behavior information of the consumption behavior data to be trained, determining a BPR loss function;

In a possible embodiment, the to-be-trained set acquired by the acquiring unit includes a first set and a second set; the consumption behavior data to be trained in the first set has actual intention information, and the consumption behavior data to be trained in the second set does not have actual intention information.

In a possible embodiment, the to-be-trained set acquired by the acquiring unit includes a first set and a second set; the consumption behavior data to be trained in the first set have actual intention information, and the consumption behavior data to be trained in the second set do not have actual intention information; the second updating unit is specifically configured to, when the intention features are input to a decoder of the initial model for training to update the decoder and obtain the intention recognition model:

In a possible implementation manner, when the second updating unit performs, according to the decoder of the initial model, decoding processing on the intention features corresponding to the consumption behavior data to be trained in the second set, so as to classify the consumption behavior data to be trained in the second set into the obtained plurality of intention information, and obtain a two-class cross-entropy optimization loss function, the second updating unit is specifically configured to:

according to the decoder of the initial model, decoding intention features corresponding to the consumption behavior data to be trained in the second set so as to classify the consumption behavior data to be trained in the second set into the plurality of intention information; for each seed feature, performing similarity calculation on the sub-features of each consumption behavior data to be trained in the second set to obtain similarity information of each consumption behavior data to be trained in the second set on each sub-feature;

for each seed feature, determining score information of each consumption behavior data to be trained in the second set on each sub-feature, wherein the score information represents a score of the consumption behavior data to be trained belonging to the same intention information in the dimension of the sub-feature;

In a fifth aspect, an embodiment of the present application provides an electronic device, including: a memory, a processor;

a memory; a memory for storing the processor-executable instructions;

wherein the processor is configured to perform the method of the first aspect or the second aspect.

In a sixth aspect, embodiments of the present application provide a computer-readable storage medium, in which computer-executable instructions are stored, and when executed by a processor, the computer-executable instructions are configured to implement the method according to the first aspect or the second aspect.

In a seventh aspect, an embodiment of the present application provides a computer program product, where the computer program product includes: a computer program, stored in a readable storage medium, from which at least one processor of an electronic device can read the computer program, execution of the computer program by the at least one processor causing the electronic device to perform the method of the first or second aspect.

According to the intention recognition method, the intention recognition training method and the intention recognition equipment based on the intention recognition model, a set to be recognized is obtained; the set to be identified comprises a plurality of consumption behavior data to be identified, the consumption behavior data to be identified comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information; generating a plurality of dual hypergraph information according to the set to be identified; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user; inputting a plurality of dual hypergraph information into an encoder of an intention recognition model, and generating intention characteristics of consumption behavior data to be recognized; the intention characteristics characterize the sub-characteristics of the consumption behavior data to be recognized on each characteristic information; and inputting the intention characteristics into a decoder of an intention recognition model to obtain intention information of the consumption behavior data to be recognized. The intention identification method based on the intention identification model provided by the embodiment can be used for fully modeling the consumption behavior data and discovering unknown intention in the consumption behavior data based on a small amount of marked intention information, so that the accuracy and interpretability of e-commerce recommendation are improved, and the user experience is further improved. And the dual hypergraph is constructed, the corresponding intention characteristics are obtained by processing the dual hypergraph through the hypergraph convolution module, and the influence of the place information, the time information and the article information on the intention of the user is fully considered. And clustering the intention characteristics of the consumption behavior data to be identified to obtain a plurality of intention information, and classifying the consumption behavior data to be identified into the plurality of intention information by using a decoder of an intention identification model, so that the accuracy and the reliability of intention discovery can be improved. Obtaining an intention recognition model for recognizing consumption behavior data based on a model training mode; the intention information of the consumption behavior data can be accurately and quickly identified based on the model. And based on the intention characteristics of the sample pairs which are not marked with the intention, the intention decoder is trained in a mode of constructing the pseudo labels, so that the classification effect of the intention recognition model can be improved, and the accuracy and precision of the model are further improved.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and together with the description, serve to explain the principles of the application.

Fig. 1 is a schematic flowchart of an intention recognition method based on an intention recognition model according to an embodiment of the present application;

FIG. 2 is a schematic diagram of dual hypergraph information generation provided in an embodiment of the present application;

FIG. 3 is a schematic flowchart of another intent recognition method based on intent recognition models according to an embodiment of the present application;

FIG. 4 is a flowchart illustrating a model training method applied to intent recognition according to an embodiment of the present disclosure;

FIG. 5 is a schematic flowchart of another method for training a model applied to intent recognition according to an embodiment of the present disclosure;

FIG. 6 is a schematic diagram of an intent recognition model provided by an embodiment of the present application;

fig. 7 is a schematic structural diagram of an identification apparatus based on intention identification according to an embodiment of the present application;

FIG. 8 is a schematic structural diagram of a model training apparatus applied to intent recognition according to an embodiment of the present disclosure;

fig. 9 is a schematic structural diagram of an electronic device according to an embodiment of the present application;

fig. 10 is a block diagram of an electronic device according to an embodiment of the present application.

With the above figures, there are shown specific embodiments of the present application, which will be described in more detail below. These drawings and written description are not intended to limit the scope of the inventive concepts in any manner, but rather to illustrate the inventive concepts to those skilled in the art by reference to specific embodiments.

Detailed Description

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

With the development of internet technology, users can complete the purchase of goods based on the internet. And may recommend items to the user based on the internet. In order to facilitate recommendation of an item to a user, it is necessary to determine user intention information. The intention information characterizes the use and reason of the user after purchasing the item; for example, when a user purchases an item and applies the item to a family party, the intention information at this time is "family party". However, most existing models in the current industrial recommendation engine are based on the black box principle, the specific reasons and intentions of the user consumption behaviors cannot be clearly modeled, the accuracy of discovering the user consumption behavior intentions is difficult to meet the business requirements, and the consumption behaviors in the e-commerce business are difficult to process.

In one example, new ideas hidden in conversational sentences are discovered through adaptive clustering and cluster refinement. Or, the alignment strategy is used for enhancing the deep clustering to solve the problem of label inconsistency. Or, the clustering effect of the learning based on the deep embedded characterization is improved by introducing the front-back consistency constraint of the intention characteristics. Or generating paired pseudo labels based on the rank statistics of the feature vectors, so as to convert the clustering task into a binary task.

However, in the above manner, the learning about the intention characterization is mainly based on natural language processing and image feature extraction, and it is difficult to handle consumption behaviors in e-commerce business; related optimization is not performed on user behavior characteristics under the electric business service, effective service migration cannot be directly performed, and the accuracy of discovering the consumption intention information cannot meet the service requirement.

In one example, the corpora of the corresponding users are collected according to the service requirements and preprocessed; defining corresponding intentions according to user corpora; constructing and training an intention recognition model according to the intention, and optimizing; identifying the intention of the user corpus through an intention identification model; judging the intention type according to the current intention, and if the judged intention type is the same as the current service type, not updating the current intention; if the judged intention type is not consistent with the current service type, updating through an intention discovery model to obtain a new intention; the intention discovery model is constructed and trained according to user corpora and is used for clustering the user corpora and then redefining the intention according to the user corpora in the category and the service requirement. Or acquiring a plurality of texts to be processed from the saved dialog logs, and processing the plurality of texts to be processed to obtain a plurality of corresponding sentence vectors; clustering a plurality of texts to be processed based on a plurality of sentence vectors to obtain A-type texts to be processed; a is an integer of 1 or more; extracting keywords from each type of texts to be processed in the type A texts to be processed to obtain at least one keyword; determining at least one intention for the new addition based on the at least one keyword. Or, performing further pre-training on a transform-based bidirectional coder (BERT) model based on the data on the service; collecting user-generated conversation records on the project; obtaining an embedding vector by using a BERT model which is further pre-trained for a dialog text of a user; reducing the dimension of the embedding vector to a low-dimensional vector by using a machine learning algorithm, so that the embedding vector is a vector with representative characteristic information; using a machine learning algorithm for the reduced imbedding vector, and adjusting relevant hyper-parameters of the algorithm to obtain clustered text information; and (5) giving clustered linguistic data to an intention library maintainer, and providing reference for adding new intentions to the linguistic data.

However, in the above manner, the modeling of the user consumption behavior is not sufficient; the method of enabling knowledge migration from known intent to unknown intent is not applicable to e-commerce scenarios.

The application provides an intention recognition method, an intention recognition training method and intention recognition training equipment based on an intention recognition model, and aims to solve the technical problems in the prior art.

The following describes the technical solutions of the present application and how to solve the above technical problems with specific embodiments. The following several specific embodiments may be combined with each other, and details of the same or similar concepts or processes may not be repeated in some embodiments. Embodiments of the present application will be described below with reference to the accompanying drawings.

Fig. 1 is a schematic flowchart of an intention recognition method based on an intention recognition model according to an embodiment of the present application, and as shown in fig. 1, the method includes:

s101, acquiring a set to be identified; the set to be identified comprises a plurality of consumption behavior data to be identified, the consumption behavior data to be identified comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information.

For example, the execution subject of this embodiment may be an electronic device, or a terminal device, or a server, or a controller, or other devices or devices that may execute this embodiment, which is not limited in this respect.

Before a plurality of consumption behavior data in a set to be recognized need to be recognized, an intention recognition model needs to be obtained first, and the intention recognition model is used for recognizing the plurality of consumption behavior data in the set to be recognized.

The embodiment describes a manner of identifying a plurality of consumption behavior data in a set to be identified by the intention identification model. First, a set to be recognized for an intent recognition model needs to be obtained. The set to be recognized comprises a plurality of consumption behavior data to be recognized, and the consumption behavior data in the set to be recognized come from an e-commerce platform.

Each consumption behavior data to be identified comprises user information U and a plurality of characteristic information. The feature information includes location information, time information, article information, and the like, and is referred to as location information L, time information T, and article information C, respectively.

The location information L refers to a location where the user generates consumption behavior, for example, an office location, or a home; time information T, which refers to a period of time during which the user generates consumption behavior, for example, 24 hours per day is divided by half an hour, and 9; the item information C refers to an item purchased by the user, for example, a movie ticket, or clothes.

The intention information of the user for generating the consumption behavior is highly related to three characteristic information, namely, the location information, the time information, the article information and the like. The place information, the consumption behavior of the user and the place information are related, for example, when the user is near an office building, the consumption behavior related to the place information, such as fast food, may be consumed; time information, wherein the consumption behavior of the user is related to the time information, for example, the user may consume wine at night and other consumption behaviors related to the time information; the item information and the consuming behavior of the user are related to the item information, i.e. the user is interested in preferences, e.g. the user likes to watch movies and may have a consuming behavior of buying movie tickets. Therefore, the user information and each feature information have a binary relationship, and the user information U, the location information L, the time information T and the article information C in each consumption behavior data are decomposed into three types of binary relationships, namely, user information-location information (U-L), user information-time information (U-T) and user information-article information (U-C).

S102, generating a plurality of dual hypergraph information according to a set to be identified; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user.

Illustratively, the set to be identified includes a plurality of consumption behavior data, wherein each consumption behavior data has three types of binary relations, and three sets of dual hypergraph information are generated respectively for the user information and the feature information in each binary relation, wherein each set of dual hypergraph information includes two hypergraphs, a first hypergraph is an association relation of the user information on the feature information, and a second hypergraph is an association relation of the feature information on the user information.

Therefore, the application generates three groups of dual hypergraph information, namely the dual hypergraph information under the user information-location information (U-L), the dual hypergraph information under the user information-time information (U-T) and the dual hypergraph information under the user information-article information (U-C). In dual hypergraph information under user information-location information (U-L), a first hypergraph is an incidence relation of the user information on the location information, and a second hypergraph is an incidence relation of the location information on the user information; in dual hypergraph information under user information-time information (U-T), a first hypergraph is an incidence relation of the user information on the time information, and a second hypergraph is an incidence relation of the time information on the user information; in dual hypergraph information under user information-article information (U-C), a first hypergraph is an association relationship of the user information on the article information, and a second hypergraph is an association relationship of the article information on the user information.

