Cross-domain recommendation method, device, equipment and medium based on multi-task learning

Abstract

The present invention discloses a cross-domain recommendation method, device, equipment and medium based on multi-task learning. The method determines the category of the user according to the user's personal information, and the categories include product domain users, content domain users and public domain users; extracts individual attribute information of the product domain user category to form a product domain user vector, and extracts individual attribute information of the public domain user category to form a public domain user vector; inputs the product domain user vector, the public domain user vector, the product information of historical consumption, the user access information of historical consumption and the target consumption product embedding vector into a trained cross-domain recommendation model, and outputs the probability of the user consuming each target consumption product; recommends the target consumption product to the user according to the probability ranking of the user consuming each target consumption product. The present application can improve the prediction accuracy of the probability of the user consuming the target product or the user subscribing to the target content, and effectively alleviate the problems of data sparsity and user cold start.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a cross-domain recommendation method, device, equipment and medium based on multi-task learning. Background Art

[0002] In the existing technology, personalized recommendation systems help users identify their potential interests and needs, and can better discover long-tail interests, and have been widely used. Recommendation algorithms are the core part of recommendation systems. The main methods are collaborative filtering (CF)-based and content-based methods. Among them, CF-based recommendation methods use the interaction data between users and items, such as user ratings. The content-based method discovers the correlation between items and content based on the metadata of items or content, and then recommends similar items to users based on the user's previous preference records. Currently, the CF-based method is widely used because it can capture the user's most intuitive evaluation of items and can be easily extended to multiple scenarios, while the content-based method requires a certain amount of effort in different fields to build a suitable feature set for recommendation. However, the CF-based method currently has two problems:

[0003] 1) Data is sparse, that is, the user rating matrix contains a large number of null values;

[0004] 2) Cold start: there is a lack of sufficient historical evaluation data for new users.

[0005] In reality, we often face the cold start and data sparsity problems of new users or new products, and cannot accurately model users to achieve personalized recommendations, so cross-domain recommendation methods have emerged. Cross-domain recommendation uses relatively richer information from richer domains to alleviate the data sparsity and cold start problems of single-domain recommendations and improve the accuracy of recommendations in sparse domains. At present, there are many technical implementation methods for cross-domain recommendation, such as matrix decomposition-based methods, methods that use transfer learning to migrate information from the source domain to the target domain, etc., but the recommendation accuracy of these methods is not ideal and needs to be further improved. Summary of the invention

[0006] Based on this, it is necessary to provide a cross-domain recommendation method, device, equipment and medium based on multi-task learning to address the above technical problems, so as to solve the problem of low recommendation accuracy of the cross-domain recommendation method in the prior art.

[0007] Based on the above purpose, a solution 1 of a cross-domain recommendation method based on multi-task learning includes:

[0008] Obtaining the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information;

[0009] According to the personal information of the user, the category of the user is determined, and the category of the user includes product domain user, content domain user, and public domain user; extracting individual attribute information of the product domain user category to form a product domain user vector, and extracting individual attribute information of the public domain user category to form a public domain user vector;

[0010] The product domain user vector, the public domain user vector, the historical consumption product information, the historical consumption user access information and the set target consumption product embedding vector are input into the trained cross-domain recommendation model, and the probability of the user consuming each target consumption product is output; the target consumption product embedding vector is composed of each target consumption product;

[0011] According to the probability ranking of the user consuming each target consumption product, the target consumption product is recommended to the user.

[0012] Based on the above purpose, a second solution of a cross-domain recommendation method based on multi-task learning includes:

[0013] Obtaining the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information;

[0014] According to the personal information of the user, the category of the user is determined, and the category of the user includes product domain users, content domain users, and public domain users; individual attribute information of the content domain users is extracted to form a content domain user vector, and individual attribute information of the public domain users is extracted to form a public domain user vector;

[0015] Input the content domain user vector, the public domain user vector, the historically subscribed reading content information, the historically subscribed user access information, and the set target subscription content embedding vector into the trained cross-domain recommendation model, and output the probability of the user subscribing to each target subscription content; the target subscription content embedding vector is composed of each target subscription content;

[0016] According to the probability ranking of the user subscribing to each target subscription content, the target subscription content is recommended to the user.

