CN115470409A

CN115470409A - Interest recommendation method, apparatus, electronic device, medium, and program product

Info

Publication number: CN115470409A
Application number: CN202211154999.6A
Authority: CN
Inventors: 王雅楠
Original assignee: Industrial and Commercial Bank of China Ltd ICBC; ICBC Technology Co Ltd
Current assignee: Industrial and Commercial Bank of China Ltd ICBC; ICBC Technology Co Ltd
Priority date: 2022-09-21
Filing date: 2022-09-21
Publication date: 2022-12-13

Abstract

The application provides an interest recommendation method, an interest recommendation device, electronic equipment, a medium and a program product, which relate to the field of financial science and technology, and the method comprises the following steps: acquiring historical behavior data of a target user in a preset time period, and acquiring preference propagation information of the target user based on the historical behavior data; inputting the preference propagation information and the preset item feature vector into an attention mechanism neural network model; acquiring a user characteristic vector corresponding to the preference propagation information based on the attention mechanism neural network model, and splicing the user characteristic vector and a preset item characteristic vector to obtain a spliced characteristic vector; and recommending related interest items to the target user in response to the prediction result of the spliced feature vector. According to the method and the device, the user preference propagation information and the events are input into the attention neural network for deep learning, so that the accuracy of event recommendation of a government affair platform can be effectively improved, and a recommendation result with higher reference value is provided for the user.

Description

Interest recommendation method, device, electronic equipment, medium and program product

Technical Field

The present application relates to the field of financial technology, and in particular, to an interest recommendation method, apparatus, electronic device, medium, and program product.

Background

With the continuous development of internet technology, government affair platforms provide a service mode of one internet access, that is, users can handle matters in different fields only by logging in the platforms.

The current government affair platform can cover thousands of items, and the government affair platform also contains items such as living service, recruitment examination and the like, so that great convenience is brought to users, but corresponding problems are brought, for example, the information demands of the users are various and have large differences, and each government affair information is independent, multi-source and heterogeneous, so that the difficulty of transacting affairs by the users through the platform is undoubtedly increased.

In the related technology, a coordination filtering algorithm is adopted for carrying out personalized recommendation on a user, the work difficulty of the user on a government affair platform is reduced to a certain extent, but due to the fact that the coordination filtering algorithm is sparse in user data and limited in cold start processing capacity of a recommendation system, the expected effect of a recommendation scene cannot be achieved, the actual intention of the user cannot be reflected sometimes in a recommendation result, and the recommendation accuracy is low.

Disclosure of Invention

In order to solve the problems, namely the problem of low accuracy of the event recommendation of the government affair platform, the application provides an interest recommendation method, device, electronic equipment, medium and program product.

In order to achieve the above object, the present application provides the following technical solutions:

according to an aspect of the present application, there is provided an interest recommendation method including:

acquiring historical behavior data of a target user in a preset time period, and acquiring preference propagation information of the target user based on the historical behavior data;

inputting the preference propagation information and the preset item feature vector into an attention mechanism neural network model;

acquiring a user characteristic vector corresponding to the preference propagation information based on the attention mechanism neural network model, and splicing the user characteristic vector and a preset item characteristic vector to obtain a spliced characteristic vector;

and recommending related interest items to the target user in response to the prediction result of the spliced feature vector.

In one embodiment, the obtaining of the preference propagation information of the target user based on the historical behavior data includes:

converting the historical behavior data into a knowledge triple form, and establishing an item knowledge graph based on the historical behavior data subjected to data conversion;

and acquiring the preference propagation information of the target user based on the item knowledge graph.

In one embodiment, the converting the historical behavior data into a knowledge triple form includes:

acquiring an item category dictionary table and an item attribute relation table, wherein the item category dictionary table carries item IDs (identities) and belonged categories, and the item attribute relation table carries attribute relations among the items;

acquiring the interaction times of the target user and each item in the item category dictionary table based on the historical behavior data;

and acquiring historical behavior data corresponding to the knowledge triple form based on the interaction times of the target user and each item in the item category dictionary table and the attribute relationship.

