CN109544306B - Cross-domain recommendation method and device based on user behavior sequence characteristics - Google Patents
Cross-domain recommendation method and device based on user behavior sequence characteristics Download PDFInfo
- Publication number
- CN109544306B CN109544306B CN201811455642.5A CN201811455642A CN109544306B CN 109544306 B CN109544306 B CN 109544306B CN 201811455642 A CN201811455642 A CN 201811455642A CN 109544306 B CN109544306 B CN 109544306B
- Authority
- CN
- China
- Prior art keywords
- domain
- target
- user
- target user
- preference
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Finance (AREA)
- Accounting & Taxation (AREA)
- Biophysics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a cross-domain recommendation method based on user behavior sequence characteristics, which comprises the following steps: the method comprises the steps of respectively obtaining interactive item sequences of a target user in an auxiliary domain and a target domain, respectively obtaining a user embedded matrix and an item embedded matrix of the target user in the auxiliary domain and the target domain through an embedded layer of a cross-domain recommendation model, respectively obtaining total preference characteristics of the target user in the auxiliary domain and the target domain through an LSTM layer, finally migrating the total preference characteristics of the user in the auxiliary domain to the target domain through MLP layer processing, obtaining a preference score of the target user for each item in the target domain through calculation, and recommending items which are possibly purchased or accessed within a next period of time for the target user in the target domain. By applying the technical scheme provided by the implementation of the invention, the recommendation accuracy of the target domain is improved, and the recommendation performance is improved. The invention also discloses a cross-domain recommendation device based on the user behavior sequence characteristics, and the cross-domain recommendation device has corresponding technical effects.
Description
Technical Field
The invention relates to the technical field of recommendation, in particular to a cross-domain recommendation method and device based on user behavior sequence characteristics.
Background
With the continuous development of the mobile internet technology, the amount of information in the network is rapidly expanded and increased in an exponential manner, and the problems of information overload and information lost on the network are increasingly serious. Recommendation systems have been developed to provide users with satisfactory information and services, and have become a research area of interest to many researchers. The recommendation system performs information filtering by predicting the user's preference for information resources.
In the field of recommendation systems, data sparsity and cold start are still and challenging problems, and researchers have proposed a wide variety of solutions. In recent years, a new research trend, namely cross-domain recommendation, is emerging, aiming at alleviating the influence of data sparsity and cold start on the performance of a recommendation system. In reality, these problems are encountered in different recommendation fields. For example, an online shopping website has more than one commodity field, such as books, beauty products, electronic products, movies, etc. The preferences of the same user in different fields may be similar, so that it is a good choice to migrate the preference characteristics of the user in a certain field to the target field to improve the recommendation performance of the target field.
In recent years, there has been research work that has begun to focus on cross-domain recommendations, but the amount of research has been very small. The existing cross-domain recommendation research method is mainly an asymmetric mode, and the recommendation performance of a target domain with sparse data is relieved by utilizing information in an auxiliary domain. In the method, mainly, knowledge and patterns in the information-rich field are learned, and then the information is migrated to the target field in a regular term or a priori manner to improve the accuracy of recommendation, and the key point is to find out which information is suitable to be migrated to the target field. However, this method has a certain limitation because the data information of the two domains is not fully utilized, and the recommendation performance is not good.
Disclosure of Invention
The invention aims to provide a cross-domain recommendation method and device based on user behavior sequence characteristics so as to improve the recommendation accuracy of a target domain and improve the recommendation performance.
In order to solve the technical problems, the invention provides the following technical scheme:
a cross-domain recommendation method based on user behavior sequence features comprises the following steps:
obtaining a first interactive project sequence of a target user in an auxiliary domain and a second interactive project sequence of the target user in the target domain;
mapping the target user, the first interaction project sequence and the second interaction project sequence to continuous low-dimensional space respectively on an embedding layer of a cross-domain recommendation model obtained through pre-training, and obtaining a first user embedding matrix and a first interaction project embedding matrix of the target user in the auxiliary domain and a second user embedding matrix and a second interaction project embedding matrix of the target user in the target domain;
determining the total preference characteristics of the target user in the auxiliary domain and the total preference characteristics of the target user in the target domain according to the first user embedding matrix, the first interactive item embedding matrix, the second user embedding matrix and the second interactive item embedding matrix respectively on an LSTM layer of the cross-domain recommendation model;
on the MLP layer of the cross-domain recommendation model, migrating the total preference characteristics of the target user in the auxiliary domain to the target domain by utilizing an MLP mapping function based on the total preference characteristics of the target user in the auxiliary domain and the total preference characteristics of the target user in the target domain;
and obtaining the preference score of the target user for each item in the target domain through calculation, and determining the item recommended to the target user in the target domain according to the preference score of the target user for each item in the target domain.
In one embodiment of the present invention, the total preference characteristic of the target user in the auxiliary domain is determined by the following steps:
based on an LSTM updating mode, obtaining the sequence preference of the target user;
determining a combination of the first user embedding matrix and the sequence preference as a total preference profile of the target user in the auxiliary domain.
In one embodiment of the present invention, the overall preference characteristic of the target user in the target domain is determined by the following steps:
determining a combination of the second user embedding matrix and the sequence preference as a total preference profile of the target user in the target domain.