[0017] Based on the above purpose, a solution 1 of a cross-domain recommendation device based on multi-task learning includes:

[0018] An information collection module is used to obtain the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information;

[0019] A user identification module is used to determine the category of the user according to the user's personal information, wherein the category of the user includes product domain user, content domain user, and public domain user; extract individual attribute information of the product domain user category to form a product domain user vector, and extract individual attribute information of the public domain user category to form a public domain user vector;

[0020] A cross-domain recommendation model is used to input the product domain user vector, the public domain user vector, the historically consumed product information, the historically consumed user access information, and a set target consumption product embedding vector, and output the probability of the user consuming each target consumption product;

[0021] The recommendation module is used to recommend target consumer products to users based on the probability ranking of the users consuming the target consumer products.

[0022] Based on the above purpose, a second solution of a cross-domain recommendation device based on multi-task learning includes:

[0023] An information collection module is used to obtain the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information;

[0024] A user identification module is used to determine the category of the user according to the user's personal information, wherein the category of the user includes product domain users, content domain users, and public domain users; extract individual attribute information of content domain users to form a content domain user vector, and extract individual attribute information of public domain users to form a public domain user vector;

[0025] A cross-domain recommendation model is used to input the content domain user vector, the public domain user vector, the historically subscribed reading content information, the historically subscribed user access information, and the set target subscription content embedding vector, and output the probability of the user subscribing to each target subscription content;

[0026] The recommendation module is used to recommend target subscription content to the user according to the probability ranking of the user subscribing to each target subscription content.

[0027] Based on the above purpose, a technical solution of a computer device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor. When the processor executes the computer-readable instructions, the above-mentioned cross-domain recommendation method based on multi-task learning is implemented.

[0028] Based on the above purpose, one or more readable storage media storing computer-readable instructions, when the computer-readable instructions are executed by one or more processors, enable the one or more processors to execute the cross-domain recommendation method as described above.

[0029] The beneficial effects of the above-mentioned cross-domain recommendation method, device, computer equipment and storage medium based on multi-task learning are: the present application analyzes the user's personal information, distinguishes the user's behavioral characteristics, classifies and characterizes the user information according to the user's behavioral characteristics, and classifies the user into product domain users, content domain users and public domain users, and applies the user's classification characterization to the cross-domain recommendation model, that is, the relevant information extracted into the categories of product domain users and public domain users is input into the cross-domain recommendation model to predict the probability of the user consuming the target product; the relevant information extracted into the categories of content domain users and public domain users is input into the cross-domain recommendation model to predict the probability of the user subscribing to the target content; the present application can improve the prediction accuracy of the probability of the user consuming the target product or the user subscribing to the target content, and effectively alleviate the problems of data sparsity and user cold start. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0031] Figure 1 is a schematic diagram of an application environment of a cross-domain recommendation method based on multi-task learning in one embodiment of the present invention;

[0032] Figure 2 is a flow chart of a cross-domain recommendation method based on multi-task learning in one embodiment of the present invention;

[0033] Figure 3 is a schematic diagram of the structure of a first task recommendation model in one embodiment of the present invention;

[0034] Figure 4 is a schematic diagram of the structure of the first attention layer I1 in one embodiment of the present invention;

[0035] Figure 5 is a schematic diagram of the structure of a second task recommendation model in one embodiment of the present invention;

[0036] Figure 6 is a flow chart of a cross-domain recommendation method based on multi-task learning in one embodiment of the present invention;

[0037] Figure 7 is a structural diagram of a cross-domain recommendation device based on multi-task learning in one embodiment of the present invention;

[0038] Figure 8 is a structural diagram of a cross-domain recommendation device based on multi-task learning in one embodiment of the present invention;

[0039] Fig. 9 is a schematic diagram of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] The cross-domain recommendation method based on multi-task learning provided in this embodiment can be applied to Figure 1 In an application environment, a client communicates with a server. The client includes but is not limited to various personal computers, laptops, smart phones, tablet computers, and portable wearable devices. The server can be implemented with an independent server or a server cluster consisting of multiple servers.

[0042] In one embodiment, if Figure 2 As shown in the figure, a cross-domain recommendation method based on multi-task learning is provided. Figure 1 The server in the example is used as an example to illustrate the following steps:

[0043] S10: Obtaining the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information.

[0044] In this step, historically consumed products refer to insurance products that users have historically purchased on specific pages in the APP (such as the insurance list page), and historically subscribed reading content refers to articles, videos, health headlines and other content resources that users have historically browsed on relevant content pages in the APP.

[0045] In this step, the user access information of historical consumption refers to the transaction time, transaction product type, etc. when the user has purchased certain products, and the user access information of historical subscription refers to the browsing time, browsing times, etc. when the user has browsed certain content.

[0046] It is understandable that in this step, the user's personal information is stored in a set user information database, and the user information database dynamically updates the user's historical consumption product information, historical subscription reading content information, and scene context information in real time.