In one embodiment, obtaining the preference propagation information of the target user based on the item knowledge graph comprises:

taking the feature set interested by the target user as a seed in the event knowledge graph, extending along the links of the event knowledge graph to form a plurality of ripple groups, and searching all sets of vertices with the shortest path of which is not more than a preset number based on the ripple groups to serve as knowledge triple sets;

and acquiring the preference propagation information of the target user based on the knowledge triple set.

In one embodiment, obtaining the user feature vector corresponding to the preference propagation information based on the attention mechanism neural network model includes:

iteratively embedding the preference propagation information and the preset item feature vector into an attention neural network model, and acquiring a user feature vector based on the knowledge triple set and the preset item feature vector in the neural network model.

In one embodiment, the obtaining, in the neural network model, a user feature vector based on the knowledge triples and the event feature vector comprises:

and acquiring the correlation probability of the knowledge triple set and the preset item feature vector in the neural network model, and acquiring the user feature vector based on the correlation probability.

In one embodiment, the obtaining, in the neural network model, the probability of the knowledge triple set being related to the default event feature vector includes:

and acquiring the related probability of the knowledge triplet set and the preset item feature vector in the neural network model based on the attention weight between the node corresponding to the knowledge triplet of each attention network layer and the neighborhood node thereof, the weight matrix of each attention network layer and a preset activation function.

In one embodiment, the obtaining the user feature vector based on the correlation probability includes:

and acquiring user feature vectors of the users in all orders corresponding to the plurality of corrugated groups based on the correlation probability, and acquiring final user feature vectors based on the user feature vectors of all orders.

According to another aspect of the present application, there is provided an interest recommendation apparatus including:

the data acquisition module is set to acquire historical behavior data of a target user within a preset time period;

a preference acquisition module configured to acquire preference propagation information of the target user based on the historical behavior data;

an input module configured to input the preference propagation information and the pre-defined event feature vector into an attention mechanism neural network model;

the machine learning module is configured to acquire a user feature vector corresponding to the preference propagation information based on the attention mechanism neural network model, and splice the user feature vector and a preset item feature vector to obtain a spliced feature vector;

and the interest recommending module is used for recommending related interest items to the target user in response to the prediction result of the spliced feature vector.

According to yet another aspect of the present application, there is provided an electronic device including: a memory and a processor;

the memory stores computer execution instructions;

the processor executes computer-executable instructions stored in the memory to cause the electronic device to perform the interest recommendation method.

According to still another aspect of the present application, there is provided a computer-readable storage medium having stored therein computer-executable instructions for implementing the interest recommendation method when executed by a processor.

According to yet another aspect of the present application, there is provided a computer program product comprising computer program code which, when run on a computer, causes the computer to perform the interest recommendation method.

It can be understood that, according to the interest recommendation method, apparatus, electronic device, medium, and program product provided in the embodiments of the present application, historical behavior data of a target user within a preset time period is obtained, preference propagation information of the target user is obtained based on the historical behavior data, then the preference propagation information and a preset item feature vector are input into a attention mechanism neural network model, a user feature vector corresponding to the preference propagation information is obtained based on the attention mechanism neural network model, the user feature vector and the preset item feature vector are spliced to obtain a spliced feature vector, and a prediction result of the spliced feature vector is responded to, and a relevant interest item is recommended to the target user. By the method, the incidence relation between the preference propagation information and the items of the user is considered, the preference propagation information and the items of the user are input into the attention neural network for deep learning, robustness is enhanced, the item recommendation result is more accurate, and the actual requirements of the user can be met when interest recommendation is carried out.

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 diagram of a possible scenario provided in an embodiment of the present application;

fig. 2 is a schematic flowchart of an interest recommendation method according to an embodiment of the present application;

FIG. 3 is an exemplary diagram of a knowledge triplet in an embodiment of the present application;

FIG. 4 is an exemplary diagram of a transaction knowledge graph in an embodiment of the present application;

FIG. 5a is an exemplary diagram of step S203 in the embodiment of the present application;

FIG. 5b is an exemplary diagram of an embedded embedding vector representation of item v in the embodiment of the present application;

FIG. 5c is a diagram illustrating an example of computing node attention weights in an embodiment of the present application;

FIG. 5d is a graph of the Leaky ReLU function in the example of the present application;