In a specific embodiment of the present invention, the migrating the total preference characteristics of the target user in the auxiliary domain to the target domain by using an MLP mapping function based on the total preference characteristics of the target user in the auxiliary domain and the total preference characteristics of the target user in the target domain includes:
migrating the overall preference characteristics of the target user in the auxiliary domain to the target domain by the following formula:
wherein the content of the first and second substances,for the target user uiIn the overall preference feature of the auxiliary domain,for affine features of the target user in the target domain, fmlpTheta represents the MLP mapping function, and is its parameter set, including the weight matrix and bias terms.
In an embodiment of the present invention, the determining, according to the favorite score of the target user for each item in the target domain, an item recommended to the target user in the target domain includes:
and sorting the items according to the preference score of the target user for each item in the target domain, and selecting the items with the preset number to recommend to the target user.
A cross-domain recommendation device based on user behavior sequence features comprises:
the project sequence obtaining module is used for obtaining a first interactive project sequence of a target user in an auxiliary domain and a second interactive project sequence of the target user in a target domain;
the embedded layer processing module is used for mapping the target user, the first interactive project sequence and the second interactive project sequence to continuous low-dimensional space respectively on an embedded layer of a cross-domain recommendation model obtained through pre-training to obtain a first user embedded matrix and a first interactive project embedded matrix of the target user in the auxiliary domain and a second user embedded matrix and a second interactive project embedded matrix of the target user in the target domain;
the LSTM layer processing module is used for determining the total preference characteristics of the target user in the auxiliary domain and the total preference characteristics of the target user in the target domain according to the first user embedding matrix, the first interactive item embedding matrix, the second user embedding matrix and the second interactive item embedding matrix respectively on the LSTM layer of the cross-domain recommendation model;
an MLP layer processing module, configured to, at an MLP layer of the cross-domain recommendation model, migrate the total preference feature of the target user in the auxiliary domain into the target domain using an MLP mapping function based on the total preference feature of the target user in the auxiliary domain and the total preference feature of the target user in the target domain;
and the item recommending module is used for obtaining the preference score of the target user for each item in the target domain through calculation and determining the item recommended to the target user in the target domain according to the preference score of the target user for each item in the target domain.
In an embodiment of the present invention, the LSTM layer processing module is specifically configured to determine the total preference characteristics of the target user in the auxiliary domain by:
based on an LSTM updating mode, obtaining the sequence preference of the target user;
determining a combination of the first user embedding matrix and the sequence preference as a total preference profile of the target user in the auxiliary domain.
In a specific embodiment of the present invention, the LSTM layer processing module is specifically configured to determine the total preference characteristics of the target user in the target domain by the following steps:
determining a combination of the second user embedding matrix and the sequence preference as a total preference profile of the target user in the target domain.
In an embodiment of the present invention, the MLP layer processing module is specifically configured to:
migrating the overall preference characteristics of the target user in the auxiliary domain to the target domain by the following formula:
wherein the content of the first and second substances,for the target user uiIn the overall preference feature of the auxiliary domain,for affine features of the target user in the target domain, fmlpTheta represents the MLP mapping function, and is its parameter set, including the weight matrix and bias terms.
In an embodiment of the present invention, the item recommendation module is specifically configured to:
and sorting the items according to the preference score of the target user for each item in the target domain, and selecting the items with the preset number to recommend to the target user.
By applying the technical scheme provided by the embodiment of the invention, a first interactive item sequence of a target user in an auxiliary domain and a second interactive item sequence of the target user in the target domain are obtained, the target user and each item sequence are respectively mapped to continuous low-dimensional spaces in an embedding layer of a cross-domain recommendation model obtained by pre-training, a user embedding matrix and an interactive item embedding matrix of the target user in the auxiliary domain and a user embedding matrix and an interactive item embedding matrix of the user in the target domain are obtained, the total preference characteristic of the target user in the auxiliary domain and the total preference characteristic of the target user in the target domain are determined on the basis of the matrixes in an LSTM layer of the cross-domain recommendation model, the total preference characteristic of the target user in the auxiliary domain is transferred to the target domain on the basis of the total preference characteristic on the basis of the MLP layer of the cross-domain recommendation model by using an MLP mapping function, the preference score of the target user on each item in the target domain is obtained by calculation, and further determining the items recommended to the target user in the target domain. The behavior sequence characteristics of the user in the auxiliary domain are transferred to the target domain for cross-domain recommendation, so that the recommendation accuracy of the target domain is improved, and the recommendation performance is improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a diagram illustrating a sensor model according to an embodiment of the present invention;
FIG. 2 is a diagram illustrating a neuron model in a sensor according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of an LSTM network according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of an MLP structure in accordance with an embodiment of the present invention;
FIG. 5 is a flowchart illustrating an implementation of a cross-domain recommendation method based on user behavior sequence characteristics according to an embodiment of the present invention;
fig. 6 is a schematic structural diagram of a cross-domain recommendation device based on user behavior sequence characteristics in an embodiment of the present invention.
Detailed Description
In order that those skilled in the art will better understand the disclosure, the invention will be described in further detail with reference to the accompanying drawings and specific embodiments. It is to be understood that the described embodiments are merely exemplary of the invention, and not restrictive of the full scope of the invention. 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 invention.
The core of the invention is to provide a cross-domain recommendation method based on user behavior sequence characteristics, and the method carries out cross-domain recommendation by utilizing a cross-domain recommendation model obtained by pre-training.