For n consumption behavior data, a set to be identified I = { x } may be constructed ₁ ，x ₂ ，x ₃ ，x ₄ ，...，x _k ，...，x _n N is a positive integer of 2 or more, and k is a positive integer of 1 or more and n or less. Kth consumption behavior data x _k Including user information, location information, time information, and article information: x is the number of _k ＝(u _k ，l _k ，t _k ，c _k ) Indicating that user k purchased item k at location k, time k; wherein u is _k User information for the kth consumption behavior data, l _k Location information for the kth consumption behavior data, t _k Time information for the kth consumption behavior data, c _k Item information that is the kth consumption behavior data.

For example, fig. 2 is a schematic diagram of generation of dual hypergraph information according to an embodiment of the present application. As shown in FIG. 2, a set l to be identified ₁ Including 5 consumption behavior data I ₁ ＝{x ₁ ，x ₂ ，x ₃ ，x ₄ ，x ₅ }. In generating a set I to be identified ₁ In the case of dual hypergraph information under user information-location information (U-L), as to the association of user information on location information, a first hypergraph corresponding to location information

When the user 1, the user 2 and the user 5 generate consumption behaviors at the place 1, the user information u ₁ User information u ₂ User information u ₅ Using a super-edge connection, user 1 and user 4 both generate consumption behavior at location 3, and user information u ₁ User information u ₄ Using a super edge connection; for the association of the location information on the user information, i.e. the second hypergraph G corresponding to the location information ^L If the user 1 has consumed the behavior at the location 1, the location 2, or the location 3, the location information l is obtained ₁ Location information l ₂ Location information l ₃ Using a super-edge connection, the user 5 generates consumption behaviors at the place 1 and the place 4, and the place information l ₁ Location information l ₄ A super edge connection is used.

In generating a set I to be identified ₁ For the dual hypergraph information under the user information-time information (U-T), the correlation relationship of the user information on the time information, that is, the first hypergraph corresponding to the time information

When the user 1, the user 2 and the user 4 generate consumption behaviors at the time 1, the user information u ₁ User information u ₂ User information u ₄ Using a super-edge connection, user 1, user 5 generate consumption behavior at time 2, and user information u ₁ User information u ₅ Using a super edge connection; for the association of the time information on the user information, i.e. the second hypergraph G corresponding to the time information ^T When the user 1 generates consumption behaviors at the time 1 and the time 2, the time information t ₁ Time information t ₂ Using a super edge connection, the user 5 generates consumption behavior at time 2, time 3, and time 4, and the time information t is ₂ Time information t ₃ Time information t ₄ A super edge connection is used.

When dual hypergraph information under user information-article information (U-C) is generated, the incidence relation of the user information on the article information, namely a first hypergraph corresponding to the article information

When the user 2, the user 3, and the user 4 purchase the item 2, the user information u ₂ User information u ₃ User information u ₄ Using a super edge connection, user information u is provided when user 1 and user 5 purchase item 4 ₁ User information u ₅ Using a super edge connection; for the association relationship of the article information on the user information, i.e. the second hypergraph G corresponding to the article information ^C When the user 1 purchases the item 1, the item 3, or the item 4, the item information c ₁ Article information c ₃ Article information c ₄ When the user 3 purchases the items 2 and 5 using a super-edge connection, the item information c ₂ Article information c ₅ A super edge connection is used.

S103, inputting the information of the dual hypergraphs into an encoder of an intention recognition model to generate intention characteristics of consumption behavior data to be recognized; wherein the intention characteristics characterize the sub-characteristics of the consumption behavior data to be recognized on each characteristic information.

Illustratively, the dual hypergraph information under the user information-location information (U-L), the dual hypergraph information under the user information-time information (U-T), and the dual hypergraph information under the user information-article information (U-C) are respectively input into an encoder of an intention recognition model, and three intention features corresponding to three sets of dual hypergraph information of each consumption behavior data in the set to be recognized, namely an intention sub-feature of a location subspace, an intention sub-feature of a time subspace, and an intention sub-feature of an article subspace, are generated.

And S104, inputting the intention characteristics into a decoder of an intention recognition model to obtain intention information of consumption behavior data to be recognized.

Illustratively, three intention features output by an encoder of the intention recognition model, namely an intention sub-feature of a place subspace, an intention sub-feature of a time subspace and an intention sub-feature of an article subspace are respectively input into a decoder of the intention recognition model, and intention information of each consumption behavior data in the set to be recognized is obtained.

In the embodiment, a set to be identified is obtained; the set to be identified comprises a plurality of consumption behavior data to be identified, the consumption behavior data to be identified comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information; generating a plurality of dual hypergraph information according to the set to be identified; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user; inputting a plurality of dual hypergraph information into an encoder of an intention recognition model, and generating intention characteristics of consumption behavior data to be recognized; the intention characteristics characterize the sub-characteristics of the consumption behavior data to be recognized on each characteristic information; and inputting the intention characteristics into a decoder of an intention recognition model to obtain intention information of the consumption behavior data to be recognized. The intention identification method based on the intention identification model provided by the embodiment can be used for fully modeling the consumption behavior data and discovering unknown intention information in the consumption behavior data based on a small amount of marked intention information, so that the accuracy and interpretability of e-commerce recommendation are improved, and the user experience is further improved.

Fig. 3 is a schematic flowchart of another intent recognition method based on an intent recognition model according to an embodiment of the present application, and as shown in fig. 3, the method includes:

s301, acquiring a set to be identified; the set to be identified comprises a plurality of consumption behavior data to be identified, the consumption behavior data to be identified comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information.

For this step, reference may be made to the description of step S101, which is not described again.

S302, generating a plurality of dual hypergraph information according to the set to be identified; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user.

In one example, the first hypergraph has a first correlation matrix characteristic, a first node degree matrix characteristic, a first hyper-edge matrix characteristic, and a first node characteristic.

The first incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the first hypergraph belongs; the first node degree matrix characteristic represents the connection relation of user nodes in the first hypergraph; the first hyper-edge matrix characteristic represents the node corresponding relation of edges in the first hyper-graph; the first node features characterize node information of user nodes in the first hypergraph. The second hypergraph has a second incidence matrix characteristic, a second node degree matrix characteristic, a second hyper-edge degree matrix characteristic and a second node characteristic; the second incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the second hypergraph belongs; the second node degree matrix characteristic represents the connection relation of characteristic information nodes in the second hypergraph; the second hyperedge matrix characteristic represents the node corresponding relation of edges in the second hypergraph; the second node characteristics characterize node information of the characteristic information nodes in the second hypergraph.

Illustratively, multiple sets of dual hypergraph information are generated based on multiple consumption behavior data in the set to be identified, wherein each set of dual hypergraph information comprises a first hypergraph of the incidence relation of the user information on the feature information and a second hypergraph of the incidence relation of the feature information on the user information. When the dual hypergraph information is generated, each hypergraph in the dual hypergraph information has an incidence matrix characteristic H, a node degree matrix characteristic D, a hyperedge degree matrix B and a node characteristic X; the first hypergraph has a first incidence matrix characteristic H ₁ First node degree matrix characteristic D ₁ First overcritical matrix feature B ₁ And a first node characteristic X ₁ (ii) a The second hypergraph has a second correlation matrix characteristic H ₂ Second node degree matrix characteristic D ₂ Second super-edge matrix characteristic B ₂ And a second node characteristic X ₂ 。

The first incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the first hypergraph belongs; the first node degree matrix characteristic represents the connection relation of user nodes in the first hypergraph; the first hyper-edge matrix characteristic represents the node corresponding relation of edges in the first hyper-graph; the first node features characterize node information of user nodes in the first hypergraph. The second incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the second hypergraph belongs; the second node degree matrix characteristic represents the connection relation of characteristic information nodes in the second hypergraph; the second hyperedge matrix characteristic represents the node corresponding relation of edges in the second hypergraph; the second node characteristics represent node information of characteristic information nodes in the second hypergraph; the characteristic information is location information, or time information, or article information.

For a specific process of generating the dual hypergraph information, reference may be made to the description of S102, which is not described again.

S303, inputting the information of the plurality of dual hypergraphs into the encoder, and generating convolution characteristics corresponding to each hypergraph convolution layer in the encoder.

In one example, step S303 includes the following process:

the first step of step S303: a plurality of dual hypergraph information is input to an encoder, and a first initial embedding feature corresponding to the feature information is generated according to the feature of the first hypergraph corresponding to the feature information and the partial feature of the second hypergraph.

The second step of step S303: and generating a second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph.

The second step of step S303: and processing based on each hypergraph convolutional layer according to the first initial embedding characteristic and the second initial embedding characteristic corresponding to the characteristic information to obtain a convolutional characteristic corresponding to each hypergraph convolutional layer.

In one example, the first step of step S303 includes the following processes: determining the propagation characteristics in the first graph corresponding to the characteristic information according to the first incidence matrix characteristics, the first node degree matrix characteristics, the first super-edge degree matrix characteristics and the first node characteristics of the first hypergraph corresponding to the characteristic information; wherein, the propagation characteristics in the first graph represent the characteristic expression of the user on the characteristic information; determining the first inter-graph propagation characteristics corresponding to the characteristic information according to the second incidence matrix characteristics, the second super-edge matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; the first inter-graph propagation characteristics represent characteristic expressions between users and characteristic information; and determining a first initial embedding feature corresponding to the feature information according to the first intra-map propagation feature corresponding to the feature information and the first inter-map propagation feature corresponding to the feature information.

In one example, the second step of step S303 includes the following processes: determining the propagation characteristics in the second graph corresponding to the characteristic information according to the second incidence matrix characteristics, the second node degree matrix characteristics, the second super-edge degree matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; the feature expression of the feature characterization feature information on the user is propagated in the second graph; determining a second inter-graph propagation characteristic corresponding to the characteristic information according to the first incidence matrix characteristic, the first hyper-edge matrix characteristic and the first node characteristic of the first hyper-graph corresponding to the characteristic information; the second inter-graph propagation characteristics represent characteristic expressions between the user and the characteristic information; and determining a second initial embedding feature corresponding to the feature information according to the second intra-map propagation feature corresponding to the feature information and the second inter-map propagation feature corresponding to the feature information.

Illustratively, the generated three sets of dual Hypergraph information are input into an encoder of the intention recognition model, and a Hypergraph Convolution layer (Joint-HGC) in the encoder processes each set of dual Hypergraph information respectively to generate Convolution characteristics corresponding to each set of dual Hypergraph information. And each supergraph convolutional layer processes the even supergraph information of each group, including an intra-graph propagation process and an inter-graph propagation process, and generates an intra-graph propagation characteristic and an inter-graph propagation characteristic. In the intra-graph propagation process of the first hypergraph, the incidence matrix characteristic, the node degree matrix characteristic, the super-edge degree matrix characteristic and the node characteristic are utilized to generate a first intra-graph propagation characteristic; generating a first inter-graph propagation characteristic by utilizing the incidence matrix characteristic, the super-edge matrix characteristic and the node characteristic of the second hypergraph in the inter-graph propagation process of the first hypergraph; and generating a first initial embedded feature according to the generated first intra-graph propagation feature and the first inter-graph propagation feature. Generating a second intra-graph propagation characteristic by utilizing the incidence matrix characteristic, the node degree matrix characteristic, the super-edge degree matrix characteristic and the node characteristic of the second hyper-graph in the intra-graph propagation process; and in the inter-graph propagation process, generating a second inter-graph propagation characteristic by using the incidence matrix characteristic, the super-edge matrix characteristic and the node characteristic of the first hypergraph, and generating a second initial embedding characteristic according to the generated second intra-graph propagation characteristic and the second inter-graph propagation characteristic.