FIG. 5e is an exemplary diagram of a multi-layer attention network layer representation in an embodiment of the present application;

FIG. 6 is a diagram illustrating an example of an output data format of a prediction result in an embodiment of the present application;

fig. 7 is a schematic structural diagram of an interest recommendation apparatus according to an embodiment of the present application;

fig. 8 is a schematic structural 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

In the related technology, a coordinated Filtering algorithm (CF for short) is adopted, and potential common hobbies of a user are screened according to historical interaction data of the user to recommend interests of the user, so that the problem that the difficulty of handling the user on a government platform is high is solved. However, the coordination filtering algorithm has limited processing capacity for user data sparseness and cold start of a recommendation system, and cannot achieve the expected effect of a recommendation scene, and sometimes, the recommendation result cannot reflect the actual intention of the user, so that the recommendation accuracy cannot meet the requirement, the item recommendation efficiency is not high, and sometimes, even a certain confusion is brought to the user.

In view of this, embodiments of the present application provide an interest recommendation method, apparatus, electronic device, medium, and program product, which fuse item preference information of a user on the basis of collaborative filtering, and may perform deep learning on user characteristics in a neural network based on attention mechanism by combining with a Knowledge Graph (KG), so as to output a result, thereby completing interest recommendation. In the process, the attention mechanism neural network is adopted for deep learning, neighbor aggregation can be better realized, the robustness of the model is improved, and the problems of sparse user data and limited cold start processing capacity of the recommendation system are solved. In addition, the item preference information of the user can obtain the potential preference of the user by taking the propagation of water flow fluctuation (Ripple) as reference, so that the recommendation result is more accurate.

In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be described in more detail below with reference to the accompanying drawings in the embodiments of the present application. In the drawings, the same or similar reference numerals denote the same or similar components or components having the same or similar functions throughout. The described embodiments are a subset of the embodiments in the present application and not all embodiments in the present application. The embodiments described below with reference to the drawings are exemplary and intended to be used for explaining the present application and should not be construed as limiting the present application. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

Fig. 1 is a schematic view of a possible scenario provided by the embodiment of the present application, and it should be noted that the interest recommendation method disclosed in the present application may be used in the field of financial technology, and may also be used in any other field.

Taking a matter recommendation scenario of a government affair platform as an example, as shown in fig. 1, the matter recommendation scenario includes a terminal device 110 and a server 120, and the terminal device 110 and the server 120 are connected to each other through a wired or wireless network. In some embodiments, the terminal device 110 is configured to provide historical behavior data of the target user within a preset time period and a trained attention mechanism neural network model to the server 120, and the server 120 is configured to obtain preference propagation information of the target user based on the historical behavior data provided by the terminal device 110, input the preference propagation information and a preset feature vector to the attention mechanism neural network model for training, and complete interest recommendation. Optionally, in the process of performing interest recommendation, the server 120 undertakes primary computing work, and the terminal device 110 undertakes secondary computing work; alternatively, server 120 undertakes the secondary computing work and terminal device 110 undertakes the primary computing work; alternatively, the server 120 or the terminal device 110 can be capable of undertaking the computing work individually.

The terminal device 110 may include, but is not limited to, a computer, a smart phone, a tablet computer, an e-book reader, a motion Picture experts group audio layer III (MP 3) player, a motion Picture experts group audio layer 4 (MP 4) player, a portable computer, a vehicle-mounted computer, a wearable device, a desktop computer, a set-top box, a smart television, and the like.

The server 120 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as cloud service, a cloud database, cloud computing, a cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, content Delivery Network (CDN), big data, and an artificial intelligence platform.

Alternatively, the number of the terminals 110 or the servers 120 may be more or less, and the embodiment of the present application is not limited thereto.

The above briefly describes a scene schematic diagram of the present application, and the following takes the server 120 applied in fig. 1 as an example to describe in detail the interest recommendation method provided in the embodiment of the present application.

Referring to fig. 2, fig. 2 is a flowchart illustrating an interest recommendation method according to an embodiment of the present disclosure, where the method includes steps S201 to S204.

Step S201, obtaining historical behavior data of a target user in a preset time period, and obtaining preference propagation information of the target user based on the historical behavior data.