In the embodiment of the invention, the cross-domain recommendation model combines a Long-Short Term Memory (LSTM) network and a multi-layer Perceptron (MLP) network to form a hierarchical deep network framework. The cross-domain recommendation model comprises a first layer, an Embedding layer, an LSTM layer and an MLP layer, wherein the first layer is used for mapping user and item sequences to a continuous low-dimensional space, the second layer is used for extracting sequence characteristics of user behaviors and inherent preference characteristics of a user, and the third layer is used for migrating the characteristics extracted from an auxiliary domain (also called a source domain) to a target domain for characteristic fusion.
LSTM is a variant of RNN (Recurrent Neural Network) that is capable of learning both short-term and long-term dependencies. LSTM is an efficient and scalable sequence prediction problem model. In embodiments of the present invention, the basic LSTM model is used for simplicity and general purpose, and can be easily extended to other LSTM variants. The basic update equation for LSTM is as follows:
wherein it、ft、otAn input gate, a forgetting gate and an output gate respectively representing the t-th item (item) determine the storage, forgetting and output of the information. c. CtThe state of the delegate unit is the unit activation vector, which is the key of the LSTM. x is the number oftAnd htRepresenting the input feature vector and the hidden output vector, respectively. σ is a sigmoid layer, and maps values between 0 and 1, with 1 representing "fully reserved" and 0 representing "fully abandoned". Wi、Wf、Wo、WcIs the weight of the door, bi、bf、bo、bcIs a corresponding offset to the offset of the first,representing the product of pairwise elements (Hadamard). Cell state ctHas two parts, one part being the previous cell state ct-1For controlling the forgetting door ftAnother part is that the user decides how many state values to add to define a range for new candidate values.
MLP is a multi-layered perceptron, an artificial neural network of forward architecture, mapping a set of input vectors to a set of output vectors. An MLP can be viewed as a directed graph, consisting of multiple layers of nodes, each layer being fully connected to the next. Each node, except the input nodes, is a neuron (or processing unit) with a nonlinear activation function. Supervised learning methods of back-propagation algorithms are often used to train MLPs. The MLP is the popularization of the sensor, and the defect that the sensor cannot recognize linear irreparable data is overcome. As shown in fig. 1, from left to right, there are an input layer, a hidden layer and an output layer, wherein the number of hidden layers is not specified by MLP, so that the appropriate number of hidden layers can be selected according to actual requirements. And there is no limit to the number of output layer neurons. FIG. 2 is a diagram of a neuron model in a perceptron.
The training process of the cross-domain recommendation model is described first below.
In the embodiment of the invention, it is assumed that the two domains of the auxiliary domain and the target domain have a common user, and U ═ U is used1,u2,…,uMDenotes a set of M users,represents a collection of N items of the auxiliary domain,representing a set of L items in the target domain. For user u in the target domain, at time ti-1A list of sequential collections of items that have been previously interacted with (e.g., purchased or accessed, etc.)Shown asWhereinIndicating user u at time tiThe item that interacted with. The task of the training is to recommend items for the user in the target domain that may interact during the next period of time, in particular to recommend items to the user at time ti-1Followed by a collection of items that may be of interest. In order to improve the accuracy of prediction, the embodiment of the invention considers the behavior sequence characteristics of the user.
The interactive item sequence of the user in the auxiliary domain and the interactive item sequence of the user in the target domain can be collected through various ways, and a cross-domain recommendation model is constructed.
Similar to the discrete vocabulary notation in natural language processing, the original user and item Identification (ID) representation capabilities are limited. In an embedding layer of the cross-domain recommendation model, a user and an interaction item sequence of the user can be mapped into two continuous low-dimensional spaces which are respectively marked as U belonging to RK*|U|,V∈RK*|V|The user embedding matrix U and the interactive item embedding matrix V respectively represent user _ embedding and item _ embedding obtained through mapping, K represents the dimension of the low-dimensional space, and | U | and | V | respectively represent the number of users (users) and items (items). The user of the auxiliary domain and the interactive item sequence of the user can obtain a first user embedding matrix through the embedding layerAnd a first interaction item embedding matrixThe user of the target domain and the interactive item sequence of the user can obtain a second user embedding matrix through the embedding layerAnd a second interaction item embedding matrix
At the LSTM layer of the cross-domain recommendation model, because the LSTM network can characterize non-linear user interaction and can well model the dynamic sequence features of users, a standard LSTM can be selected for extracting the sequence features of user behaviors, and the network structure is shown in fig. 3. XtItems representing user interactions at time t, htRepresenting the dynamic interest preferences of the user. The size of LSTM may be set to K, where K coincides with the low dimensional spatial dimension size K in the embedding layer. In the embodiment of the present invention, user _ embedding may be regarded as inherent preference of a user, and h represents hidden layer hidden of LSTM, which may be regarded as sequence preference of the user and may be obtained based on an LSTM updating manner (formula (1)). The auxiliary domain is obtained after passing through the LSTM layerThe target domain is obtained through the LSTM layerNamely, the sum of the first user embedding matrix and the sequence preference is determined as the total preference characteristic of the user in the auxiliary domain, and the sum of the second user embedding matrix and the sequence preference is determined as the total preference characteristic of the user in the target domain.
Through the upper two layers obtain { Us,Ut,Vs,VtAnd describing the relationship between different fields through a mapping function in an MLP layer of the cross-field recommendation model. The MLP can be chosen as the mapping function because the MLP can describe this input-output structure well, as shown in fig. 4, which is more flexible than the linear model, and the MLP can be trained and optimized easily by using the back propagation algorithm.