[0047] Optionally, the user information database is a database storing real identities of users, and a process for establishing the user information database is as follows:

[0048] (1) The user applies for an identity number from the identity information verification system through the communication network; the identity information verification system assigns an identity number to the user through the communication network and prompts the user to submit identity information;

[0049] (2) The user submits various identity information to the identity verification system via the communication network; the identity information verification system temporarily stores the identity information submitted by the user and verifies the authenticity of the identity information submitted by the user based on the different types of information;

[0050] (3) After the identity authenticity verification is passed, the identity information verification system stores the temporarily stored identity information in the user information database and associates the identity pass number with the identity information, thereby establishing a database that can verify the user's true identity through the identity pass number.

[0051] Among them, the personal information submitted by the user includes the user's communication number information and the user's certificate information. Among them, the user's communication number information includes the user's mobile phone number, landline number, email address, home address, QQ number, MSN number, various network account numbers of the user and other information; the user's certificate information includes bank card, ID card, academic certificate and other information.

[0052] S20: Determine the category of the user according to the personal information of the user, the category of the user including product domain user, content domain user, and public domain user; extract individual attribute information of the product domain user to form a product domain user vector, and extract individual attribute information of the public domain user to form a public domain user vector.

[0053] It is understandable that for a certain user, the user may have more than one category. For example, the user has purchased insurance products and subscribed to some reading content for browsing. The user's behavioral characteristics meet the three categories of users. Therefore, the categories are product domain users, content domain users, and public domain users. If a user has only purchased insurance products and has not subscribed to any reading content, the user's behavioral characteristics only meet the category of content domain users, and the user's category is content domain users. Similarly, if a user has only subscribed to some reading content for browsing and has not purchased insurance products, the user's behavioral characteristics only meet the category of product domain users, and the user's category is product domain users.

[0054] The reason why this step judges the category of users is that the ultimate goal of the recommendation method described in this embodiment is to recommend target products (such as insurance products). Target product recommendations and target reading content recommendations are usually regarded as two independent recommendation scenarios, but a considerable number of users have interacted in both scenarios. For example, a user has purchased insurance products and browsed articles and other content resources. This part of users becomes public domain users.

[0055] It can be understood that after classifying each user, the individual attribute information of the product domain users is composed into a product domain user vector, and each element in the vector is the individual attribute information under this category; similarly, the individual attribute information of the public domain users is composed into a public domain user vector, and each element in the public domain user vector is the individual attribute information under this category.

[0056] S30: Input the product domain user vector, the public domain user vector, the historical consumption product information, the historical consumption user access information and the set target consumption product embedding vector into the trained cross-domain recommendation model, and output the probability of the user consuming each target consumption product; the target consumption product embedding vector is composed of each target consumption product.

[0057] In this step, the cross-domain recommendation model adopts a multi-task learning network structure, including a first task recommendation model and a second task recommendation model. The two recommendation models need to be trained simultaneously. The structure of the first task recommendation model is as follows: Figure 3 As shown, it includes a first attention layer I1, a first pooling layer I2, a first gated network layer I3 and a first fully connected layer I4 which are connected in sequence.

[0058] Among them, the first attention layer I1 is used to input the target consumption product embedding vector, the product domain user vector, the historical consumption product vector composed of the historical consumption product information, and the first context vector composed of the historical consumption user access information in the scene context information; the first attention layer is used to output the weighted value of each historical consumption product.

[0059] Specifically, the structure of the first attention layer I1 is as follows: Figure 4 As shown, it includes a mixed product module U1, an inner product module U2 and a scalar addition module U3, wherein the mixed product module U1 is used to calculate the vector mixed product of the product domain user vector, the first context vector and the target consumption product embedding vector, the product domain user vector is obtained by stacking the embedding vectors of the user's individual attribute information; the first context vector is obtained by stacking the embedding vectors of the user's historical consumption access information in the scene context information.

[0060] The inner product module U2 is used to input the target consumption product embedding vector and the historical consumption product vector, and then calculate the inner product between the target consumption product embedding vector and the historical consumption product vector, where the historical consumption product vector is obtained based on the product information of historical consumption.

[0061] The scalar addition module U3 is used to input the vector mixed product obtained by the mixed product module U1 and the inner product obtained by the inner product module U2, and calculate the sum of the vector mixed product and the inner product to obtain the weighted value of each historical consumption product. For example, there are N historical consumption products, and through the first attention layer I1, the scalar addition module U3 finally outputs the weighted values ​​w01, w02, ..., w0n of the N historical consumption products.