In this embodiment, historical behavior data of a target user within a preset time period is obtained, and illustratively, the historical behavior data is obtained by extracting a behavior log table in which the target user logs in an APP "my XX (province)" and operations such as clicking, submitting, searching, commenting and the like are performed within 3 months. It is understood that the preset time period can be adaptively set by those skilled in the art in combination with practical applications, and in other embodiments, may not be limited to 3 months.

In the embodiment, the preference propagation information of the target user is obtained, the preference propagation information is obtained by establishing the item knowledge graph, and further, the potential preference of the user can be obtained by combining the propagation of water flow fluctuation. It is understood that a knowledge graph is a graph-based data structure, consisting of "entities" and "relationships" between entities, and that a knowledge graph is essentially a semantic network. An entity refers to things in the real world, such as events and the like; relationships are used to express some kind of relationship between different entities, which is an attribute relationship in this embodiment. "Water flow fluctuation" has two meanings: first, similar to the real ripples created by multiple raindrops, the user's potential interest in an entity is generated by his historical preferences and then propagates layer by layer (k) along the links of the knowledge-graph, from near to far. Second, the intensity of the potential preference of the user for ripple concentration diminishes as the number of hop water layers increases, similar to the magnitude of the gradual attenuation of the actual ripple.

In one embodiment, the obtaining of the preference propagation information of the target user based on the historical behavior data in step S201 may include the following steps:

converting the historical behavior data into a knowledge triple form, and establishing a transaction knowledge graph based on the historical behavior data subjected to data conversion;

Specifically, historical behavior data is converted into a knowledge triple form, a (item) knowledge triple is constructed, a knowledge graph relation table and an adjacency matrix are formed, the relation in each triple instance (head, relation, tail) is expressed as a relation (r) from head (h) to tail (t), and a logic structure processed in the knowledge graph relation table is as follows: { item _ id _ h, relation, item _ id _ t }, such as { social security query, parent class, social security }. And processing the knowledge graph relation table into an item adjacency matrix representing the adjacent relation between the vertexes, establishing an item knowledge graph, and acquiring the preference propagation information of the target user according to the item knowledge graph.

Further, by obtaining the item category dictionary table and the item attribute relationship table, the logic processing is performed on the historical behavior data, and the historical behavior data is converted into a knowledge triple form, where the conversion of the historical behavior data into the knowledge triple form may include the following steps:

Specifically, the item dictionary table can be processed into item _ index as an index table about item id { item _ id, index }, so as to find the corresponding item quickly.

Specifically, the attribute relationship between the transactions, for example, the relationship between the transaction a and the transaction B belongs to the subordinate/parent/child/descendant, or the like.

Specifically, in the embodiment, as shown in fig. 3, the numbers corresponding to the head and tail in the triple of item knowledge are the number of interactions of the item. In some embodiments, a scoring table is built to represent the number of user interactions with an item: in view of no scoring item collection in the government affair scene, the interaction times of the user to the matters are specially used for replacing the numerical representation of the preference of the user, namely, the behavior log table of the user is further processed, and the interaction times of the user to the matters are added and recorded to form a rating table: { user _ id, item _ id, rating }.

In the process of acquiring the preference propagation information of the knowledge graph, although the preference information of the user can be accurately screened out, the potential preference of the user is ignored to a certain extent. For this purpose, in an implementation manner of this embodiment, the obtaining of the preference propagation information of the target user based on the item knowledge graph may include the following steps:

Specifically, the events interested by the user is regarded as seeds (seeds), and the seeds are spread to other events by one circle on the event knowledge map, which is called Preference Propagation. The outer-layer items also belong to the potential preferences of the user, and when the user is analyzed in the item knowledge graph, the items are taken into consideration instead of only using the observed items to obtain the preference of the user.

In particular, for a target user u, his (historical) set of features of interest V _u Considered as seeds in the knowledge-graph and then extended along the links to form a plurality of corrugated sets

A corrugated unit

Is a set of distance seed sets V _u The k-hop (k-hop) algorithm in the graph algorithm refers to a knowledge triple set which starts from a certain starting point and finds a set of all vertices whose shortest paths do not exceed k, as shown in fig. 4.