The back propagation algorithm, namely the BP algorithm, is a learning algorithm suitable for a multilayer neuron network under the guidance of a mentor, and is established on the basis of a gradient descent method. The input-output relationship of the BP network is substantially a mapping relationship: the function performed by an input-m-output BP neural network is a continuous mapping from n-dimensional euclidean space to a finite field in m-dimensional euclidean space, which is highly non-linear. Its information processing ability comes from multiple composition of simple non-linear function, so it has strong function reproduction ability. This is the basis on which the BP algorithm is applied. The back propagation algorithm is mainly iterated by two links (excitation propagation and weight updating) repeatedly and circularly until the response of the network to the input reaches a preset target range.
The learning process of the BP algorithm consists of a forward propagation process and a backward propagation process. In the forward propagation process, input information passes through the hidden layer through the input layer, is processed layer by layer and is transmitted to the output layer. If the expected output value cannot be obtained in the output layer, taking the square sum of the output and the expected error as an objective function, turning into reverse propagation, calculating the partial derivative of the objective function to the weight of each neuron layer by layer to form the gradient of the objective function to the weight vector, and finishing the learning of the network in the weight modifying process as the basis for modifying the weight. And when the error reaches the expected value, the network learning is finished.
A specific optimization function for the MLP mapping function may be defined as follows:
wherein f ismlpTheta represents the MLP mapping function, and is its parameter set, including the weight matrix and bias terms.
For the convenience of training, in the embodiment of the present invention, a layer of MLP is selected as a mapping function, where the dimensions of both the input layer and the output layer are set to K, specifically, K may be set to 128, which is consistent with K mentioned above, the dimension of the hidden layer is set to 2K, the activating function selects a tan-sigmoid function, and these parameters may be learned by a random gradient descent method. Wherein a back-propagation algorithm is used to calculate gradients of these parameters, updating these parameters of the MLP with users common in the auxiliary domain and the target domain.
For a user in the target domain, there is noWith sufficient information, it is difficult to make accurate recommendations for it. However, with the inherent preferences of the user and the sequence preferences learned in the secondary domain, the accuracy of the recommendations can be improved. E.g. uiRepresenting a user in the target domain, the affine characteristics of the user in the target domain can be obtained according to the following formula
Indicating the overall preference characteristics of the user in the auxiliary domain. The next step can be based on this obtainedTo make recommendations. Each user gets oneEventually form a matrix U't∈RK*|U|Representing the user characteristics fusing the auxiliary domain and the target domain information according to U'tVtTAnd calculating the preference score of the user for each item in the target domain, and recommending the item which is possibly interested in the user in the next period of time for the user according to the calculation result.
After the cross-domain recommendation model is obtained through training, cross-domain recommendation can be performed by using the cross-domain recommendation model.
Referring to fig. 5, an implementation flowchart of a cross-domain recommendation method based on user behavior sequence features according to an embodiment of the present invention is shown, where the method may include the following steps:
s510: and obtaining a first interactive item sequence of the target user in the auxiliary domain and a second interactive item sequence of the target user in the target domain.
In an embodiment of the invention, the target user is a user to whom an item in the target domain is to be recommended. The target user has interactive projects in both the auxiliary domain and the target domain, and a first interactive project sequence of the target user in the auxiliary domain and a second interactive project sequence of the target user in the target domain can be obtained through multiple ways. The sequence of interactive items is a collection of items that have interacted with the target user in the past, arranged in chronological order.
S520: and mapping the target user, the first interactive project sequence and the second interactive project sequence to continuous low-dimensional space respectively on an embedded layer of a cross-domain recommendation model obtained by pre-training, and obtaining a first user embedded matrix and a first interactive project embedded matrix of the target user in an auxiliary domain, and a second user embedded matrix and a second interactive project embedded matrix of the target user in the target domain.
In the embedding layer of the cross-domain recommendation model obtained by training, users and items can be mapped to continuous low-dimensional space, namely, a target user, a first interactive item sequence and a second interactive item sequence are respectively mapped to the continuous low-dimensional space, a first user embedding matrix and a first interactive item embedding matrix of an auxiliary domain are obtained, and a second user embedding matrix and a second interactive item embedding matrix of the target domain are obtained at the same time.
S530: and determining the total preference characteristics of the target user in the auxiliary domain and the total preference characteristics of the target user in the target domain according to the first user embedding matrix, the first interactive item embedding matrix, the second user embedding matrix and the second interactive item embedding matrix respectively on an LSTM layer of the cross-domain recommendation model.
At the LSTM layer of the cross-domain recommendation model, sequence features of user behaviors and preference features inherent to the user can be extracted.
After the first user embedding matrix and the first interaction item embedding matrix of the auxiliary domain and the second user embedding matrix and the second interaction item embedding matrix of the target domain are obtained, the total preference characteristic of the target user in the auxiliary domain and the total preference characteristic of the target user in the target domain can be determined according to the obtained matrixes.
In one embodiment of the present invention, the total preference characteristic of the target user in the auxiliary domain may be determined by the following steps:
based on an LSTM updating mode, obtaining sequence preference of a target user;
and determining the combination of the first user embedding matrix and the sequence preference as the total preference characteristic of the target user in the auxiliary domain. Specifically, the sum of the first user embedding matrix and the sequence preference can be determined as the total preference characteristic of the target user in the auxiliary domain. Namely, it is
Accordingly, the overall preference characteristics of the target user in the target domain may be determined by:
and determining the combination of the second user embedding matrix and the sequence preference as the total preference characteristic of the target user in the target domain. Specifically, the sum of the second user embedding matrix and the sequence preference can be determined as the total preference characteristic of the target user in the target domain. Namely, it is
S540: on the MLP layer of the cross-domain recommendation model, based on the total preference characteristics of the target users in the auxiliary domain and the total preference characteristics of the target users in the target domain, the MLP mapping function is utilized to transfer the total preference characteristics of the target users in the auxiliary domain to the target domain.