Before the hypergraph convolution layer of the encoder processes each pair of even hypergraph information, each pair of even hypergraph information needs to be embedded to generate a dual hypergraphThe initial node features of the first hypergraph and the initial node features of the second hypergraph within the information. Wherein, the first hypergraph corresponding to the location information in the dual hypergraph information under the user information-location information (U-L)

Is not present>

Second hypergraph G corresponding to place information _L Is not present>

The first hypergraph corresponding to the time information in the dual hypergraph information under the user information-time information (U-T)>

Is not present>

Second hypergraph G corresponding to time information _T Is not present>

The first hypergraph corresponding to the item information in the dual hypergraph information under the user information-item information (U-C)>

Is not present>

Second hypergraph G corresponding to article information _C Is not present>

The encoder has a total of W layers for the hypergraph convolution layer, the first layer of the hypergraph convolution layer, for the embedded initial node feature X ⁽⁰⁾ Updating to generate the first layerCorresponding convolution feature X ⁽¹⁾ (ii) a Convolution characteristics X of Q layer of hypergraph convolution layer to Q-1 layer ^(Q-1) Updating to generate convolution characteristic X corresponding to the Q-th layer ^(Q) Wherein Q is a positive integer greater than or equal to 1 and less than or equal to W; convolution characteristics X of W-1 layer of hypergraph convolution layer ^(W-1) Updating to generate convolution characteristic X corresponding to W-th layer ^(W) 。

For example, the hypergraph convolutional layer processes dual hypergraph information under user information-location information (U-L), and the specific process of processing the Q-th hypergraph convolutional layer is as follows: for the first hypergraph corresponding to location information

In particular, use is made of>

Is/are as follows

Convolution feature generated at layer Q-1->

Generating a first intra-map propagation feature and using a second hypergraph G corresponding to location information _L H of (A) to (B) _L 、B _L The convolution characteristic generated by layer Q-1->

Generating a first inter-graph propagation characteristic, wherein the convolution characteristic of the first hypergraph corresponding to the location information and output by the Q-th layer hypergraph convolution layer is the sum of the first intra-graph propagation characteristic and the first inter-graph propagation characteristic; for the second hypergraph G corresponding to the location information _L In other words, using G _L H of (A) to (B) _L 、D _L 、B _L Layer Q-1 generated convolution features->

Generating a second intra-map propagation feature and utilizing the first hypergraph @, which corresponds to location information>

Is/are as follows

Convolution feature generated at layer Q-1->

And generating a second inter-graph propagation characteristic, wherein the convolution characteristic of the second hypergraph corresponding to the location information and output by the Q-th layer hypergraph convolution layer is the sum of the second intra-graph propagation characteristic and the second inter-graph propagation characteristic.

The hypergraph convolutional layer processes dual hypergraph information under user information-time information (U-T), and the specific process of processing the hypergraph convolutional layer at the Q layer is as follows: for the first hypergraph corresponding to the time information

In particular, use is made of>

Is/are>

Convolution features generated at layer Q-1 @>

Generating a first intra-graph propagation feature and using a second hypergraph G corresponding to time information _T H of (A) to (B) _T 、B _T The convolution characteristic generated by layer Q-1->

Generating a first inter-graph propagation characteristic, wherein the convolution characteristic of the first hypergraph corresponding to the time information and output by the Q-th layer hypergraph convolution layer is the sum of the first intra-graph propagation characteristic and the first inter-graph propagation characteristic; for the second hypergraph G corresponding to the time information _T In other words, using G _T H of (A) to (B) _T 、D _T 、B _T Layer Q-1 formationCharacteristic of convolution of

Generating propagation features in a second graph and utilizing a first hypergraph corresponding to time information>

In:>

convolution features generated at layer Q-1 @>

And generating a second inter-graph propagation characteristic, wherein the convolution characteristic of the second hypergraph corresponding to the time information and output by the Q-th layer hypergraph convolution layer is the sum of the second intra-graph propagation characteristic and the second inter-graph propagation characteristic.

The hypergraph convolution layer processes dual hypergraph information under user information-article information (U-C), and the specific process of processing the Q-th hypergraph convolution layer is as follows: for the first hypergraph corresponding to the article information

In particular, use is made of>

Is/are>

Convolution feature generated at layer Q-1->

Generating a first intra-map propagation feature and utilizing a second hypergraph G corresponding to item information _C H of (A) _C 、B _C Layer Q-1 generated convolution features->

Generating a first inter-graph propagation characteristic, wherein the convolution characteristic of the first hypergraph corresponding to the article information and output by the Q-th layer hypergraph convolution layer is the sum of the first intra-graph propagation characteristic and the first inter-graph propagation characteristic; for the second hypergraph G corresponding to the article information _C In other words, using G _C H of (A) _C 、D _C 、B _C The convolution characteristic generated by layer Q-1->

Generating a second intra-map propagation feature and utilizing the first hyper-map corresponding to the item information>

In:>

convolution feature generated at layer Q-1->

And generating a second inter-graph propagation characteristic, wherein the convolution characteristic of the second hypergraph corresponding to the article information and output by the Q-th layer hypergraph convolution layer is the sum of the second intra-graph propagation characteristic and the second inter-graph propagation characteristic.

The Q-th layer hypergraph convolution layer respectively processes convolution characteristics output by the Q-1 th layer of the three groups of dual hypergraph information to generate propagation characteristics in the graph

The calculation formula of (a) is as follows:

wherein the content of the first and second substances,

representing the calculation process of the matrix, and taking the reciprocal after carrying out open square root calculation on each element in the matrix; -1 represents the inverse of the matrix; t represents the transpose of the matrix; formula 1 shows that the Q-th hypergraph convolution layer processes dual hypergraph information under user information-location information to generate a first hypergraph in-first-graph propagation characteristic corresponding to the location information

Equation 2 shows that the Q-th hypergraph convolutional layer processes the dual hypergraph information under the user information-location information to generate the propagation characteristic ^ based on the location information in the second map of the second hypergraph>

Formula 3 shows that the Q-th hypergraph convolution layer processes dual hypergraph information under user information-time information to generate a first graph propagation characteristic of a first hypergraph corresponding to the time information; formula 4 shows that the Q-th hypergraph convolutional layer processes the dual hypergraph information under the user information-time information to generate the second graph internal propagation characteristics of the second hypergraph corresponding to the time information; equation 5 shows that the Q-th hypergraph convolution layer processes the dual hypergraph information under the user information-article information to generate the productPropagating features within a first graph of a first hypergraph to which information corresponds; equation 6 shows that the Q-th hypergraph convolution layer processes the dual hypergraph information under the user information-article information to generate the second in-map propagation characteristic of the second hypergraph corresponding to the article information.

The Q-th layer of hypergraph convolution layer respectively processes the convolution characteristics output by the last layer of three groups of dual hypergraph information to generate inter-graph propagation characteristics

The calculation formula of (a) is as follows:

formula 7 shows that the Q-th hypergraph convolutional layer processes dual hypergraph information under user information-location information to generate first inter-graph propagation characteristics of a first hypergraph corresponding to the location information; formula 8 shows that the Q-th hypergraph convolutional layer processes the dual hypergraph information under the user information-location information to generate the inter-second-graph propagation characteristics of the second hypergraph corresponding to the location information; formula 9 shows that the Q-th hypergraph convolution layer processes dual hypergraph information under user information-time information to generate first inter-graph propagation characteristics of a first hypergraph corresponding to the time information; formula 10 shows that the Q-th hypergraph convolutional layer processes the dual hypergraph information under the user information-time information to generate the inter-second-graph propagation characteristics of the second hypergraph corresponding to the time information; formula 11 shows that the Q-th hypergraph convolutional layer processes the dual hypergraph information under the user information-article information to generate the first inter-graph propagation characteristics of the first hypergraph corresponding to the article information; equation 12 shows that the Q-th layer hypergraph convolution layer processes the dual hypergraph information under the user information-article information to generate the inter-second-graph propagation characteristics of the second hypergraph corresponding to the article information.

The Q-th layer of hypergraph convolution layer respectively processes the convolution characteristics output by the last layer of three groups of dual hypergraph information to generate convolution characteristics X ^(Q) The calculation formula of (a) is as follows:

wherein, formula 13 indicates that the Q-th hypergraph convolution layer processes the dual hypergraph information under the user information-location information to generate a Q-th convolution characteristic of the first hypergraph corresponding to the location information; formula 14 shows that the Q-th hypergraph convolution layer processes the dual hypergraph information under the user information-location information to generate a Q-th convolution characteristic of the second hypergraph corresponding to the location information; formula 15 shows that the Q-th hypergraph convolution layer processes the dual hypergraph information under the user information-time information to generate a Q-th convolution characteristic of the first hypergraph corresponding to the time information; formula 16 shows that the Q-th hypergraph convolution layer processes dual hypergraph information under user information-time information to generate a Q-th convolution characteristic of the second hypergraph corresponding to the time information; formula 17 shows that the Q-th hypergraph convolution layer processes dual hypergraph information under user information-article information to generate a Q-th convolution characteristic of the first hypergraph corresponding to the article information; equation 18 shows that the Q-th hypergraph convolution layer processes the dual hypergraph information under the user information-article information to generate the Q-th convolution characteristic of the second hypergraph corresponding to the article information.

Since D and B can be calculated by H, the calculation process of the hypergraph convolution layer is simplified to a function expression, which is specifically expressed as follows:

/>

wherein Joint-HGC () represents the process by which the hypergraph convolutional layer processes data to generate convolutional features.

And S304, aggregating the convolution features to obtain an aggregated feature.

Illustratively, the convolution features generated by each of the W-layer hypergraph convolution layers are aggregated to obtain aggregated features corresponding to the three sets of dual hypergraph information.

Specifically, after the W-layer hypergraph convolution layer performs convolution processing on the three sets of dual hypergraph information, W convolution features are generated for each hypergraph of each set of dual hypergraph information. Therefore, it is necessary to connect a pooling layer after the W layers of hypergraph convolution layers, and perform pooling aggregation processing on the W convolution features generated for each hypergraph to generate an aggregated feature corresponding to each hypergraph. After the pooling polymerization treatment, each group of the three groups of dual hypergraph information generates two corresponding polymerization characteristics to the dual hypergraph information. Convolution characteristic X of dual hypergraph generation under user information-location information (U-L) _U，L 、X _L The corresponding aggregate characteristics are represented as

Convolution characteristic X of dual hypergraph generation under user information-time information (U-T) _U，T 、X _T The corresponding aggregate characteristic is expressed as->

Convolution feature X of dual hypergraph generation under user information-time information (U-C) _U，C 、X _C The corresponding aggregate characteristic is expressed as->

S305, determining the intention characteristics of the consumption behavior data to be identified according to the aggregation characteristics.

In one example, step S305 includes the following process:

Illustratively, three groups of dual hypergraph information are aggregated to generate an aggregation feature, and the aggregation feature is input into the multi-layer perceptron algorithm model to be processed to generate an intention feature.

Specifically, two per dual hypergraph information generationAnd aggregating the features, and generating an intention feature after splicing the features by a multilayer perceptron. Three sets of dual hypergraph information corresponding to each consumption behavior data to be identified generate three disentangled intention sub-features, namely an intention sub-feature in a place sub-space, an intention sub-feature in a time sub-space and an intention sub-feature in an article sub-space. I consumption behavior data x _i Where i is a positive integer greater than or equal to 1, the corresponding intent features may be expressed as follows:

wherein MLP represents an algorithmic model of the multi-layered perceptron.

The de-entanglement processing refers to processing of dual hypergraph information of location information, dual hypergraph information of time information, and dual hypergraph information of article information.

S306, the set to be identified comprises a first set and a second set; the consumption behavior data to be identified in the first set has actual intention information, and the consumption behavior data to be identified in the second set does not have actual intention information. And clustering the intention characteristics of the consumption behavior data to be identified to obtain a plurality of intention information.