Step S202, inputting the preference propagation information and the preset item feature vector into an attention mechanism neural network model.

In practical applications, the prediction and recommendation are usually performed simultaneously for a plurality of preset item feature vectors. It is understood that the default event feature, i.e. the feature vector of the given event to be predicted, can be determined by one skilled in the art in combination with the actual application.

Specifically, with reference to fig. 5a, the preference propagation information (knowledge triples of each ripple group) of the user u and the feature vector of the (preset) event V are used as the input of the attention mechanism neural network model, deep learning is performed in the attention neural network model, and finally the predicted probability that the user u will click on (or otherwise operate) the event V is output.

Step S203, obtaining a user feature vector corresponding to the preference propagation information based on the attention mechanism neural network model, and splicing the user feature vector and a preset item feature vector to obtain a spliced feature vector.

In the embodiment, the preference propagation information and the preset item feature vector of the user are input into the attention neural network model, the user feature vector corresponding to the preference propagation information is obtained, the relevance between the user feature vector and the item can be well reflected, and the accuracy of the recommendation result is effectively improved.

In one embodiment, the obtaining the user feature vector corresponding to the preference propagation information based on the attention mechanism neural network model in step S203 may include the following steps:

In connection with fig. 5a, the set of knowledge triplets (knowledge triplets of each ripple set) is iteratively embedded with the transaction into the attention mechanism neural network model, interacted with in the model to obtain a representation relationship of the user u with respect to the transaction V, further, attention weight representations of the nodes are computed in the representation relationship (multi-level graph attention network), and then combined to form the end-user embedding, i.e., user feature vector.

In this embodiment, the embedded embedding vector representation of a transaction item V is as shown in fig. 5b, each item V being associated with an item embedding V. Item embedding may incorporate one hot ID, attributes, context information, etc. of a transaction, based on the application scenario.

In one embodiment, the obtaining, in the neural network model, a user feature vector based on the knowledge triples and the event feature vector may include the following steps:

In the embodiment, each triplet (hi, ri, ti) in 1 ripple group Su1, su1 of a given event embedding V and a user u allocates a relevant probability by comparing tail event V with a head hi and a relation Ri in the triplet, and in the process of calculating the probability, softmax probability is not directly output, but a multilayer attention-seeking neural network idea is introduced, and probability prediction is performed on a method for extracting node hidden layer vector representation.

Further, the obtaining of the correlation probability between the knowledge triple set and the preset item feature vector in the neural network model specifically includes:

and acquiring the correlation probability of the knowledge triple set and the preset item feature vector in the neural network model based on the attention weight between the node corresponding to the knowledge triple of each attention network layer and the neighbor node thereof, the weight matrix of each attention network layer and a preset activation function.

Specifically, the above process may satisfy the following formula:

in the formula, p _i Representing the correlation probability of i nodes corresponding to the knowledge triple and the feature vector of the preset item, sigma representing an activation function, K' representing the number of layers of the attention network, W ^k’ A weight matrix representing the k' th layer,

denotes the attention weight, N, between node i and node j of the k' th layer _i Represents a domain composed of all nodes adjacent to node i,

representing the feature vector of node j.

The derivation process of the above formula is: with single-layer graph attention, for a knowledge triplet tail entry (node) i, the calculation process considering the attention weight of its neighbor node j is as follows:

in connection with fig. 5c, the input is two item event vectors hi, hj, in order to transform them to RF, a weight matrix W to be learned is introduced, and a vector a to be learned gives an event embedding V and each triplet (hi, ri, ti) in the 1 ripple group Su1, su1 of the user u assigns a relevant probability by comparing tail event V with the head hi and the relation Ri in the triplet, and during the calculation of the probability, the softmax probability is not directly output, but rather, the multi-layer attention-seeking neural network concept is introduced, and the probability prediction is performed on the method of extracting node hidden layer vector representation; two operations are performed: w chi and W chi hj, and obtaining two RF dimensional vectors; calculating attention value e of node i on node j _ij = α (Whi, whj). α is a mapping of RF x RF → R;

finally, after e is obtained for all neighborhood nodes of the node i, the attention weight normalization operation is completed by utilizing softmax, and the specific calculation process is as follows:

wherein alpha is _ij To normalize the attention weight after all nodes, e _ij The activation function selects the LeakyReLU function for attention weight of node i on its neighbor node j, and the Leaky ReLU with leakage modified Linear Unit (Leaky ReLU) function is a variant of the classical ReLu activation function, with the output of the function having a small slope to the negative input. Since the derivative is always non-zero, this can reduce the occurrence of silent neurons, allowing gradient-based learning (albeit slowly), solving the problem of non-learning of neurons after the Relu function enters the negative interval. The Leaky ReLU function is shown in FIG. 5 d.

After the attention weights of all the nodes are normalized, the information extraction of the nodes can be carried out through the attention layer. The output value pi of the entire network is calculated as follows (where σ represents the activation function), where the probability output of the single-layer attention network is:

for the multi-layer attention network layer, as shown in connection with fig. 5e, the multi-layer attention mechanism exists in the following sense: different features may need to be assigned different attention weights, and if only a single attention layer is used, the same attention weight is used for all attributes of the neighborhood node, which may weaken the learning ability of the model.

After introducing the multilayer attention network, K 'represents the number of attention layers (e.g. the number of wavy lines and straight lines is 3, defining K' = 3), double vertical lines | | is the concatenation, and the adjustment formula of the network middle layer is as follows:

based on the formula, the correlation probability of the output layer is obtained:

in one embodiment, the obtaining the user feature vector based on the correlation probability may include:

Specifically, the above process may satisfy the following formula:

in the formula (I), the compound is shown in the specification,

indicating groups of corrugations

Corresponding first order user feature vectors, (hi, r) _i ,t _i ) Indicating groups of corrugations

Knowledge triplets of neutralization, t _i To represent

U denotes the user feature vector,

representing user feature vectors of order 1 to H, respectively.

Specifically, an input vector, that is, a vector represented by the second vertical rectangle in fig. 5a is calculated, and after the correlation probability is obtained, the tail (ti) in Su1 is multiplied by the corresponding correlation probability to perform weighted summation, so as to obtain a vector ou1:

as shown in connection with FIG. 5a, the user's interest is transferred along the links in Su1 from its history set Vu to its set of 1-hop related entities u1, which is called preference propagation in the waterflow fluctuation model, and this process is repeated, obtaining the user's 2 nd order

… …. Obtaining the preference of each-order user through H-order propagation from the history click item preference of the user

H, as understood, corresponds to k above.

And finally, obtaining the final embedding of the user u to the item v, namely the last vertical rectangle in the graph, wherein the user feature vector is as follows:

and step S204, recommending related interest items to the target user in response to the prediction result of the spliced feature vector.

Specifically, user embedding and item embedding are spliced to obtain a splicing feature vector u ^T v, by piecing together feature vectors u ^T v, in this embodiment, the prediction process may be obtained according to the following formula, and the probability of the click (or other operation) to be predicted is obtained finally:

wherein the content of the first and second substances,

illustratively, if the probability is greater than 0.5, we determine that the positive sample represents a click, and if the probability is less than or equal to 0.5, we determine that the negative sample is a no click.

In some embodiments, to improve the recommendation efficiency, the output data form of the prediction result is shown in fig. 6, where three columns are user _ id, item _ id and label (1 represents positive sample, and 0 is negative sample). label represents the labeled results of predicting clicks and clicks of users and things.

Further, in the process of training the attention neural network model, user behavior data of a certain day in the historical behavior data can be extracted as a sample verification tag table for predicting the event click probability result of the model training end user: { user _ id, item _ id, label }.

Further, the recommendation result is verified: by adopting an offline AUC (Area Under the user, model evaluation index) observation, the user item click problem can be regarded as a two-classification problem, and 0,1 represents whether the user clicks the item or not. The AUC value of the model is 7.2% higher than that of the existing model through the off-line model test. As can be understood, AUC is an important index for measuring the classification effect of the model in the industry, and the corresponding AUC value is larger and the classification effect is better; and performing AB-Test result comparison on the historical online recommendation result and the recall rate result of the click effect of the user by adopting online effect verification and the recommendation result of the coordination filtering model method in the related technology.