[0062] In the first task recommendation model, the first pooling layer I2 is used to perform weighted summation on the historical consumption product vectors according to the weighted values ​​of the various historical consumption products to obtain a first sum-pooled vector.

[0063] In the first task recommendation model, the first gating network layer I3 is used to update the product domain user vector according to the first sum pooling vector to obtain an updated product domain user vector.

[0064] It can be understood that the structure of the first gated network layer I3 is a threshold fusion subnetwork in the prior art, which is used to input the target consumer product embedding vector, the first context vector, the first and pooled vector, the product domain user vector and the public domain user vector, and by determining the different importance weights of each vector, each vector is linearly weighted summed with the corresponding importance weight to obtain the weight of the product domain user, and the product domain user weight is used to update the product domain user vector, that is, the calculated product domain user weight is superimposed on the corresponding element values ​​in the product domain user vector to obtain the updated product domain user vector.

[0065] The first fully connected layer I4 is used to input the target consumer product embedding vector, the first context vector, the first and pooling vectors, the updated product domain user vector and the public domain user vector, where the public domain user vector is a user vector classified as a public domain user, and the first fully connected layer I4 is used to output the probability of the user consuming each target consumer product.

[0066] The structure of the second task recommendation model is as follows Figure 5 As shown, it includes a second attention layer P1, a second pooling layer P2, a second gated network layer P3 and a second fully connected layer P4 connected in sequence. Among them, the second attention layer P1 is used to input the set target subscription content embedding vector, the content domain user vector is composed of the individual attribute information of the content domain user, the historical subscription content vector is composed of the historical subscription reading content information, and the second context vector is composed of the historical subscription user access information; the second attention layer is used to output the weighted value of each historical subscription content.

[0067] The structure of the second attention layer P1 is the same as that of the first attention layer I1. Figure 4 As shown, the second attention layer P1 also includes a mixed product module, an inner product module and a scalar addition module. The mixed product module is used to calculate the vector mixed product of the content domain user vector, the second context vector and the target subscribed content embedding vector. The content domain user vector is obtained by stacking the embedding vectors of the user's individual attribute information; the second context vector is obtained by stacking the embedding vectors of the user's access information who has historically subscribed to the reading content in the scene context information.

[0068] The inner product module of the second attention layer P1 is used to input the target subscription content embedding vector and the historical subscription content vector, and calculate the inner product between the target consumption product embedding vector and the historical consumption product vector. The historical subscription content vector is obtained based on the reading content information of the historical subscription.

[0069] The scalar addition module of the second attention layer P1 is used to input the vector mixed product and the inner product, calculate the sum of the vector mixed product and the inner product, and obtain the weighted value of each historical subscription content. For example, among N historical subscription contents, through the second attention layer P1, the scalar addition module finally outputs the weighted values ​​w11, w12, ..., w1n of the N historical subscription contents.

[0070] In the second task recommendation model, the second pooling layer P2 is used to perform weighted summation on the historical subscription content vectors according to the weighted values ​​of the various historical subscription contents to obtain a second sum pooling vector.

[0071] In the second task recommendation model, the second gated network layer P3 is used to update the content domain user vector according to the second sum pooling vector to obtain an updated content domain user vector. Specifically, similar to the first gated network layer I3, the structure of the second gated network layer P3 is also a threshold fusion subnetwork in the prior art, except that the second gated network layer P3 is used to input the target subscription content embedding vector, the second context vector, the second sum pooling vector, the content domain user vector and the public domain user vector, determine the different importance weights of each vector, perform linear weighted summation of each vector and the corresponding importance weight, obtain the weight of the content domain user, superimpose the weight of the content domain user to each element in the content domain user vector, update the content domain user vector, and obtain the updated content domain user vector.

[0072] In the second task recommendation model, the second fully connected layer P4 is used to input the target subscription content embedding vector, the second context vector, the second and pooling vector, the updated content domain user vector and the public domain user vector, and the second fully connected layer P4 is used to output the probability of the user subscribing to each target subscription content.

[0073] According to the network structure of the above cross-domain recommendation model, the network training process is as follows:

[0074] Extract the required user information from the above user information database, including the user's individual attribute information, historical product information, historical subscribed reading content information and scene context information, and construct the target label and sample data. The method for determining the target label is: according to the user's log data in the user information database, set labels for the consumed products and subscribed content. Set to 1 for unconsumed products and unsubscribed content. is 0.

[0075] The sample data is divided into appropriate training sets and test sets. The training set is trained using the above cross-domain recommendation model, and the test set is tested. The model hyperparameters are automatically adjusted during the training process until the cross-domain recommendation model can fit the training set well and achieve the expected accuracy in the test set.