Illustratively, consumption behavior data to be identified in the set to be identified are divided into two sets, one set is a first set composed of M consumption behavior data labeled with intention information, and the other set is a second set composed of N consumption behavior data not labeled with intention information, wherein M, N are positive integers greater than or equal to 1. The consumption behavior data marked with the intention information is based on the suggestion of experts in the related field, and the marked intention information category is marked as a K 'category, wherein K' is a positive integer greater than or equal to 1. And processing the intention characteristics of the M + N consumption behavior data to be identified in the set to be identified by utilizing K-means clustering, and estimating to obtain all intention information categories corresponding to all consumption behavior data.

The intention information refers to the reason why the user has made a consumption action, i.e., the use of the purchased article. For example, a user generates a consumption behavior of purchasing movie tickets, whose consumption intention may be "family party".

The encoder of the intention recognition model processes the M + N consumption behavior data to be recognized in the set to be recognized, and respectively generates intention sub-features in M + N place subspaces, intention sub-features in M + N time subspaces and intention sub-features in M + N article subspaces. And respectively processing 3 (M + N) intention features by utilizing K-means clustering, taking f times of the known intention K 'as the number of all kinds of intention information, wherein f is a positive integer greater than or equal to 1, and clustering the 3 (M + N) intention features into f multiplied by K' clusters. After clustering, the number of abandoned intention features is less than

Cluster of individuals, retaining the number of intended features more than ^ greater than ^>

And the number of the obtained intention information types of the clusters is marked as K.

S307, according to the decoder of the intention identification model, decoding the intention characteristics corresponding to the consumption behavior data to be identified so as to classify the consumption behavior data to be identified into a plurality of pieces of intention information to obtain the intention information of the consumption behavior data to be identified.

In one example, step S307 includes the following process:

first step of step S307: according to the decoder of the intention identification model, decoding intention characteristics corresponding to consumption behavior data to be identified so as to classify the consumption behavior data to be identified into a plurality of acquired intention information and acquire probability distribution information of the consumption behavior data to be identified on each characteristic information; wherein the probability distribution information represents the probability distribution condition that the consumption behavior data to be identified belongs to the intention information in the plurality of intention information.

Second step of step S307: and determining the weight information of the consumption behavior data to be identified on each characteristic information according to the sub-characteristics of the consumption behavior data to be identified on each characteristic information.

The third step of step S307: and determining the initial intention of the consumption behavior data to be identified on each characteristic information according to the probability distribution information and the weight information of the consumption behavior data to be identified on each characteristic information.

The fourth step of step S307: and determining intention information of the consumption behavior data to be identified according to each initial intention corresponding to the consumption behavior data to be identified.

Exemplarily, an intention sub-feature in a place sub-space, an intention sub-feature in a time sub-space and an intention sub-feature in an article sub-space corresponding to consumption behavior data to be recognized are respectively input into a decoder of an intention recognition model, and decoding processing is performed to obtain probability distribution information of place information under K intention information types, probability distribution information of time information under K intention information types and probability distribution information of article information under K intention information types; calculating weight information of consumption behavior data to be identified on the place information, the time information and the article information by utilizing the normalized index function; respectively selecting intention information with the highest probability on the place information, intention information with the highest probability on the time information and intention information with the highest probability on the article information as initial intention information of the consumption behavior data to be identified by utilizing the weight information on the place information, the weight information on the time information, the weight information on the article information, and the probability distribution information on the place information, the probability distribution information on the time information and the probability distribution information on the article information output by the decoder, and then selecting the intention information with the highest probability from the three initial intention information as the intention information corresponding to the consumption behavior data to be identified.

And the decoder of the intention identification model respectively decodes intention sub-features of the consumption behavior data to be identified in a place sub-space, intention sub-features in a time sub-space and intention sub-features in an article sub-space, and outputs three K-dimensional probability distribution information, wherein each dimension represents the probability that the intention features belong to the intention information.

According to the intention sub-feature in the place sub-space, the intention sub-feature in the time sub-space and the intention sub-feature in the article sub-space, the weight information on the place information, the weight information on the time information and the weight information on the article information are calculated by utilizing the normalized exponential function, and the specific process is as follows:

α ^L ＝softmax(q ^T Φ ^L (x _i ))

α ^T ＝softmax(q ^T Φ ^T (x _i ))

α ^C ＝softmax(q ^T Φ ^C (x _i ))

wherein alpha is ^L Weight information, alpha, representing location information ^T Representing weight information on time information, alpha ^C Weight information on the article information is represented, and the sum of the three weight information is equal to 1; softmax represents a normalized exponential function; q represents a learnable attention vector, T represents a transposition calculation; phi denotes the encoder of the intention recognition model, x _i Representing the ith consumption behavior data to be identified, i is a positive integer greater than or equal to 1 and less than or equal to M + N, phi ^L (x _i ) Representing the intention characteristics of the ith consumption behavior data to be recognized in the place subspace; phi ^T (x _i ) Representing the intention sub-feature of the ith consumption behavior data to be recognized in a time subspace; phi ^T (x _i ) And representing the intention sub-characteristic of the ith consumption behavior data to be recognized in the item subspace.

And multiplying the probability distribution information based on the K dimensions and the normalized weight of the probability distribution information, and selecting intention information corresponding to the maximum probability in the K probability information as initial intention information. Based on the process, the initial intention information on the place information, the initial intention information on the time information and the initial intention information on the article information are respectively obtained, and the specific process is as follows:

wherein the content of the first and second substances,

indicating the initial intention information of the ith consumption behavior data to be identified on the place information; />

Indicating the initial intention information of the ith consumption behavior data to be identified on the time information; />

Indicating that the ith consumption behavior data to be identified is initial intention information on the article information; argmax represents a function of taking the maximum; />

Probability distribution information representing the ith consumption behavior data to be identified on the place information; />

Probability distribution information representing the ith consumption behavior data to be identified on the time information; />

Indicating the ith to be recognizedProbability distribution information of the fee behavior data on the article information; i is a positive integer of not less than 1 and not more than M + N.

Selecting initial intention information of ith to-be-identified consumption behavior data on location information, initial intention information of ith to-be-identified consumption behavior data on time information, and initial intention information with the highest probability of the ith to-be-identified consumption behavior data on article information, wherein i is a positive integer which is greater than or equal to 1 and less than or equal to M + N, and taking the positive integer as intention information corresponding to the to-be-identified consumption behavior data, and the specific process is as follows:

wherein, the first and the second end of the pipe are connected with each other,

and indicating intention information corresponding to the ith consumption behavior data to be identified, wherein i is a positive integer which is greater than or equal to 1 and less than or equal to M + N.

And S308, recommending articles to the user according to the intention information of the consumption behavior data to be identified.

Illustratively, the items are recommended to the user according to intention information corresponding to the consumption behavior data to be recognized and output by the intention recognition model.

The intention recognition model can be applied to all business scenes related to user behavior understanding, including but not limited to recommended articles, user understanding, user portrait construction and targeted market release.

The e-commerce platform collects the consumption behavior data of the user and obtains intention information of part of the consumption behavior data through a webpage or an application program. The intention recognition model provided by the application can process the data to discover unknown intention information. In the process of identifying the intention, the item information with the highest frequency of occurrence under each kind of intention information can be counted to generate the mapping relation between the intention information and the item information. Recall strategies in recommendations can be extended based on intent information-item information. In addition, intention characteristics generated by an encoder of the intention recognition model can be used as input to be added to the recommendation model, and recommendation efficiency is improved; and adding the discovered intention information into an item information list pushed to the user as a recommendation reason of the item.

The e-commerce platform collects consumption behavior data of the user and partial acquired intention information through a webpage or an application program, and the intention identification model provided by the application can process the data and discover unknown intention information. Furthermore, the attribute of intention information can be added to each user in actual service, so that the portrait information of the user is enriched, and the understanding of the consumption behavior of the user is promoted.

The E-commerce platform collects consumption behavior data of the user and acquired partial intention information through a webpage or an application program, and the intention identification model provided by the application can process the data and find unknown intention information. Because each user consumption behavior comprises user information and article information, after the intention information is obtained, a relation index of intention information-user information and intention information-article information can be established, and then article matching is carried out according to the intention information-article information index, so that targeted market release for a specific user is realized.

In the embodiment, a set to be identified is obtained; the set to be identified comprises a plurality of consumption behavior data to be identified, the consumption behavior data to be identified comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information; generating a plurality of dual hypergraph information according to the set to be identified; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user; inputting a plurality of dual hypergraph information into an encoder of an intention recognition model, and generating intention characteristics of consumption behavior data to be recognized; the intention characteristics represent sub characteristics of the consumption behavior data to be recognized on each characteristic information; and inputting the intention characteristics into a decoder of an intention recognition model to obtain intention information of the consumption behavior data to be recognized. The intention identification method based on the intention identification model provided by the embodiment can be used for fully modeling the consumption behavior data and discovering unknown intention information in the consumption behavior data based on a small amount of marked intention information, so that the accuracy and interpretability of e-commerce recommendation are improved, and the user experience is further improved. And by constructing the dual hypergraph and processing the dual hypergraph by using a hypergraph convolution module, the corresponding intention characteristics are obtained, and the influence of the place information, the time information and the article information on the intention information is fully considered. And clustering the intention characteristics of the consumption behavior data to be identified to obtain a plurality of intention information, and classifying the consumption behavior data to be identified into the plurality of intention information by using a decoder of an intention identification model, so that the accuracy and the reliability of intention discovery can be improved.

Fig. 4 is a schematic flowchart of a model training method applied to intent recognition according to an embodiment of the present application, and as shown in fig. 4, the method includes:

s401, acquiring a set to be trained; the to-be-trained set comprises a plurality of to-be-trained consumption behavior data, the to-be-trained consumption behavior data comprise a user and a plurality of feature information, and a binary relation exists between the user and each feature information.

Firstly, an intention recognition initial model is needed to be obtained, the intention recognition initial model is trained by utilizing a set to be trained, and then the intention recognition model is obtained. Initial model hyper-parameters including negative sampling number, batch size, embedding size, learning rate, regular terms and the like need to be set in the initial model training process. In the process of training the network, the weights and bias values of each layer of the network can be updated by a random gradient descent method in the process of back propagation.

This embodiment describes how to train the initial model of intent recognition to obtain the model of intent recognition. Firstly, a to-be-trained set for initial model training needs to be acquired. The consumption behavior data in the set to be trained come from an e-commerce platform.

The "consumption behavior data includes a user and a plurality of feature information, and the user has a binary relationship with each feature information" may be referred to the introduction of S101, and will not be described herein again.

S402, generating a plurality of dual hypergraph information according to a set to be trained; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the association relationship of the user on the feature information, and the second hypergraph represents the association relationship of the feature information on the user.

Illustratively, the set to be trained comprises a plurality of consumption behavior data to be trained, wherein each consumption behavior data has three types of binary relations, and three sets of dual hypergraph information are generated respectively aiming at the user information and the feature information in each binary relation, wherein each set of dual hypergraph information comprises two hypergraphs, the first hypergraph is the incidence relation of the user information on the feature information, and the second hypergraph is the incidence relation of the feature information on the user information.

The process of "generating multiple dual hypergraph information" can be referred to the introduction of S102, and is not described herein again.

S403, inputting the information of the dual hypergraph into an encoder of an initial model for training, and generating intention characteristics of consumption behavior data to be trained; the intention features represent sub-features of the consumption behavior data to be trained on each feature information; and updates the encoder according to the intent characteristics.

Illustratively, the dual hypergraph information under the user information-location information (U-L), the dual hypergraph information under the user information-time information (U-T) and the dual hypergraph information under the user information-article information (U-C) are respectively input into an encoder of an intention recognition initial model, three intention sub-features corresponding to three groups of dual hypergraphs of each consumption behavior data to be trained, namely, an intention sub-feature of a location subspace, an intention sub-feature of a time subspace and an intention sub-feature of an article subspace are generated, and the encoder of the intention recognition initial model is updated based on the intention features.