In the MLP layer of the cross-domain recommendation model obtained by pre-training, the MLP mapping function has also been trained. Based on the total preference characteristics of the target user in the auxiliary domain and the total preference characteristics of the target user in the target domain, the MLP mapping function can be used for feature fusion of the auxiliary domain and the target domain, and the total preference characteristics of the target user in the auxiliary domain can be migrated into the target domain, specifically, the migration can be performed through a formula (3).
S550: and obtaining the preference score of the target user for each item in the target domain through calculation, and determining the item recommended to the target user in the target domain according to the preference score of the target user for each item in the target domain.
After the MLP mapping function is used for migrating the total preference characteristics of the target user in the auxiliary domain into the target domain, the preference score of the target user for each item in the target domain can be obtained through calculation. In particular, can be according to U'tVtTAnd calculating the preference score of the target user for each item in the target domain. In this connection, it is possible to use,a matrix is embedded for the second interactive item.
After the favorite score of the target user for each item in the target domain is obtained, the item recommended to the target user in the target domain can be determined according to the favorite score of the target user for each item in the target domain.
Specifically, the items may be sorted according to the preference score of the target user for each item in the target domain, and a set number of items may be selected and recommended to the target user. The number can be set and adjusted according to actual conditions, and the embodiment of the present invention is not limited thereto.
By applying the method provided by the embodiment of the invention, a first interactive item sequence of a target user in an auxiliary domain and a second interactive item sequence of the target user in the target domain are obtained, the target user and each item sequence are respectively mapped to continuous low-dimensional spaces in an embedding layer of a cross-domain recommendation model obtained by pre-training, a user embedding matrix and an interactive item embedding matrix of the target user in the auxiliary domain and a user embedding matrix and an interactive item embedding matrix of the user in the target domain are obtained, the total preference characteristic of the target user in the auxiliary domain and the total preference characteristic of the target user in the target domain are determined on the basis of the matrixes in an LSTM layer of the cross-domain recommendation model, the total preference characteristic of the target user in the auxiliary domain is transferred to the target domain on the basis of the total preference characteristic on the basis of the MLP layer of the cross-domain recommendation model by using an MLP mapping function, the preference score of the target user on each item in the target domain is obtained by calculation, and further determining the items recommended to the target user in the target domain. The behavior sequence characteristics of the user in the auxiliary domain are transferred to the target domain for cross-domain recommendation, so that the recommendation accuracy of the target domain is improved, and the recommendation performance is improved.
Corresponding to the above method embodiment, the embodiment of the present invention further provides a cross-domain recommendation device based on the user behavior sequence feature, and a cross-domain recommendation device based on the user behavior sequence feature described below and a cross-domain recommendation method based on the user behavior sequence feature described above may be referred to correspondingly.
Referring to fig. 6, the apparatus includes the following modules:
the project sequence obtaining module 610 is configured to obtain a first interactive project sequence of the target user in the auxiliary domain and a second interactive project sequence of the target user in the target domain;
an embedded layer processing module 620, configured to map the target user, the first interaction item sequence, and the second interaction item sequence to continuous low-dimensional spaces respectively at an embedded layer of a cross-domain recommendation model obtained through pre-training, so as to obtain a first user embedded matrix and a first interaction item embedded matrix of the target user in the auxiliary domain, and a second user embedded matrix and a second interaction item embedded matrix of the target user in the target domain;
an LSTM layer processing module 630, configured to determine, at an LSTM layer of the cross-domain recommendation model, a total preference feature of the target user in the auxiliary domain and a total preference feature of the target user in the target domain according to the first user embedding matrix, the first interactive item embedding matrix, the second user embedding matrix, and the second interactive item embedding matrix, respectively;
an MLP layer processing module 640, configured to migrate, at an MLP layer of the cross-domain recommendation model, the total preference feature of the target user in the auxiliary domain into the target domain using an MLP mapping function based on the total preference feature of the target user in the auxiliary domain and the total preference feature of the target user in the target domain;
and the item recommending module 650 is configured to obtain a preference score of the target user for each item in the target domain through calculation, and determine an item recommended to the target user in the target domain according to the preference score of the target user for each item in the target domain.
By applying the device provided by the embodiment of the invention, a first interactive item sequence of a target user in an auxiliary domain and a second interactive item sequence of the target user in the target domain are obtained, the target user and each item sequence are respectively mapped to continuous low-dimensional spaces in an embedding layer of a cross-domain recommendation model obtained by pre-training, a user embedding matrix and an interactive item embedding matrix of the target user in the auxiliary domain and a user embedding matrix and an interactive item embedding matrix of the user in the target domain are obtained, the total preference characteristic of the target user in the auxiliary domain and the total preference characteristic of the target user in the target domain are determined on the basis of the matrixes in an LSTM layer of the cross-domain recommendation model, the total preference characteristic of the target user in the auxiliary domain is transferred to the target domain on the basis of the total preference characteristic on the basis of the MLP layer of the cross-domain recommendation model by using an MLP mapping function, and the preference score of the target user for each item in the target domain is obtained by calculation, and further determining the items recommended to the target user in the target domain. The behavior sequence characteristics of the user in the auxiliary domain are transferred to the target domain for cross-domain recommendation, so that the recommendation accuracy of the target domain is improved, and the recommendation performance is improved.