[0076] The joint loss function during model training is:

[0077]

[0078] Among them, L(θ1,…,θ K ,θ s ) is the joint loss function, K is the number of tasks, K = 2, ω k is the weight hyperparameter of each task, θ s is a shared parameter, θ k The unique parameter of the corresponding task, ω k ,θ s ,θ k Determined by the model training process. For a specific task k, the corresponding loss L k (θ k ,θ s )for:

[0079]

[0080] Among them, L k (θ k ,θ s ) is the loss function of task k, θ s is a shared parameter, θ k Is a unique parameter of the corresponding task, loss k is the cross entropy loss function, is a binary label indicating whether the current sample belongs to the task k. k = 1 indicates that the task is to predict the user's consumption of the target product. k = 2 indicates that the task is to predict the user's subscription to the target content. i indicates the sample number. is the predicted value (i.e., probability value) output by the model, is the label of the sample.

[0081] S40: Recommending target consumption products to the user based on the probability ranking of the user consuming each target consumption product.

[0082] In this step, after predicting the probability of the user consuming each target consumer product, the target consumer products are sorted according to the probability value, and consumption recommendations for the target products are made based on this sorting.

[0083] It is understandable that since the ultimate goal of the recommendation method of this embodiment is to recommend the target product, the user's personal information may contain less product information about historical consumption and user access information about historical consumption. Therefore, it is necessary to use the public domain user vector and update the product domain user vector in combination with the first gated network layer to highlight the characteristic differences between users in each product domain and realize reasonable recommendation of target products for different users.

[0084] In one embodiment, a cross-domain recommendation method based on multi-task learning is provided. The method can recommend target subscription content to users by using the above-trained cross-domain recommendation model, such as Figure 6 As shown, the method comprises the following steps:

[0085] S11: Obtaining the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information.

[0086] S21: Determine the category of the user according to the personal information of the user, the category of the user including product domain user, content domain user, and public domain user; extract individual attribute information of the content domain user category to form a content domain user vector, and extract individual attribute information of the public domain user category to form a public domain user vector.

[0087] It can be understood that after classifying each user, the individual attribute information of the content domain user category is composed into a content domain user vector, and each element in the vector is the individual attribute information under this category; similarly, the individual attribute information of the public domain user category is composed into a public domain user vector, and each element in the public domain user vector is the individual attribute information under this category.

[0088] S31: Input the content domain user vector, the public domain user vector, the historically subscribed reading content information, the historically subscribed user access information, and the set target subscription content embedding vector into a trained cross-domain recommendation model, and output the probability of the user subscribing to each target subscription content; the target subscription content embedding vector is composed of each target subscription content.

[0089] It can be understood that this step uses the second task recommendation model in the trained cross-domain recommendation model to predict the probability of users subscribing to each target subscription content. Figure 5 As shown, it includes a second attention layer P1, a second pooling layer P2, a second gated network layer P3 and a second fully connected layer P4 connected in sequence. The structure of the second task recommendation model refers to the relevant records in the aforementioned embodiment, and will not be repeated in this embodiment.

[0090] It should be noted that the difference between this step and step S30 of the aforementioned embodiment lies in the input data and the purpose of implementation. In step S30, the first task recommendation model in the cross-domain recommendation model is used, and the input data are the product domain user vector, the public domain user vector, the historical consumption product information, the historical consumption user access information and the set target consumption product embedding vector, and what is predicted is the probability of the user consuming each target consumption product; while in this step, the second task recommendation model in the cross-domain recommendation model is used, and the input data are the content domain user vector, the public domain user vector, the historical subscription reading content information, the historical subscription user access information and the set target subscription content embedding vector, and what is predicted is the probability of the user subscribing to each target subscription content.

[0091] S41: Recommending target subscription content to the user according to the probability ranking of the user subscribing to each target subscription content.

[0092] In this step, after predicting the probability of the user subscribing to each target subscription content, the target subscription content is sorted according to the probability value, and the target subscription content is recommended based on the sorting.

[0093] It is understandable that since the ultimate goal of the recommendation method of this embodiment is to recommend target subscription content, the user's personal information may contain less information about historically subscribed reading content and historically subscribed user access information. Therefore, it is necessary to use the public domain user vector and combine the second gated network layer to update the content domain user vector, highlight the characteristic differences between users in each content domain, and achieve reasonable recommendation of target subscription content for different users.