S404, inputting the intention characteristics into a decoder of the initial model for training so as to update the decoder and obtain an intention recognition model; the intention recognition model is used for recognizing consumption behavior data to be recognized to obtain intention information.

Illustratively, three groups of intention features output by an encoder of the intention recognition initial model, namely intention sub-features of a place subspace, intention sub-features of a time subspace and intention sub-features of an article subspace are input into a decoder of the intention recognition initial model, and the decoder is updated to obtain a trained intention recognition model. The trained intention recognition model can recognize consumption behavior data to be recognized to obtain intention information corresponding to the consumption behavior data.

Steps S401-S404 may be repeatedly performed until a preset stop condition is reached. And presetting a stopping condition, namely finishing processing the acquired consumption behavior data in the set to be trained.

In the embodiment, a set to be trained is obtained; the system comprises a to-be-trained set and a training system, wherein the to-be-trained set comprises a plurality of to-be-trained consumption behavior data, the to-be-trained consumption behavior data comprise a user and a plurality of feature information, and a binary relation exists between the user and each feature information; generating a plurality of dual hypergraph information according to a set to be trained; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user; inputting a plurality of dual hypergraph information into an encoder of an initial model for training, and generating intention characteristics of consumption behavior data to be trained; the intention features represent sub-features of the consumption behavior data to be trained on each feature information; updating the encoder according to the intention characteristics; inputting the intention characteristics into a decoder of the initial model for training so as to update the decoder and obtain an intention recognition model; the intention identification model is used for identifying consumption behavior data to be identified to obtain intention information. Obtaining an intention recognition model for recognizing consumption behavior data based on a model training mode; the intention information of the consumption behavior data can be accurately and quickly identified based on the model.

Fig. 5 is a schematic flowchart of another method for training a model applied to intent recognition according to an embodiment of the present application, and as shown in fig. 5, the method includes:

s501, acquiring a set to be trained; the set to be trained comprises a plurality of consumption behavior data to be trained, the consumption behavior data to be trained comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information.

For this step, reference may be made to the description of step S401, which is not described again.

S502, generating a plurality of dual hypergraph information according to a set to be trained; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user.

In one example, the first hypergraph has a first correlation matrix characteristic, a first node degree matrix characteristic, a first hyper-edge degree matrix characteristic, and a first node characteristic.

The first incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the first hypergraph belongs; the first node degree matrix characteristic represents the connection relation of user nodes in the first hypergraph; the first hyper-edge matrix characteristic represents the node corresponding relation of edges in the first hyper-graph; the first node characteristics represent node information of user nodes in the first hypergraph; the second hypergraph has a second incidence matrix characteristic, a second node degree matrix characteristic, a second hyper-edge degree matrix characteristic and a second node characteristic; the second incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the second hypergraph belongs; the second node degree matrix characteristic represents the connection relation of characteristic information nodes in the second hypergraph; the second hyperedge matrix characteristic represents the node corresponding relation of edges in the second hypergraph; the second node characteristics characterize node information of the characteristic information nodes in the second hypergraph.

Illustratively, a plurality of dual hypergraph information is generated based on a plurality of consumption behavior data in the set to be trained, wherein each dual hypergraph information comprises a first hypergraph of the incidence relation of the user information on the characteristic information and a second hypergraph of the incidence relation of the characteristic information on the user information. When the dual hypergraph information is generated, each hypergraph in the dual hypergraph information has an incidence matrix characteristic H, a node degree matrix characteristic D, a hyperedge degree matrix B and a node characteristic X; the first hypergraph has a first incidence matrix characteristic H ₁ First node degree matrix characteristic D ₁ First overcritical matrix feature B ₁ And a first node characteristic X ₁ (ii) a The second hypergraph has a second correlation matrix characteristic H ₂ Second node degree matrix characteristic D ₂ Second super-edge matrix characteristic B ₂ And a second node characteristic X ₂ 。

S503, inputting the information of the plurality of dual hypergraph into the encoder, and generating convolution characteristics corresponding to each hypergraph convolution layer in the encoder.

In one example, step S503 includes the following process:

the first step of step S503: and inputting a plurality of dual hypergraph information into the encoder, and generating a first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and partial feature in the second hypergraph.

The second step of step S503: and generating a second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph.

The third step of step S503: and processing based on each hypergraph convolutional layer according to the first initial embedding characteristic and the second initial embedding characteristic corresponding to the characteristic information to obtain a convolutional characteristic corresponding to each hypergraph convolutional layer.

In one example, the first step of step S503 includes the following processes: determining the propagation characteristics in the first graph corresponding to the characteristic information according to the first incidence matrix characteristics, the first node degree matrix characteristics, the first super-edge degree matrix characteristics and the first node characteristics of the first hypergraph corresponding to the characteristic information; wherein, the propagation characteristics in the first graph represent the characteristic expression of the user on the characteristic information; determining the first inter-graph propagation characteristics corresponding to the characteristic information according to the second incidence matrix characteristics, the second super-edge matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; the first inter-graph propagation characteristics represent characteristic expressions between users and characteristic information; and determining a first initial embedding feature corresponding to the feature information according to the first intra-map propagation feature corresponding to the feature information and the first inter-map propagation feature corresponding to the feature information.

In one example, the second step of step S503 includes the following processes: determining the propagation characteristics in the second graph corresponding to the characteristic information according to the second incidence matrix characteristics, the second node degree matrix characteristics, the second super-edge degree matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; the feature expression of the feature characterization feature information on the user is propagated in the second graph; determining a second inter-graph propagation characteristic corresponding to the characteristic information according to the first incidence matrix characteristic, the first hyper-edge matrix characteristic and the first node characteristic of the first hyper-graph corresponding to the characteristic information; the second inter-graph propagation characteristics represent characteristic expression between the user and the characteristic information; and determining a second initial embedding feature corresponding to the feature information according to the second intra-map propagation feature corresponding to the feature information and the second inter-map propagation feature corresponding to the feature information.

Illustratively, the generated three sets of dual hypergraph information are input into an encoder of the intention recognition model, and a hypergraph convolution layer in the encoder processes each set of dual hypergraph information respectively to generate convolution characteristics corresponding to each set of dual hypergraph information. And each supergraph convolutional layer processes the even supergraph information of each group, including an intra-graph propagation process and an inter-graph propagation process, and generates an intra-graph propagation characteristic and an inter-graph propagation characteristic. The method comprises the steps that an intra-graph propagation process of a first hypergraph generates a first intra-graph propagation characteristic by utilizing an incidence matrix characteristic, a node degree matrix characteristic, a super-edge degree matrix characteristic and a node characteristic of the first hypergraph, an inter-graph propagation process of the first hypergraph generates a first inter-graph propagation characteristic by utilizing an incidence matrix characteristic, a super-edge degree matrix characteristic and a node characteristic of a second hypergraph, and a first initial embedding characteristic is generated according to the generated first intra-graph propagation characteristic and the first inter-graph propagation characteristic; and in the intra-graph propagation process of the second hypergraph, generating a second intra-graph propagation characteristic by utilizing the incidence matrix characteristic, the node degree matrix characteristic, the super-edge degree matrix characteristic and the node characteristic of the second hypergraph, in the inter-graph propagation process, generating a second inter-graph propagation characteristic by utilizing the incidence matrix characteristic, the super-edge degree matrix characteristic and the node characteristic of the first hypergraph, and generating a second initial embedded characteristic according to the generated second intra-graph propagation characteristic and the second inter-graph propagation characteristic.

For a specific process of inputting a plurality of dual hypergraph information into the encoder to generate convolution characteristics corresponding to each hypergraph convolution layer in the encoder, reference may be made to the description of S303, which is not described again.

And S504, aggregating the volume features to obtain an aggregated feature.

The specific process of "aggregating each convolution feature to obtain an aggregated feature" may refer to the introduction of S304, and is not described again.

And S505, determining the intention characteristics of the consumption behavior data to be trained according to the aggregation characteristics.

In one example, step S505 includes the following process: and processing the aggregation features based on a multilayer perception mode to generate intention features of consumption behavior data to be trained.

Illustratively, the aggregation features generated by aggregating the three sets of dual hypergraph information are input into a multilayer perceptron to be processed, and the intention features are generated.

The specific process of "determining the intention characteristic of the consumption behavior data to be trained according to the aggregated characteristic" may be referred to the introduction of S305, and is not described in detail.

And S506, updating the encoder according to the intention characteristics.

In one example, the consumption behavior data to be trained has actual behavior information, and the actual behavior information represents whether the user purchases an item in the consumption behavior data. Step S506 includes the following processes:

the first step of step S506: and determining an independent loss function according to each pair of sub-features corresponding to the consumption behavior data to be trained and the total number of the consumption behavior data to be trained in the set to be trained.

The second step of step S506: and determining the predicted behavior information of the consumption behavior data to be trained according to the sub-characteristics corresponding to the consumption behavior data to be trained, and determining the BPR loss function according to the predicted behavior information and the actual behavior information of the consumption behavior data to be trained.

The third step of step S506: and determining an overall loss function according to the BPR loss function, the independent loss function and the preset hyper-parameter, and updating the decoder according to the overall loss function.

In one example, the set to be trained includes a first set and a second set; the consumption behavior data to be trained in the first set has actual intention information, and the consumption behavior data to be trained in the second set does not have actual intention information.

Exemplarily, based on the number of consumption behavior data to be trained and each pair of intent sub-features, determining an independent loss function; determining a BPR loss function based on the actual behavior information and the predicted behavior information; and determining an overall loss function according to the BPR loss function, the independent loss function and the preset hyper-parameter, and updating the decoder according to the overall loss function.

The consumption behavior data to be identified in the set to be trained are divided into two sets, wherein one set is a first set formed by M consumption behavior data marked with intention information, and the other set is a second set formed by N ' consumption behavior data not marked with intention information, and M ' and N ' are positive integers greater than or equal to 1. The encoder for recognizing the initial model processes the M '+ N' consumption behavior data to be trained, and respectively generates M '+ N' intention characteristics of place information, M '+ N' intention characteristics of time information and M '+ N' intention characteristics of item information. Determining an independent loss function based on the number of consumption behavior data to be trained and corresponding intention characteristics, wherein the specific process is as follows:

wherein the content of the first and second substances,

an independence loss to ensure independence of any two intended sub-features; m'Representing the amount of consumption behavior data labeled with intention information; n' represents the quantity of consumption behavior data which are not marked with intention information; m '+ N' represents the total number of consumption behavior data to be trained in the set to be trained; s = { L, T, C }, representing a set of intent sub-features of a location subspace, intent sub-features of a time subspace, intent sub-features of an item subspace; s represents an intention sub-feature in S, S 'represents another intention sub-feature in S, e.g., if S is an intention sub-feature of a location sub-space, S' is an intention sub-feature of a temporal sub-space, or an intention sub-feature of an item sub-space; dCov (·) denotes distance covariance; dVar (·) represents the distance variance;

represents an intention sub-feature corresponding to the b-th consumption behavior data to be trained, wherein b is a positive integer greater than or equal to 1 and less than or equal to M' + N>

Representing another intention sub-feature corresponding to the b-th consumption behavior data to be trained, e.g. if

Is an intended sub-feature of the location subspace, then +>

Is an intended sub-feature of a temporal sub-space, or, alternatively, an intended sub-feature of an item sub-space.