In conclusion, the embodiment strengthens the logic relevance between the introduced items, is beneficial to discovering the potential relation between the introduced items and improves the precision of the recommended items; a multilayer attention-drawing neural network is introduced, and the representation calculation result of the drawing neural network is enhanced; the diversity of recommendation results is increased, and based on various types of relation composition, the method is beneficial to reasonably maintaining the benefits of users and increasing the diversity of recommendation items; meanwhile, the system has interpretability, and the knowledge graph links the historical records of the user with the recommended items, so that the interpretability is brought to a recommendation system; the recall rate of the recommendation result is improved, and through online testing and data analysis, the related evaluation indexes of the recommendation model are effectively improved, wherein the recall rate is improved by nearly 25%.

Correspondingly, an interest recommendation apparatus is further provided in an embodiment of the present application, as shown in fig. 7, including:

a data acquisition module 71, configured to acquire historical behavior data of a target user within a preset time period;

a preference obtaining module 72 configured to obtain preference propagation information of the target user based on the historical behavior data;

an input module 73 configured to input the preference propagation information and the pre-set event feature vector into an attention mechanism neural network model;

a machine learning module 74, configured to obtain a user feature vector corresponding to the preference propagation information based on the attention mechanism neural network model, and splice the user feature vector and a preset item feature vector to obtain a spliced feature vector;

an interest recommendation module 75 configured to recommend relevant interest items to the target user in response to the prediction result of the stitched feature vector.

In one embodiment, the preference obtaining module 73 includes:

the conversion establishing unit is arranged for converting the historical behavior data into a knowledge triple form and establishing a transaction knowledge graph based on the historical behavior data subjected to data conversion;

a preference acquisition unit configured to acquire preference propagation information of the target user based on the item knowledge graph.

In one embodiment, the conversion establishing unit is specifically configured to obtain an item category dictionary table and an item attribute relationship table, where the item category dictionary table carries item IDs and belonging categories, and the item attribute relationship table carries attribute relationships between items; acquiring the interaction times of the target user and each item in the item category dictionary table based on the historical behavior data; and acquiring historical behavior data corresponding to the knowledge triple form based on the interaction times of the target user and each item in the item category dictionary table and the attribute relationship.

In one embodiment, the preference obtaining unit is specifically configured to regard the feature set of interest of the target user as a seed in the event knowledge graph, extend along links of the event knowledge graph to form a plurality of ripple groups, and find, based on the ripple groups, all sets of vertices whose shortest paths to the set of vertices do not exceed a preset number as a set of knowledge triples; and acquiring the preference propagation information of the target user based on the knowledge triple set.

In one embodiment, the machine learning module 74 is specifically configured to iteratively embed the preference propagation information and the default event feature vector into an attention neural network model, and obtain a user feature vector based on the knowledge triplet sets and the default event feature vector in the neural network model.

In one embodiment, the obtaining, in the neural network model, a user feature vector based on the knowledge triples and the event feature vector includes: and acquiring the correlation probability of the knowledge triple set and the preset item feature vector in the neural network model, and acquiring a user feature vector based on the correlation probability.

In an embodiment, the obtaining, in the neural network model, a probability of a correlation between the knowledge triple set and the default feature vector includes: and acquiring the related probability of the knowledge triplet set and the preset item feature vector in the neural network model based on the attention weight between the node corresponding to the knowledge triplet of each attention network layer and the neighborhood node thereof, the weight matrix of each attention network layer and a preset activation function.

In an embodiment, the obtaining the user feature vector based on the correlation probability specifically includes: and acquiring user feature vectors of the users in all orders corresponding to the plurality of corrugated groups based on the correlation probability, and acquiring final user feature vectors based on the user feature vectors of all orders.

It should be noted that, the apparatus provided in the present application can correspondingly implement all the method steps implemented by the server in the foregoing method embodiment, and can achieve the same technical effect, and detailed descriptions of the same parts and beneficial effects as those in the method embodiment are not repeated herein.

Correspondingly, an electronic device is further provided in an embodiment of the present application, as shown in fig. 8, including: a memory 81 and a processor 82;

the memory 81 stores computer-executable instructions;

the processor 82 executes the computer-executable instructions stored by the memory 81 to cause the electronic device to perform the interest recommendation method.