[0094] This application obtains a cross-domain recommendation model based on a multi-task learning method, including a first task recommendation model and a second task recommendation model. Combined with the user's personal information and the user's category, it can more accurately predict the probability of the user consuming the target product or the user subscribing to the target content. According to the probability of the user consuming the target product or the user subscribing to the target content, the target product or target content with a larger probability value is recommended first.

[0095] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0096] In one embodiment, a cross-domain recommendation device based on multi-task learning is provided, and the cross-domain recommendation device based on multi-task learning corresponds one-to-one to the cross-domain recommendation method based on multi-task learning in the first embodiment. Figure 7As shown, the cross-domain recommendation device includes an information collection module L1, a user identification module L2, a product domain recommendation module L31 and a first ranking recommendation module L41. The functional modules are described in detail as follows:

[0097] The information collection module L1 is used to obtain the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information.

[0098] The user identification module L2 is used to determine the category of the user based on the user's personal information, and the user categories include product domain users, content domain users, and public domain users; extract individual attribute information of product domain users to form a product domain user vector, and extract individual attribute information of public domain users to form a public domain user vector.

[0099] The product domain recommendation module L31 is used to input the product domain user vector, the public domain user vector, the historical consumption product information, the historical consumption user access information and the set target consumption product embedding vector, and output the probability of the user consuming each target consumption product.

[0100] The first ranking and recommendation module L41 is used to recommend target consumer products to users according to the probability ranking of the users consuming the target consumer products.

[0101] It can be understood that the product domain recommendation module L31 in this device corresponds to the aforementioned first task recommendation model, and the product domain recommendation module L31 includes a first attention layer, a first pooling layer, a first gated network layer and a first fully connected layer connected in sequence.

[0102] Among them, the first attention layer is used to input the target consumption product embedding vector, the product domain user vector, the historical consumption product vector composed of the historical consumption product information, and the first context vector composed of the historical consumption user access information; the first attention layer is used to output the weighted value of each historical consumption product.

[0103] The first pooling layer is used to perform weighted summation on the historical consumption product vectors according to the weighted values ​​of the various historical consumption products to obtain a first sum pooling vector; the first gated network layer is used to update the product domain user vector according to the first sum pooling vector to obtain an updated product domain user vector.

[0104] The first fully connected layer is used to input the target consumer product embedding vector, the first context vector, the first and pooling vector, the updated product domain user vector and the public domain user vector, and the first fully connected layer is used to output the probability of the user consuming each target consumer product.

[0105] In one embodiment, a cross-domain recommendation device based on multi-task learning is provided, and the cross-domain recommendation device based on multi-task learning corresponds one-to-one to the cross-domain recommendation method based on multi-task learning in the second embodiment. Figure 8 As shown, the cross-domain recommendation device includes an information collection module L1, a user identification module L2, a content domain recommendation module L32, and a second ranking recommendation module L42. The functional modules are described in detail as follows:

[0106] The information collection module L1 is used to obtain the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information.

[0107] The user identification module L2 is used to determine the category of the user according to the user's personal information, and the user categories include product domain users, content domain users, and public domain users; extract individual attribute information of content domain users to form a content domain user vector, and extract individual attribute information of public domain users to form a public domain user vector.

[0108] The content domain recommendation module L32 is used to input the content domain user vector, the public domain user vector, the historically subscribed reading content information, the historically subscribed user access information and the set target subscription content embedding vector, and output the probability of the user subscribing to each target subscription content.

[0109] The second ranking recommendation module L42 is used to recommend target subscription content to the user according to the probability ranking of the user's subscription to each target subscription content.

[0110] It can be understood that the content domain recommendation module L32 in the present device corresponds to the aforementioned second task recommendation model, and the content domain recommendation module L32 includes a second attention layer, a second pooling layer, a second gated network layer and a second fully connected layer connected in sequence.

[0111] Among them, the second attention layer is used to input the target subscription content embedding vector, the content domain user vector, the historical subscription content vector composed of the reading content information of the historical subscription, and the second context vector composed of the user access information of the historical subscription; the second attention layer is used to output the weighted value of each historical subscription content.

[0112] The second pooling layer is used to perform weighted summation on the historical subscription content vectors according to the weighted values ​​of the various historical subscription contents to obtain a second sum pooling vector; the second gated network layer is used to update the content domain user vector according to the second sum pooling vector to obtain an updated content domain user vector.

[0113] The second fully connected layer is used to input the target subscription content embedding vector, the second context vector, the second and pooling vector, the updated content domain user vector and the public domain user vector, and the second fully connected layer is used to output the probability of the user subscribing to each target subscription content.