And respectively predicting the user behavior for the three intention sub-features corresponding to each consumption behavior data in the set to be trained, wherein the specific process is as follows:

wherein the content of the first and second substances,

a probability score representing the occurrence of the user consumption behavior; x is a radical of a fluorine atom _b Representing the consumption behavior data of the b-th to-be-trained object, wherein b is a positive integer which is greater than or equal to 1 and less than or equal to M '+ N'; s = { L, T, C }, representing a set of intent sub-features of a location subspace, intent sub-features of a time subspace, intent sub-features of an item subspace; s represents an intention sub-feature in S, which is an intention sub-feature of a place sub-space, or an intention sub-feature of a time sub-space, or an intention sub-feature of an item sub-space; ρ represents a predictor; />

Representing the intended feature.

The probability score of the occurrence of the user consumption behavior is the sum of the probability score of the place information, the probability score of the time information and the probability score of the article information.

The consumption behavior data in the set to be trained are all actual behavior information obtained from the platform, the predicted behavior information can be used when the BPR loss function is determined, the predicted behavior information is obtained through sampling, three dimensions in the actual behavior information are fixed, and a sample obtained through changing one dimension is the predicted behavior information. For example, for x ₁ ＝(u ₁ ，l ₁ ，t ₁ ，c ₁ ) Represents user u ₁ At a location l ₁ Time t ₁ Down purchased article c ₁ Actual behavior information; then the sample is generated at this time

User u ₁ At location l ₁ Time t ₁ Down purchased article c ₂ Not occurring, i.e. predicted behavior information. The specific calculation process of the BPR loss function is as follows:

wherein:

loss is that actual user consumption behavior can get a higher prediction score than behavior that the user does not consume; x is the number of _b Representing actual behavior information obtained from the platform; />

Representing predicted behavior information; />

Representing a set of actual behavior information and predicted behavior information.

Determining an overall PRE loss function based on the BPR loss function, the independent loss function and a preset hyper-parameter, wherein the specific process is as follows:

wherein λ is a preset hyper-parameter.

And updating the encoder intending to identify the initial model by utilizing the PRE loss function to obtain the trained encoder.

And S507, clustering the intention characteristics of the consumption behavior data to be trained to obtain a plurality of intention information.

Exemplarily, processing the intention characteristics of the consumption behavior data to be trained in the set to be trained by utilizing K-means clustering, and estimating and obtaining all intention information categories corresponding to all consumption behavior data.

The encoder for recognizing the initial model processes the M '+ N' consumption behavior data to be trained in the set to be trained, and generates the intention sub-features of M '+ N' location subspaces, the intention sub-features in M '+ N' temporal subspaces, and the intention sub-features in M '+ N' commodity subspaces, respectively. Respectively processing 3 (M ' + N ') intention features by utilizing K-means clustering, taking f times of known intention K ' as the number of all intention information, wherein f is a positive integer greater than or equal to 1, clustering the 3 (M ' + N ') intention features into f multiplied by K' in clusters. After clustering, the number of abandoned intention features is less than

And S508, decoding the intention characteristics corresponding to the consumption behavior data to be trained in the first set according to the decoder of the initial model to obtain a cross entropy loss function.

Exemplarily, the decoder for recognizing the initial model processes the intention features corresponding to the consumption behavior data to be trained, which are labeled with intention information in the set to be trained, so as to classify the intention features into K estimated intention information, and obtain a cross entropy loss function.

S509, according to the decoder of the initial model, decoding the intention features corresponding to the consumption behavior data to be trained in the second set, so as to classify the consumption behavior data to be trained in the second set into a plurality of obtained intention information, and obtain a two-classification cross entropy optimization loss function.

In one example, step S509 includes the following process:

first step of step S509: according to the decoder of the initial model, decoding the intention characteristics corresponding to the consumption behavior data to be trained in the second set so as to classify the consumption behavior data to be trained in the second set into a plurality of obtained intention information; and for each seed feature, performing similarity calculation on the sub-features of each consumption behavior data to be trained in the second set to obtain similarity information of each consumption behavior data to be trained in the second set on each sub-feature.

Second step of step S509: and for each seed feature, determining the score information of each consumption behavior data to be trained in the second set on each sub-feature, wherein the score information represents the score of the same intention information of the consumption behavior data to be trained on the dimension of the sub-feature.

The third step of step S509: and determining a two-class cross entropy optimization loss function according to the similarity information of each to-be-trained consumption behavior data in the second set on each sub-feature and the score information of each to-be-trained consumption behavior data in the second set on each sub-feature.

Illustratively, the decoder for the intention recognition initial model processes intention features corresponding to consumption behavior data to be trained, which are not labeled with intention information, in a set to be trained, so as to classify the intention features into K estimated intention information. The consumption behavior data to be trained, which is not marked with the intention information, needs to construct a sample pair, calculate the similarity of the corresponding intention features of the sample pair, and calculate whether the corresponding intention features of the sample pair belong to the same class of scores. And determining a two-classification cross entropy optimization loss function according to the similarity of the samples to the corresponding intention characteristics and the scores of whether the samples belong to the same class.

Constructing a sample pair (x) for the consumption behavior data to be trained which is not marked with intention information in the set to be trained _a ，x _v ) Wherein x is _a Consumption behavior data to be trained, x, representing the a-th unlabeled intention information _v And (4) representing the consumption behavior data to be trained of the nth unannotated intention information, wherein a and v are both positive integers which are more than or equal to 1 and positive integers which are less than or equal to N'. Processing the obtained sample pairs based on the trained encoder to obtain intention characteristics corresponding to the a-th consumption behavior data to be trained without marked intention information

And the intention characteristic which corresponds to the consumption behavior data to be trained and is not marked with intention information>

Calculating the similarity of a sample to the corresponding intended feature, i.e. the pseudo-label, in particular, using a more robust method based on a low-pass fast Fourier transformThe process is as follows:

/>

wherein the content of the first and second substances,

representing the probability that any two tags belong to the same category; COSINE represents a COSINE similarity function; LPFFT represents fourier transform.

The pseudo label is the similarity of the intention characteristics corresponding to the two consumption behavior data to be trained in the sample pair, and is a confidence interval of [0,1 ].

And calculating corresponding intention characteristics of the consumption behavior data sample pairs to be trained

Whether the scores belong to the same category or not is determined by the following specific process:

a score representing whether the intended features in the sample pair belong to the same class; η represents a classifier that classifies consumption behavior data to be trained.

Determining a two-classification cross entropy optimization loss function according to the similarity of the corresponding intention characteristics of the sample pair and the scores of whether the samples belong to the same class, wherein the specific process comprises the following steps:

and N' represents the consumption behavior data quantity of the set to be trained without marked intention information.

And S510, updating the decoder according to the cross entropy loss function and the two-classification cross entropy optimization loss function to obtain an intention identification model.

Illustratively, the decoder of the intention recognition initial model is updated according to a cross entropy loss function obtained by marking intention information and a two-class cross entropy optimization loss function obtained by not marking intention information, so as to obtain a trained decoder, and thus obtain a trained intention recognition model.

Steps S501-S510 may be repeatedly performed until a preset stop condition is reached. The preset stopping condition is that the acquired consumption behavior data in the set to be trained are processed.

For example, fig. 6 is a schematic diagram of an intention recognition model provided in an embodiment of the present application, and as shown in fig. 6, the intention recognition model is composed of two parts, namely an encoder and a decoder. Inputting the acquired consumption behavior data set into an encoder, and generating dual hypergraph information under user information-place information, dual hypergraph information under user information-time information, and dual hypergraph information under user information-article information, wherein the mth consumption behavior data in the consumption behavior data set is x _m Including user information, location information, time information, and article information: x is the number of _m ＝(u _m ，l _m ，t _m ，c _m )，u _m User information for the mth consumption behavior data, l _m Location information for the mth consumption behavior data, t _m Time information for the mth consumption behavior data, c _m The item information is item information of the mth consumption behavior data, and m is a positive integer greater than or equal to 1. The dual hypergraph information under the user information-location information includes a first hypergraph corresponding to the location information

And a second hypergraph G corresponding to the location information _L The dual hypergraph information under the user information-time information includes a first hypergraph ≥ corresponding to the time information>

And the first corresponding to the time informationTwo super graph G _T The dual hypergraph information under the user information-item information comprises a first hypergraph->

And a second hypergraph G corresponding to the article information _C . The three groups of dual hypergraph information are respectively processed by a hypergraph convolutional layer (Joint-HGC), and the intention sub-characteristic ^ is greater than or equal to the intention sub-characteristic in the dual hypergraph information generation place subspace under the user information-place information>

Intent sub-feature in dual hypergraph information generation temporal subspace under user information-temporal information->

Intent sub-feature in user information-item information duality hypergraph information Generation item subspace->

Determining an independent loss based on the generated intent sub-feature>

And BPR loss->

Based on independent loss->

And BPR loss>

The encoder of the model is trained. And estimating the intention information types of all consumption behavior data by utilizing K-means clustering; and, the intention sub-features in the location subspace generated by the encoder are combined

Intents in a temporal subspaceSub-characteristic->

Intent sub-feature in item subspace>

Inputting the information into a decoder, processing the information by the decoder to generate probability distribution condition of intention information in a plurality of intention information on the place information>

Probability distribution of intention information in a plurality of intention information over time information->

And a probability distribution of the intention information among the plurality of intention information on the item information->

Probability distribution based on intention information in a plurality of intention information on location information>

Determining a maximum probability p of a plurality of intent information _m Corresponding intention information, and further, intention information of the mth consumption behavior data is determined.

In the embodiment, a set to be trained is obtained; the method comprises the steps that a set to be trained comprises a plurality of consumption behavior data to be trained, the consumption behavior data to be trained comprise a user and a plurality of feature information, and a binary relation exists between the user and each feature information; generating a plurality of dual hypergraph information according to a set to be trained; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user; inputting a plurality of dual hypergraph information into an encoder of an initial model for training, and generating intention characteristics of consumption behavior data to be trained; the intention features represent sub-features of the consumption behavior data to be trained on each feature information; updating the encoder according to the intention characteristics; inputting the intention characteristics into a decoder of the initial model for training so as to update the decoder to obtain an intention recognition model; the intention identification model is used for identifying consumption behavior data to be identified to obtain intention information. Obtaining an intention recognition model for recognizing consumption behavior data based on a model training mode; the intention information of the consumption behavior data can be accurately and quickly identified based on the model. And the intention decoder is trained in a mode of constructing a pseudo label for the intention characteristics of the sample pairs without the intention, so that the classification effect of the intention recognition model can be improved, and the accuracy and precision of the model are further improved.

Fig. 7 is a schematic structural diagram of an identification apparatus based on intention identification according to an embodiment of the present application, and as shown in fig. 7, the apparatus includes:

an obtaining unit 71, configured to obtain a set to be identified; the set to be identified comprises a plurality of consumption behavior data to be identified, the consumption behavior data to be identified comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information.

A processing unit 72, configured to generate a plurality of dual hypergraph information according to the set to be identified; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user.

A first determination unit 73, configured to input a plurality of dual hypergraph information into an encoder of the intention recognition model, and generate an intention feature of consumption behavior data to be recognized; wherein the intention characteristics characterize the sub-characteristics of the consumption behavior data to be recognized on each characteristic information.

A second determining unit 74, configured to input the intention characteristics into a decoder of the intention recognition model, and obtain intention information of the consumption behavior data to be recognized.

In an example, the first determining unit 73 is specifically configured to: inputting a plurality of dual hypergraph information into an encoder to generate convolution characteristics corresponding to each hypergraph convolution layer in the encoder; polymerizing each convolution characteristic to obtain a polymerization characteristic; and determining the intention characteristics of the consumption behavior data to be identified according to the aggregation characteristics.