It should be noted that, the electronic device provided in the present application can correspondingly implement all the method steps implemented by the server in the foregoing method embodiment, and can achieve the same technical effect, and detailed descriptions of the same parts and beneficial effects as those in the method embodiment are not repeated here.

The embodiment of the present application also provides a computer-readable storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed by a processor, the computer-executable instructions are used for implementing the interest recommendation method.

The embodiment of the present application also provides a computer program product, where the computer program product includes computer program code, and when the computer program code runs on a computer, the computer is caused to execute the interest recommendation method.

The chip comprises a memory and a processor, wherein the memory is used for storing a computer program, and the processor is used for calling and running the computer program from the memory and executing the interest recommendation method.

It should be noted that, the computer-readable storage medium, the chip, and the product provided in the present application can correspondingly implement all the method steps implemented by the server in the foregoing method embodiments, and can achieve the same technical effects, and detailed descriptions of the same parts and beneficial effects as those of the method embodiments in this embodiment are not repeated herein.

It will be understood by those of ordinary skill in the art that all or some of the steps of the methods, systems, functional modules/units in the devices disclosed above may be implemented as software, firmware, hardware, and suitable combinations thereof. In a hardware implementation, the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be performed by several physical components in cooperation. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on computer readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media).

The term computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, as is well known to those skilled in the art. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital Versatile Disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can accessed by a computer.

In addition, communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media as known to those skilled in the art.

In the description of the embodiments of the present application, the term "and/or" merely represents an association relationship describing an associated object, and means that three relationships may exist, for example, a and/or B may represent: a exists alone, A and B exist simultaneously, and B exists alone. In addition, the term "at least one" means any one of a variety or any combination of at least two of a variety, for example, including at least one of A, B, and may mean any one or more elements selected from the group consisting of A, B and C communication. Further, the term "plurality" means two or more unless specifically stated otherwise.

In the description of the embodiments of the present application, the terms "first," "second," "third," "fourth," and the like (if any) are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the application described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

Finally, it should be noted that: the above embodiments are only used for illustrating the technical solutions of the present application, and not for limiting the same; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art will understand that: the technical solutions described in the foregoing embodiments may be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An interest recommendation method, comprising:

2. The method of claim 1, wherein the obtaining the preference propagation information of the target user based on the historical behavior data comprises:

3. The method of claim 2, wherein converting the historical behavior data into a knowledge triplet form comprises:

4. The method of claim 2, wherein the obtaining preference propagation information of the target user based on the transaction knowledge-graph comprises:

5. The method of claim 4, wherein obtaining the user feature vector corresponding to the preference propagation information based on an attention mechanism neural network model comprises:

6. The method of claim 5, wherein obtaining a user feature vector based on the set of knowledge triples and the transaction feature vector in the neural network model comprises:

7. The method of claim 6, wherein obtaining the probability of the knowledge triplet sets being associated with the default event feature vector in the neural network model comprises:

8. The method of claim 6, wherein the obtaining a user feature vector based on the correlation probability comprises:

and acquiring user feature vectors of the users in each order corresponding to the plurality of corrugated groups based on the correlation probability, and acquiring final user feature vectors based on the user feature vectors of each order.

9. An interest recommendation apparatus, comprising:

the data acquisition module is arranged for acquiring historical behavior data of a target user within a preset time period;

an input module configured to input the preference propagation information and the pre-set event feature vector into an attention mechanism neural network model;

and the interest recommendation module is used for recommending related interest items to the target user in response to the prediction result of the spliced feature vector.

10. An electronic device, comprising: a memory and a processor;

the memory stores computer-executable instructions;

the processor executes the computer-executable instructions stored by the memory to cause the electronic device to perform the interest recommendation method of any of claims 1-8.

11. A computer-readable storage medium having computer-executable instructions stored thereon, which when executed by a processor, are configured to implement the interest recommendation method of any one of claims 1-8.

12. A computer program product, characterized in that the computer program product comprises computer program code which, when run on a computer, causes the computer to perform the interest recommendation method according to any one of claims 1-8.