[0114] For the specific definition of the cross-domain recommendation device, please refer to the definition of the cross-domain recommendation method above, which will not be repeated here. Each module in the above two cross-domain recommendation devices can be implemented in whole or in part by software, hardware and a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0115] The present application analyzes the user's personal information, distinguishes the user's behavioral characteristics, classifies and represents the user information according to the user's behavioral characteristics, and classifies the user into product domain users, content domain users, and public domain users. The user's classification representation is applied to the cross-domain recommendation model, that is, the relevant information extracted into the categories of product domain users and public domain users is input into the cross-domain recommendation model to predict the probability of the user consuming the target product; the relevant information extracted into the categories of content domain users and public domain users is input into the cross-domain recommendation model to predict the probability of the user subscribing to the target content; the present application can improve the prediction accuracy of the probability of the user consuming the target product or the user subscribing to the target content, and effectively alleviate the problems of data sparsity and user cold start.

[0116] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig. 9As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a readable storage medium and an internal memory. The readable storage medium stores an operating system, computer-readable instructions and a database. The internal memory provides an environment for the operation of the operating system and computer-readable instructions in the readable storage medium. The database of the computer device is used to store data involved in the cross-domain recommendation method. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer-readable instructions are executed by the processor, a cross-domain recommendation method based on multi-task learning is implemented. The readable storage medium provided in this embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.

[0117] In one embodiment, one or more computer-readable storage media storing computer-readable instructions are provided, and the readable storage media provided in this embodiment include non-volatile readable storage media and volatile readable storage media. The readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the cross-domain recommendation method based on multi-task learning is implemented.

[0118] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through computer-readable instructions, and the computer-readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When the computer-readable instructions are executed, they may include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0119] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0120] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A cross-domain recommendation method based on multi-task learning, characterized in that: The following steps are involved: Obtaining the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scenario context information; the scenario context information includes historical consumption user access information, historical subscription user access information, the individual attribute information includes communication number information and user certificate information, the user's communication number information includes the user's mobile phone number, landline phone number, email address, home address, QQ number, MSN number, and various network account numbers of the user; the user's certificate information includes bank card, ID card, and academic certificate; According to the personal information of the user, the category of the user is determined, and the category of the user includes product domain user, content domain user, and public domain user; extracting individual attribute information of the product domain user category to form a product domain user vector, and extracting individual attribute information of the public domain user category to form a public domain user vector; The product domain user vector, the public domain user vector, the historical consumption product information, the historical consumption user access information and the set target consumption product embedding vector are input into the trained cross-domain recommendation model, and the probability of the user consuming each target consumption product is output; the target consumption product embedding vector is composed of each target consumption product; Recommending target consumer products to users based on the probability ranking of the users consuming the target consumer products; The cross-domain recommendation model adopts a multi-task learning network structure, including a first task recommendation model, which includes a first attention layer, a first pooling layer, a first gated network layer and a first fully connected layer connected in sequence; The first attention layer is used to input the target consumption product embedding vector, the product domain user vector, the historical consumption product vector composed of the historical consumption product information, and the first context vector composed of the historical consumption user access information; the first attention layer is used to output the weighted value of each historical consumption product; The first pooling layer is used to perform weighted summation on the historical consumption product vectors according to the weighted values ​​of the various historical consumption products to obtain a first sum pooling vector; the first gating network layer is used to update the product domain user vector according to the first sum pooling vector to obtain an updated product domain user vector; The first fully connected layer is used to input the target consumer product embedding vector, the first context vector, the first sum pooling vector, the updated product domain user vector and the public domain user vector, and the first fully connected layer is used to output the probability of the user consuming each target consumer product.

2. The cross-domain recommendation method based on multi-task learning as claimed in claim 1, characterized in that: The first attention layer includes a mixed product module, an inner product module and a scalar addition module. The mixed product module is used to calculate the vector mixed product of the updated product domain user vector, the first context vector and the target consumption product embedding vector; the inner product module is used to input the target consumption product embedding vector and the historical consumption product vector, and calculate the inner product between the target consumption product embedding vector and the historical consumption product vector; the scalar addition module is used to input the vector mixed product and the inner product, and calculate the sum of the vector mixed product and the inner product to obtain the weighted value of each historical consumption product.