In one example, the first hypergraph has a first correlation matrix characteristic, a first node degree matrix characteristic, a first hyper-edge matrix characteristic, and a first node characteristic; the first incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the first hypergraph belongs; the first node degree matrix characteristic represents the connection relation of user nodes in the first hypergraph; the first hyper-edge matrix characteristic represents the node corresponding relation of edges in the first hyper-graph; the first node characteristics represent node information of user nodes in the first hypergraph; the second hypergraph has a second incidence matrix characteristic, a second node degree matrix characteristic, a second hyper-edge degree matrix characteristic and a second node characteristic; the second incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the second hypergraph belongs; the second node degree matrix characteristic represents the connection relation of characteristic information nodes in the second hypergraph; the second hyperedge matrix characteristic represents the node corresponding relation of edges in the second hypergraph; the second node characteristics characterize node information of the characteristic information nodes in the second hypergraph.

In one example, when the first determining unit 73 inputs a plurality of dual hypergraph information into the encoder to generate convolution characteristics corresponding to each hypergraph convolution layer in the encoder, it is specifically configured to: inputting a plurality of dual hypergraph information into an encoder, and generating a first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and the partial feature of the second hypergraph; generating a second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph; and processing based on each hypergraph convolutional layer according to the first initial embedding characteristic and the second initial embedding characteristic corresponding to the characteristic information to obtain a convolutional characteristic corresponding to each hypergraph convolutional layer.

In one example, when the first determining unit 73 generates the first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and the partial feature of the second hypergraph, specifically, it is configured to: determining the propagation characteristics in the first graph corresponding to the characteristic information according to the first incidence matrix characteristics, the first node degree matrix characteristics, the first super-edge degree matrix characteristics and the first node characteristics of the first hypergraph corresponding to the characteristic information; wherein, the propagation characteristics in the first graph represent the characteristic expression of the user on the characteristic information; determining the first inter-graph propagation characteristics corresponding to the characteristic information according to the second incidence matrix characteristics, the second super-edge matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; the first inter-graph propagation characteristics represent characteristic expressions between users and characteristic information; and determining a first initial embedding feature corresponding to the feature information according to the first intra-map propagation feature corresponding to the feature information and the first inter-map propagation feature corresponding to the feature information.

In one example, when the first determining unit 73 generates the second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph, specifically, it is configured to: determining the propagation characteristics in the second graph corresponding to the characteristic information according to the second incidence matrix characteristics, the second node degree matrix characteristics, the second super-edge degree matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; the feature expression of the feature characterization feature information on the user is propagated in the second graph; determining a second inter-graph propagation characteristic corresponding to the characteristic information according to the first incidence matrix characteristic, the first hyper-edge matrix characteristic and the first node characteristic of the first hyper-graph corresponding to the characteristic information; the second inter-graph propagation characteristics represent characteristic expressions between the user and the characteristic information; and determining a second initial embedding feature corresponding to the feature information according to the second intra-map propagation feature corresponding to the feature information and the second inter-map propagation feature corresponding to the feature information.

In one example, the first determining unit 73, when determining the intention characteristic of the consumption behavior data to be identified according to the aggregation characteristic, is specifically configured to: and processing the aggregation characteristics based on a multilayer perception mode to generate the intention characteristics of the consumption behavior data to be identified.

In one example, the to-be-identified sets acquired by the acquisition unit 71 include a first set and a second set; the consumption behavior data to be identified in the first set has actual intention information, and the consumption behavior data to be identified in the second set does not have actual intention information.

In one example, the to-be-identified sets acquired by the acquisition unit 71 include a first set and a second set; the consumption behavior data to be identified in the first set have actual intention information, and the consumption behavior data to be identified in the second set do not have actual intention information; the second determining unit 74, when inputting the intention characteristics into the decoder of the intention recognition model and obtaining the intention information of the consumption behavior data to be recognized, is specifically configured to: clustering the intention characteristics of the consumption behavior data to be identified to obtain a plurality of intention information; and according to the decoder of the intention identification model, decoding the intention characteristics corresponding to the consumption behavior data to be identified so as to classify the consumption behavior data to be identified into a plurality of pieces of intention information to obtain the intention information of the consumption behavior data to be identified.

In an example, when the second determining unit 74 performs decoding processing on the intention feature corresponding to the consumption behavior data to be recognized according to the decoder of the intention recognition model to classify the consumption behavior data to be recognized into the obtained plurality of intention information, so as to obtain the intention information of the consumption behavior data to be recognized, specifically, the second determining unit is configured to: according to the decoder of the intention identification model, decoding the intention characteristics corresponding to the consumption behavior data to be identified so as to classify the consumption behavior data to be identified into a plurality of acquired intention information and acquire probability distribution information of the consumption behavior data to be identified on each characteristic information; the probability distribution information represents the probability distribution condition of the intention information belonging to the consumption behavior data to be identified in the plurality of intention information; determining the weight information of the consumption behavior data to be identified on each characteristic information according to the sub-characteristics of the consumption behavior data to be identified on each characteristic information; determining the initial intention of the consumption behavior data to be identified on each characteristic information according to the probability distribution information and the weight information of the consumption behavior data to be identified on each characteristic information; and determining intention information of the consumption behavior data to be identified according to each initial intention corresponding to the consumption behavior data to be identified.

In one example, the identifying means further comprises an applying unit 75 for recommending the item to the user according to the intention information of the consumption behavior data to be identified.

Fig. 8 is a schematic structural diagram of a model training apparatus applied to intent recognition according to an embodiment of the present application, and as shown in fig. 8, the apparatus includes:

an obtaining unit 81, configured to obtain a set to be trained; the to-be-trained set comprises a plurality of to-be-trained consumption behavior data, the to-be-trained consumption behavior data comprise a user and a plurality of feature information, and a binary relation exists between the user and each feature information.

The processing unit 82 is used for generating a plurality of dual hypergraph information according to the set to be trained; the dual hypergraph information comprises a first hypergraph corresponding to each feature information and a second hypergraph corresponding to each feature information, the first hypergraph represents the incidence relation of the user on the feature information, and the second hypergraph represents the incidence relation of the feature information on the user.

The first updating unit 83 is used for inputting a plurality of dual hypergraph information into an encoder of the initial model for training, and generating the intention characteristics of consumption behavior data to be trained; the intention features represent sub-features of the consumption behavior data to be trained on each feature information; and updates the encoder according to the intent characteristics.

A second updating unit 84, configured to input the intention characteristics into the decoder of the initial model for training, so as to update the decoder, and obtain an intention recognition model; the intention recognition model is used for recognizing consumption behavior data to be recognized to obtain intention information.

In one example, the first updating unit 83 is specifically configured to: inputting a plurality of dual hypergraph information into an encoder to generate convolution characteristics corresponding to each hypergraph convolution layer in the encoder; polymerizing each convolution characteristic to obtain a polymerization characteristic; and determining the intention characteristics of the consumption behavior data to be trained according to the aggregation characteristics.

In one example, when the first updating unit 83 inputs a plurality of dual hypergraph information into the encoder to generate convolution characteristics corresponding to each hypergraph convolution layer in the encoder, it is specifically configured to: inputting a plurality of dual hypergraph information into an encoder, and generating a first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and the partial feature of the second hypergraph; generating a second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph; and processing based on each hypergraph convolutional layer according to the first initial embedding characteristic and the second initial embedding characteristic corresponding to the characteristic information to obtain a convolutional characteristic corresponding to each hypergraph convolutional layer.

In one example, when the first updating unit 83 generates the first initial embedded feature corresponding to the feature information according to the feature of the first hypergraph corresponding to the feature information and the partial feature of the second hypergraph, specifically: determining the propagation characteristics in the first graph corresponding to the characteristic information according to the first incidence matrix characteristics, the first node degree matrix characteristics, the first super-edge degree matrix characteristics and the first node characteristics of the first hypergraph corresponding to the characteristic information; wherein, the propagation characteristics in the first graph represent the characteristic expression of the user on the characteristic information; determining the first inter-graph propagation characteristics corresponding to the characteristic information according to the second incidence matrix characteristics, the second super-edge matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; the first inter-graph propagation characteristics represent characteristic expressions between users and characteristic information; and determining a first initial embedding feature corresponding to the feature information according to the first intra-map propagation feature corresponding to the feature information and the first inter-map propagation feature corresponding to the feature information.

In one example, when generating the second initial embedded feature corresponding to the feature information according to the feature of the second hypergraph corresponding to the feature information and the partial feature in the first hypergraph, the first updating unit 83 is specifically configured to: determining the propagation characteristics in the second graph corresponding to the characteristic information according to the second incidence matrix characteristics, the second node degree matrix characteristics, the second super-edge degree matrix characteristics and the second node characteristics of the second hypergraph corresponding to the characteristic information; the feature expression of the feature characterization feature information on the user is propagated in the second graph; determining second inter-graph propagation characteristics corresponding to the characteristic information according to the first incidence matrix characteristics, the first super-edge matrix characteristics and the first node characteristics of the first hyper-graph corresponding to the characteristic information; the second inter-graph propagation characteristics represent characteristic expressions between the user and the characteristic information; and determining a second initial embedding feature corresponding to the feature information according to the second intra-map propagation feature corresponding to the feature information and the second inter-map propagation feature corresponding to the feature information.

In one example, the first updating unit 83, when determining the intention characteristic of the consumption behavior data to be trained according to the aggregated characteristic, is specifically configured to: and processing the aggregation features based on a multilayer perception mode to generate intention features of consumption behavior data to be trained.

In one example, the consumption behavior data to be trained acquired by the acquiring unit 81 has actual behavior information, and the actual behavior information represents whether the user purchases an item in the consumption behavior data; the first updating unit 83 is specifically configured to, when updating the encoder according to the intention characteristic: determining an independent loss function according to each pair of sub-features corresponding to the consumption behavior data to be trained and the total number of the consumption behavior data to be trained of the set to be trained; determining predicted behavior information of the consumption behavior data to be trained according to the sub-characteristics corresponding to the consumption behavior data to be trained, and determining a BPR loss function according to the predicted behavior information and the actual behavior information of the consumption behavior data to be trained; and determining an overall loss function according to the BPR loss function, the independent loss function and the preset hyper-parameter, and updating a decoder according to the overall loss function.

In one example, the to-be-trained set acquired by the acquiring unit 81 includes a first set and a second set; the consumption behavior data to be trained in the first set has actual intention information, and the consumption behavior data to be trained in the second set does not have actual intention information.

In one example, the to-be-trained set acquired by the acquiring unit 81 includes a first set and a second set; the consumption behavior data to be trained in the first set have actual intention information, and the consumption behavior data to be trained in the second set do not have actual intention information; the second updating unit 84 is specifically configured to, when the intention characteristics are input into the decoder of the initial model and trained to update the decoder, and an intention recognition model is obtained: clustering the intention characteristics of the consumption behavior data to be trained to obtain a plurality of intention information; decoding the intention characteristics corresponding to the consumption behavior data to be trained in the first set according to a decoder of the initial model to obtain a cross entropy loss function; according to a decoder of the initial model, decoding the intention characteristics corresponding to the consumption behavior data to be trained in the second set so as to classify the consumption behavior data to be trained in the second set into a plurality of obtained intention information to obtain a two-class cross entropy optimization loss function; and updating the decoder according to the cross entropy loss function and the two-classification cross entropy optimization loss function to obtain an intention identification model.

In an example, when the second updating unit 84 performs decoding processing on the intention features corresponding to the consumption behavior data to be trained in the second set according to the decoder of the initial model, so as to classify the consumption behavior data to be trained in the second set into the obtained multiple intention information, and obtain the two-class cross entropy optimization loss function, specifically, the second updating unit is configured to: according to the decoder of the initial model, decoding the intention characteristics corresponding to the consumption behavior data to be trained in the second set so as to classify the consumption behavior data to be trained in the second set into a plurality of acquired intention information; and for each seed feature, performing similarity calculation on the sub-feature of each consumption behavior data to be trained in the second set to obtain similarity information of each consumption behavior data to be trained in the second set on each sub-feature; for each seed feature, determining score information of each consumption behavior data to be trained in the second set on each sub-feature, wherein the score information represents a score of the consumption behavior data to be trained belonging to the same intention information on the dimension of the sub-feature; and determining a two-class cross entropy optimization loss function according to the similarity information of each to-be-trained consumption behavior data in the second set on each sub-feature and the score information of each to-be-trained consumption behavior data in the second set on each sub-feature.