In an embodiment of the present invention, the LSTM layer processing module 630 is specifically configured to determine the overall preference characteristics of the target user in the auxiliary domain by the following steps:
based on an LSTM updating mode, obtaining the sequence preference of the target user;
determining a combination of the first user embedding matrix and the sequence preference as a total preference profile of the target user in the auxiliary domain.
In an embodiment of the present invention, the LSTM layer processing module 630 is specifically configured to determine the total preference characteristics of the target user in the target domain by the following steps:
and determining the combination of the second user embedding matrix and the sequence preference as the total preference characteristic of the target user in the target domain.
In an embodiment of the present invention, the MLP layer processing module 640 is specifically configured to:
migrating the overall preference characteristics of the target user in the auxiliary domain to the target domain by the following formula:
wherein the content of the first and second substances,for the target user uiIn the overall preference feature of the auxiliary domain,for affine features of the target user in the target domain, fmlpTheta represents the MLP mapping function, and is its parameter set, including the weight matrix and bias terms.
In an embodiment of the present invention, the item recommendation module 650 is specifically configured to:
and sorting the items according to the preference score of the target user for each item in the target domain, and selecting the items with the preset number to recommend to the target user.
The embodiments are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same or similar parts among the embodiments are referred to each other.
Those of skill would further appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the various illustrative components and steps have been described above generally in terms of their functionality in order to clearly illustrate this interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in Random Access Memory (RAM), memory, Read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
The principle and the implementation of the present invention are explained in the present application by using specific examples, and the above description of the embodiments is only used to help understanding the technical solution and the core idea of the present invention. It should be noted that, for those skilled in the art, it is possible to make various improvements and modifications to the present invention without departing from the principle of the present invention, and those improvements and modifications also fall within the scope of the claims of the present invention.
Claims (6)
1. A cross-domain recommendation method based on user behavior sequence features is characterized by comprising the following steps:
obtaining a first interactive project sequence of a target user in an auxiliary domain and a second interactive project sequence of the target user in the target domain;
mapping the target user, the first interaction project sequence and the second interaction project sequence to continuous low-dimensional space respectively on an embedding layer of a cross-domain recommendation model obtained through pre-training, and obtaining a first user embedding matrix and a first interaction project embedding matrix of the target user in the auxiliary domain and a second user embedding matrix and a second interaction project embedding matrix of the target user in the target domain;
determining the total preference characteristics of the target user in the auxiliary domain and the total preference characteristics of the target user in the target domain according to the first user embedding matrix, the first interactive item embedding matrix, the second user embedding matrix and the second interactive item embedding matrix respectively on an LSTM layer of the cross-domain recommendation model;
on the MLP layer of the cross-domain recommendation model, migrating the total preference characteristics of the target user in the auxiliary domain to the target domain by utilizing an MLP mapping function based on the total preference characteristics of the target user in the auxiliary domain and the total preference characteristics of the target user in the target domain;
the preference score of the target user for each item in the target domain is obtained through calculation, and the item recommended to the target user in the target domain is determined according to the preference score of the target user for each item in the target domain;
wherein the overall preference characteristics of the target user in the auxiliary domain are determined by:
based on an LSTM updating mode, obtaining the sequence preference of the target user;
determining a combination of the first user embedding matrix and the sequence preference as a total preference characteristic of the target user in the auxiliary domain;
determining a total preference profile of the target user at the target domain by:
determining a combination of the second user embedding matrix and the sequence preference as a total preference profile of the target user in the target domain.
2. The method of claim 1, wherein the migrating the overall preference profile of the target user in the auxiliary domain into the target domain using an MLP mapping function based on the overall preference profile of the target user in the auxiliary domain and the overall preference profile of the target user in the target domain comprises:
migrating the overall preference characteristics of the target user in the auxiliary domain to the target domain by the following formula:
wherein the content of the first and second substances,for the target user uiIn the overall preference feature of the auxiliary domain,for affine features of the target user in the target domain, fmlpTheta represents the MLP mapping function, and is its parameter set, including the weight matrix and bias terms.
3. The method according to any one of claims 1 to 2, wherein the determining the items recommended to the target user in the target domain according to the preference score of the target user for each item in the target domain comprises:
and sorting the items according to the preference score of the target user for each item in the target domain, and selecting the items with the preset number to recommend to the target user.
4. A cross-domain recommendation device based on user behavior sequence features is characterized by comprising:
the project sequence obtaining module is used for obtaining a first interactive project sequence of a target user in an auxiliary domain and a second interactive project sequence of the target user in a target domain;
the embedded layer processing module is used for mapping the target user, the first interactive project sequence and the second interactive project sequence to continuous low-dimensional space respectively on an embedded layer of a cross-domain recommendation model obtained through pre-training to obtain a first user embedded matrix and a first interactive project embedded matrix of the target user in the auxiliary domain and a second user embedded matrix and a second interactive project embedded matrix of the target user in the target domain;
the LSTM layer processing module is used for determining the total preference characteristics of the target user in the auxiliary domain and the total preference characteristics of the target user in the target domain according to the first user embedding matrix, the first interactive item embedding matrix, the second user embedding matrix and the second interactive item embedding matrix respectively on the LSTM layer of the cross-domain recommendation model;
an MLP layer processing module, configured to, at an MLP layer of the cross-domain recommendation model, migrate the total preference feature of the target user in the auxiliary domain into the target domain using an MLP mapping function based on the total preference feature of the target user in the auxiliary domain and the total preference feature of the target user in the target domain;
the item recommending module is used for obtaining the preference score of the target user for each item in the target domain through calculation and determining the item recommended to the target user in the target domain according to the preference score of the target user for each item in the target domain;
wherein the LSTM layer processing module is specifically configured to determine the total preference characteristics of the target user in the auxiliary domain by:
based on an LSTM updating mode, obtaining the sequence preference of the target user;
determining a combination of the first user embedding matrix and the sequence preference as a total preference characteristic of the target user in the auxiliary domain;
determining a total preference profile of the target user at the target domain by:
determining a combination of the second user embedding matrix and the sequence preference as a total preference profile of the target user in the target domain.
5. The apparatus of claim 4, wherein the MLP layer processing module is specifically configured to:
migrating the overall preference characteristics of the target user in the auxiliary domain to the target domain by the following formula:
wherein the content of the first and second substances,for the target user uiIn the overall preference feature of the auxiliary domain,for affine features of the target user in the target domain, fmlpTheta represents the MLP mapping function, and is its parameter set, including the weight matrix and bias terms.
6. The apparatus according to any one of claims 4 to 5, wherein the item recommendation module is specifically configured to:
and sorting the items according to the preference score of the target user for each item in the target domain, and selecting the items with the preset number to recommend to the target user.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811455642.5A CN109544306B (en) | 2018-11-30 | 2018-11-30 | Cross-domain recommendation method and device based on user behavior sequence characteristics |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811455642.5A CN109544306B (en) | 2018-11-30 | 2018-11-30 | Cross-domain recommendation method and device based on user behavior sequence characteristics |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109544306A CN109544306A (en) | 2019-03-29 |
CN109544306B true CN109544306B (en) | 2021-09-21 |
Family
ID=65852018
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811455642.5A Active CN109544306B (en) | 2018-11-30 | 2018-11-30 | Cross-domain recommendation method and device based on user behavior sequence characteristics |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109544306B (en) |
Families Citing this family (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110060097A (en) * | 2019-04-01 | 2019-07-26 | 苏州市职业大学 | User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks |
CN110347916B (en) * | 2019-06-11 | 2022-03-29 | 北京三快在线科技有限公司 | Cross-scene item recommendation method and device, electronic equipment and storage medium |
US10970350B2 (en) | 2019-08-15 | 2021-04-06 | Advanced New Technologies Co., Ltd. | Method and apparatus for processing user interaction sequence data |
CN110598847B (en) * | 2019-08-15 | 2020-08-28 | 阿里巴巴集团控股有限公司 | Method and device for processing interactive sequence data |
CN110837598B (en) * | 2019-11-11 | 2021-03-19 | 腾讯科技(深圳)有限公司 | Information recommendation method, device, equipment and storage medium |
CN111026970B (en) * | 2019-12-20 | 2022-07-29 | 腾讯科技(深圳)有限公司 | Artificial intelligence based migration model training method and device and storage medium |
CN111241394B (en) * | 2020-01-07 | 2023-09-22 | 腾讯科技(深圳)有限公司 | Data processing method, data processing device, computer readable storage medium and electronic equipment |
CN111258469B (en) * | 2020-01-09 | 2021-05-14 | 支付宝(杭州)信息技术有限公司 | Method and device for processing interactive sequence data |
CN112035743B (en) * | 2020-08-28 | 2021-10-15 | 腾讯科技(深圳)有限公司 | Data recommendation method and device, computer equipment and storage medium |
CN111966914B (en) * | 2020-10-26 | 2021-01-15 | 腾讯科技(深圳)有限公司 | Content recommendation method and device based on artificial intelligence and computer equipment |
CN112417293B (en) * | 2020-12-03 | 2024-06-18 | 京东科技控股股份有限公司 | Information pushing method and system, model training method and related equipment |
CN112541132B (en) * | 2020-12-23 | 2023-11-10 | 北京交通大学 | Cross-domain recommendation method based on multi-view knowledge representation |
CN112801718B (en) * | 2021-02-22 | 2021-10-01 | 平安科技(深圳)有限公司 | User behavior prediction method, device, equipment and medium |
CN112948696B (en) * | 2021-03-31 | 2022-10-18 | 青岛科技大学 | Cross-domain medical care equipment recommendation method and system with privacy protection function |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP2860672A2 (en) * | 2013-10-10 | 2015-04-15 | Deutsche Telekom AG | Scalable cross domain recommendation system |
CN105447145A (en) * | 2015-11-25 | 2016-03-30 | 天津大学 | Item-based transfer learning recommendation method and recommendation apparatus thereof |
CN106250545A (en) * | 2016-08-10 | 2016-12-21 | Tcl集团股份有限公司 | A kind of multimedia recommendation method and system searching for content based on user |
CN106897464A (en) * | 2017-03-29 | 2017-06-27 | 广东工业大学 | A kind of cross-cutting recommendation method and system |
CN107609116A (en) * | 2017-09-13 | 2018-01-19 | 星环信息科技(上海)有限公司 | A kind of method and apparatus for creating cross-domain Transfer Depth network |
CN107895000A (en) * | 2017-10-30 | 2018-04-10 | 昆明理工大学 | A kind of cross-cutting semantic information retrieval method based on convolutional neural networks |
CN108563690A (en) * | 2018-03-15 | 2018-09-21 | 中山大学 | A kind of collaborative filtering recommending method based on object-oriented cluster |
CN108763493A (en) * | 2018-05-30 | 2018-11-06 | 深圳市思迪信息技术股份有限公司 | A kind of recommendation method based on deep learning |
CN108763321A (en) * | 2018-05-02 | 2018-11-06 | 深圳智能思创科技有限公司 | A kind of related entities recommendation method based on extensive related entities network |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140074650A1 (en) * | 2012-03-01 | 2014-03-13 | Qloo, Inc. | Personalized cross-domain recommender system |