Fig. 9 is a schematic structural diagram of an electronic device according to an embodiment of the present application, and as shown in fig. 9, the electronic device includes: a memory 91, a processor 92; a memory 91; a memory for storing instructions executable by the processor 92.

Wherein the processor 92 is configured to perform the method as provided in any of the embodiments described above.

The terminal device further comprises a receiver 93 and a transmitter 94. The receiver 93 is used for receiving instructions and data transmitted from other devices, and the transmitter 94 is used for transmitting instructions and data to an external device.

Fig. 10 is a block diagram of an electronic device, which may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, an exercise device, a personal digital assistant, and the like, according to an embodiment of the present disclosure.

The apparatus 1000 may include one or more of the following components: processing component 1002, memory 1004, power component 1006, multimedia component 1008, audio component 1010, input/output (I/O) interface 1012, sensor component 1014, and communications component 1016.

The processing component 1002 generally controls the overall operation of the device 1000, such as operations associated with display, telephone calls, data communications, camera operations, and recording operations. The processing components 1002 may include one or more processors 1020 to execute instructions to perform all or a portion of the steps of the methods described above. Further, processing component 1002 may include one or more modules that facilitate interaction between processing component 1002 and other components. For example, the processing component 1002 may include a multimedia module to facilitate interaction between the multimedia component 1008 and the processing component 1002.

The memory 1004 is configured to store various types of data to support operations at the apparatus 1000. Examples of such data include instructions for any application or method operating on device 1000, contact data, phonebook data, messages, pictures, videos, and so forth. The memory 1004 may be implemented by any type or combination of volatile or non-volatile memory devices, such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disks.

The power supply component 1006 provides power to the various components of the device 1000. The power components 1006 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device 1000.

The multimedia component 1008 includes a screen that provides an output interface between the device 1000 and a user. In some embodiments, the screen may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive an input signal from a user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundary of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 1008 includes a front facing camera and/or a rear facing camera. The front camera and/or the rear camera may receive external multimedia data when the device 1000 is in an operating mode, such as a shooting mode or a video mode. Each front camera and rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.

The audio component 1010 is configured to output and/or input audio signals. For example, audio component 1010 includes a Microphone (MIC) configured to receive external audio signals when apparatus 1000 is in an operational mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signal may further be stored in the memory 1004 or transmitted via the communication component 1016. In some embodiments, audio component 1010 also includes a speaker for outputting audio signals.

I/O interface 1012 provides an interface between processing component 1002 and peripheral interface modules, which may be keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to: a home button, a volume button, a start button, and a lock button.

The sensor assembly 1014 includes one or more sensors for providing various aspects of status assessment for the device 1000. For example, sensor assembly 1014 may detect an open/closed state of device 1000, the relative positioning of components, such as a display and keypad of device 1000, sensor assembly 1014 may also detect a change in position of device 1000 or a component of device 1000, the presence or absence of user contact with device 1000, orientation or acceleration/deceleration of device 1000, and a change in temperature of device 1000. The sensor assembly 1014 may include a proximity sensor configured to detect the presence of a nearby object without any physical contact. The sensor assembly 1014 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 1014 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

The communication component 1016 is configured to facilitate communications between the apparatus 1000 and other devices in a wired or wireless manner. The device 1000 may access a wireless network based on a communication standard, such as WiFi,2G or 3G, or a combination thereof. In an exemplary embodiment, the communication component 1016 receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communications component 1016 further includes a Near Field Communication (NFC) module to facilitate short-range communications. For example, the NFC module may be implemented based on Radio Frequency Identification (RFID) technology, infrared data association (IrDA) technology, ultra Wideband (UWB) technology, bluetooth (BT) technology, and other technologies.

In an exemplary embodiment, the apparatus 1000 may be implemented by one or more Application Specific Integrated Circuits (ASICs), digital Signal Processors (DSPs), digital Signal Processing Devices (DSPDs), programmable Logic Devices (PLDs), field Programmable Gate Arrays (FPGAs), controllers, micro-controllers, microprocessors or other electronic components for performing the above-described methods.

In an exemplary embodiment, a non-transitory computer readable storage medium comprising instructions, such as the memory 1004 comprising instructions, executable by the processor 1020 of the device 1000 to perform the above-described method is also provided. For example, the non-transitory computer readable storage medium may be a ROM, a Random Access Memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like.

Embodiments of the present application also provide a non-transitory computer-readable storage medium, where instructions in the storage medium, when executed by a processor of an electronic device, enable the electronic device to perform the above-mentioned method.

The present application further provides a computer program product comprising: a computer program, stored in a readable storage medium, from which at least one processor of the electronic device can read the computer program, and the execution of the computer program by the at least one processor causes the electronic device to perform the solutions provided by any of the above embodiments.

Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the application and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the application being indicated by the following claims.

It will be understood that the present application is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims

1. An intent recognition method based on an intent recognition model, the method comprising:

2. The method of claim 1, wherein inputting the plurality of dual hypergraph information into an encoder of an intent recognition model, generating intent features for the consumption behavior data to be recognized, comprises:

aggregating the convolution characteristics to obtain aggregate characteristics; and determining the intention characteristics of the consumption behavior data to be identified according to the aggregation characteristics.

3. The method of claim 1, wherein the first hypergraph has a first correlation matrix characteristic, a first node degree matrix characteristic, a first hyper-edge matrix characteristic, and a first node characteristic;

4. The method of claim 2, wherein inputting the plurality of dual hypergraph information into the encoder, generating convolution features corresponding to each hypergraph convolution layer in the encoder, comprises:

5. The method of claim 4, wherein generating a first initial embedded feature corresponding to the feature information based on the feature of the first hypergraph corresponding to the feature information and the partial feature of the second hypergraph comprises:

6. The method of claim 4, wherein generating a second initial embedded feature corresponding to the feature information based on the feature of the second hypergraph corresponding to the feature information and the partial feature of the first hypergraph comprises:

7. The method of claim 2, wherein determining the intent characteristics of the consumption behavior data to be identified from the aggregated characteristics comprises:

8. The method of claim 1, wherein the set to be identified comprises a first set and a second set; the consumption behavior data to be identified in the first set has actual intention information, and the consumption behavior data to be identified in the second set does not have actual intention information.

9. The method according to any one of claims 1-7, wherein the set to be identified comprises a first set and a second set; the consumption behavior data to be identified in the first set have actual intention information, and the consumption behavior data to be identified in the second set do not have actual intention information; inputting the intention characteristics into a decoder of the intention recognition model to obtain intention information of the consumption behavior data to be recognized, wherein the intention information comprises:

10. The method according to claim 9, wherein decoding, according to a decoder of the intent recognition model, intent features corresponding to the consumption behavior data to be recognized to classify the consumption behavior data to be recognized into the obtained plurality of intent information, so as to obtain the intent information of the consumption behavior data to be recognized, includes:

and determining intention information of the consumption behavior data to be identified according to the initial intents corresponding to the consumption behavior data to be identified.

11. The method according to any one of claims 1-7, further comprising:

12. A method of model training for application to intent recognition, the method comprising:

acquiring a set to be trained; the to-be-trained set comprises a plurality of to-be-trained consumption behavior data, the to-be-trained consumption behavior data comprise a user and a plurality of feature information, and a binary relation exists between the user and each feature information;

inputting the dual hypergraph information into an encoder of an initial model for training, and generating intention characteristics of the consumption behavior data to be trained; the intention features represent sub-features of the consumption behavior data to be trained on each feature information; and updating the encoder according to the intention characteristics;

inputting the intention characteristics into a decoder of the initial model for training so as to update the decoder to obtain an intention recognition model; the intention recognition model is used for recognizing consumption behavior data to be recognized to obtain intention information.

13. The method of claim 12, wherein inputting the plurality of dual hypergraph information into an encoder of an initial model for training, generating intent features of the consumption behavior data to be trained, comprises:

14. The method of claim 12, wherein the first hypergraph has a first correlation matrix characteristic, a first node degree matrix characteristic, a first hyper-edge matrix characteristic, and a first node characteristic;

the second hypergraph has a second incidence matrix characteristic, a second node degree matrix characteristic, a second hyper-edge degree matrix characteristic and a second node characteristic; the second incidence matrix characteristic represents the characteristic expression of dual hypergraph information to which the second hypergraph belongs; the second node degree matrix characteristic represents the connection relation of characteristic information nodes in a second hypergraph; the second super-edge matrix characteristic represents the node corresponding relation of edges in a second super-graph; the second node features characterize node information of the feature information nodes in the second hypergraph.

15. The method of claim 13, wherein inputting the plurality of dual hypergraph information into the encoder, generating convolution features corresponding to each hypergraph convolution layer in the encoder, comprises:

16. The method of claim 15, wherein generating a first initial embedded feature corresponding to the feature information based on the feature of the first hypergraph corresponding to the feature information and the partial feature of the second hypergraph comprises:

17. The method of claim 15, wherein generating a second initial embedded feature corresponding to the feature information based on the feature of the second hypergraph corresponding to the feature information and the partial feature of the first hypergraph comprises:

18. The method of claim 13, wherein determining intent characteristics of the consumption behavior data to be trained from the aggregated characteristics comprises:

19. The method according to any one of claims 12-18, wherein the consumption behavior data to be trained has actual behavior information, the actual behavior information characterizing whether a user purchases an item in the consumption behavior data; updating the encoder according to the intent characteristics, including:

and determining a total loss function according to the BPR loss function, the independent loss function and a preset hyper-parameter, and updating the decoder according to the total loss function.

20. The method of claim 12, wherein the set to be trained comprises a first set and a second set; the consumption behavior data to be trained in the first set has actual intention information, and the consumption behavior data to be trained in the second set does not have actual intention information.

21. The method according to any one of claims 12-18, wherein the set to be trained comprises a first set and a second set; the consumption behavior data to be trained in the first set have actual intention information, and the consumption behavior data to be trained in the second set do not have actual intention information; inputting the intention characteristics into a decoder of the initial model for training so as to update the decoder to obtain an intention recognition model, wherein the intention recognition model comprises the following steps:

22. The method of claim 21, wherein decoding, according to a decoder of the initial model, the intention features corresponding to the consumption behavior data to be trained in the second set to classify the consumption behavior data to be trained in the second set into the obtained plurality of intention information, and obtaining a two-class cross-entropy optimization loss function comprises:

23. An identification apparatus of intention recognition, characterized in that the apparatus comprises:

a second determining unit, configured to input the intention feature into a decoder of the intention recognition model, so as to obtain intention information of the consumption behavior data to be recognized;

24. A model training apparatus for use in intent recognition, the apparatus comprising:

the acquisition unit is used for acquiring a set to be trained; the to-be-trained set comprises a plurality of to-be-trained consumption behavior data, the to-be-trained consumption behavior data comprises a user and a plurality of feature information, and a binary relation exists between the user and each feature information;

the first updating unit is used for inputting the dual hypergraph information into an encoder of an initial model for training and generating the intention characteristics of the consumption behavior data to be trained; wherein the intention features characterize sub-features of the consumption behavior data to be trained on each feature information; and updating the encoder according to the intention characteristics;

25. An electronic device, comprising: a processor, and a memory communicatively coupled to the processor;

the memory stores computer-executable instructions;

the processor executes computer-executable instructions stored by the memory to implement the method of any one of claims 1-11 or to implement the method of any one of claims 12-22.

26. A computer-readable storage medium having stored therein computer-executable instructions for performing the method of any one of claims 1-11 when executed by a processor, or for performing the method of any one of claims 12-22 when executed by a processor.

27. A computer program product, characterized in that it comprises a computer program which, when being executed by a processor, carries out the method of any one of claims 1-22.