-
2018
- 2018-11-30 CN CN201811455642.5A patent/CN109544306B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP2860672A2 (en) * | 2013-10-10 | 2015-04-15 | Deutsche Telekom AG | Scalable cross domain recommendation system |
CN105447145A (en) * | 2015-11-25 | 2016-03-30 | 天津大学 | Item-based transfer learning recommendation method and recommendation apparatus thereof |
CN106250545A (en) * | 2016-08-10 | 2016-12-21 | Tcl集团股份有限公司 | A kind of multimedia recommendation method and system searching for content based on user |
CN106897464A (en) * | 2017-03-29 | 2017-06-27 | 广东工业大学 | A kind of cross-cutting recommendation method and system |
CN107609116A (en) * | 2017-09-13 | 2018-01-19 | 星环信息科技(上海)有限公司 | A kind of method and apparatus for creating cross-domain Transfer Depth network |
CN107895000A (en) * | 2017-10-30 | 2018-04-10 | 昆明理工大学 | A kind of cross-cutting semantic information retrieval method based on convolutional neural networks |
CN108563690A (en) * | 2018-03-15 | 2018-09-21 | 中山大学 | A kind of collaborative filtering recommending method based on object-oriented cluster |
CN108763321A (en) * | 2018-05-02 | 2018-11-06 | 深圳智能思创科技有限公司 | A kind of related entities recommendation method based on extensive related entities network |
CN108763493A (en) * | 2018-05-30 | 2018-11-06 | 深圳市思迪信息技术股份有限公司 | A kind of recommendation method based on deep learning |
Non-Patent Citations (4)
Title |
---|
"Exploiting Hierarchical Structures for POI Recommendation";Pengpeng Zhao;《2017 IEEE International Conference on Data Mining (ICDM)》;20171121;第655-664页 * |
"基于上下文感知和个性化度量嵌入的下一个兴趣点推荐";鲜学丰 等;《计算机工程与科学》;20180415;第616-625页 * |
"基于深度学习的推荐系统研究综述";黄立威;《计算机学报》;20180305;第41卷(第7期);第1619-1647页 * |
跨领域推荐技术综述;陈雷慧 等;《华东师范大学学报(自然科学版)》;20171006(第5期);第101-116+137页 * |
Also Published As
Publication number | Publication date |
---|---|
CN109544306A (en) | 2019-03-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109544306B (en) | Cross-domain recommendation method and device based on user behavior sequence characteristics | |
CN108648049B (en) | Sequence recommendation method based on user behavior difference modeling | |
CN111061946B (en) | Method, device, electronic equipment and storage medium for recommending scenerized content | |
CN111966914B (en) | Content recommendation method and device based on artificial intelligence and computer equipment | |
CN109582956A (en) | text representation method and device applied to sentence embedding | |
CN110032679B (en) | Dynamic news recommendation method based on hierarchical attention network | |
CN113590900A (en) | Sequence recommendation method fusing dynamic knowledge maps | |
CN116249991A (en) | Neural network distillation method and device | |
Huynh et al. | Context-similarity collaborative filtering recommendation | |
CN113761359B (en) | Data packet recommendation method, device, electronic equipment and storage medium | |
Wang et al. | POI recommendation method using LSTM-attention in LBSN considering privacy protection | |
CN113609337A (en) | Pre-training method, device, equipment and medium of graph neural network | |
Shiguemori et al. | Estimation of initial condition in heat conduction by neural network | |
CN117216281A (en) | Knowledge graph-based user interest diffusion recommendation method and system | |
CN113259163B (en) | Web service quality prediction method and system based on network topology perception | |
WO2023174064A1 (en) | Automatic search method, automatic-search performance prediction model training method and apparatus | |
Zhao et al. | CapDRL: a deep capsule reinforcement learning for movie recommendation | |
Lee et al. | A Web-based collaborative filtering system | |
CN116361643A (en) | Model training method for realizing object recommendation, object recommendation method and related device | |
George et al. | Hy-MOM: Hybrid recommender system framework using memory-based and model-based collaborative filtering framework | |
CN115564532A (en) | Training method and device of sequence recommendation model | |
Wang et al. | Well production real-time intelligent monitoring based on convolutional neural network | |
CN116932862A (en) | Cold start object recommendation method, cold start object recommendation device, computer equipment and storage medium | |
CN110659962B (en) | Commodity information output method and related device | |
Samih et al. | Deep graph embeddings for content based-book recommendations |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
TR01 | Transfer of patent right |
Effective date of registration: 20220811 Address after: 215000 floor 7, North building, Guofa building, No. 3118 Renmin Road, Gusu District, Suzhou City, Jiangsu Province Patentee after: Suzhou big data Co.,Ltd. Patentee after: Suzhou Big Data Trading Service Co.,Ltd. Patentee after: Suzhou big data Research Institute Co.,Ltd. Address before: No. 8, Xiangcheng District Ji Xue Road, Suzhou, Jiangsu Patentee before: SOOCHOW University |
|
TR01 | Transfer of patent right |