3. A cross-domain recommendation method based on multi-task learning, characterized in that: The following steps are involved: Obtaining the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scenario context information; the scenario context information includes historical consumption user access information, historical subscription user access information, the individual attribute information includes communication number information and user certificate information, the user's communication number information includes the user's mobile phone number, landline phone number, email address, home address, QQ number, MSN number, and various network account numbers of the user; the user's certificate information includes bank card, ID card, and academic certificate; According to the personal information of the user, the category of the user is determined, and the category of the user includes product domain users, content domain users, and public domain users; individual attribute information of the content domain users is extracted to form a content domain user vector, and individual attribute information of the public domain users is extracted to form a public domain user vector; Input the content domain user vector, the public domain user vector, the historically subscribed reading content information, the historically subscribed user access information, and the set target subscription content embedding vector into the trained cross-domain recommendation model, and output the probability of the user subscribing to each target subscription content; the target subscription content embedding vector is composed of each target subscription content; According to the probability ranking of the user's subscription to each target subscription content, the target subscription content is recommended to the user; the cross-domain recommendation model adopts a multi-task learning network structure, including a second task recommendation model, which includes a second attention layer, a second pooling layer, a second gating network layer and a second fully connected layer connected in sequence; The second attention layer is used to input the target subscription content embedding vector, the content domain user vector, the historical subscription content vector composed of the reading content information of the historical subscription, and the second context vector composed of the user access information of the historical subscription; the second attention layer is used to output the weighted value of each historical subscription content; The second pooling layer is used to perform weighted summation on the historical subscription content vectors according to the weighted values ​​of the various historical subscription contents to obtain a second sum pooling vector; the second gating network layer is used to update the content domain user vector according to the second sum pooling vector to obtain an updated content domain user vector; The second fully connected layer is used to input the target subscription content embedding vector, the second context vector, the second sum pooling vector, the updated content domain user vector and the public domain user vector, and the second fully connected layer is used to output the probability of the user subscribing to each target subscription content.

4. The cross-domain recommendation method based on multi-task learning as claimed in claim 3, characterized in that: The second attention layer includes a mixed product module, an inner product module and a scalar addition module. The mixed product module is used to calculate the vector mixed product of the updated content domain user vector, the second context vector and the target subscription content embedding vector; the inner product module is used to input the target subscription content embedding vector and the historical subscription content vector, and calculate the inner product between the target subscription content embedding vector and the historical subscription content vector; the scalar addition module is used to input the vector mixed product and the inner product, and calculate the sum of the vector mixed product and the inner product to obtain the weighted value of each historical subscription content.

5. A cross-domain recommendation device based on multi-task learning, characterized in that: include: The information collection module is used to obtain the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information; the individual attribute information includes communication number information and user certificate information; the user's communication number information includes the user's mobile phone number, landline phone number, email address, home address, QQ number, MSN number, and various network account numbers of the user; the user's certificate information includes bank card, ID card, and academic certificate; A user identification module is used to determine the category of the user according to the user's personal information, wherein the category of the user includes product domain user, content domain user, and public domain user; extract individual attribute information of the product domain user to form a product domain user vector, and extract individual attribute information of the public domain user to form a public domain user vector; The product domain recommendation module is used to input the product domain user vector, the public domain user vector, the historical consumption product information, the historical consumption user access information and the set target consumption product embedding vector into the trained cross-domain recommendation model, and output the probability of the user consuming each target consumption product; the target consumption product embedding vector is composed of each target consumption product; A first ranking and recommendation module, used to recommend target consumer products to users according to the probability ranking of the users consuming the target consumer products; The cross-domain recommendation model adopts a multi-task learning network structure, including a first task recommendation model, which includes a first attention layer, a first pooling layer, a first gated network layer and a first fully connected layer connected in sequence; The first attention layer is used to input the target consumption product embedding vector, the product domain user vector, the historical consumption product vector composed of the historical consumption product information, and the first context vector composed of the historical consumption user access information; the first attention layer is used to output the weighted value of each historical consumption product; The first pooling layer is used to perform weighted summation on the historical consumption product vectors according to the weighted values ​​of the various historical consumption products to obtain a first sum pooling vector; the first gating network layer is used to update the product domain user vector according to the first sum pooling vector to obtain an updated product domain user vector; The first fully connected layer is used to input the target consumer product embedding vector, the first context vector, the first sum pooling vector, the updated product domain user vector and the public domain user vector, and the first fully connected layer is used to output the probability of the user consuming each target consumer product.

6. A cross-domain recommendation device based on multi-task learning, characterized in that: include: The information collection module is used to obtain the user's personal information, including the user's individual attribute information, historical consumption product information, historical subscription reading content information, and scene context information; the scene context information includes historical consumption user access information and historical subscription user access information; the individual attribute information includes communication number information and user certificate information; the user's communication number information includes the user's mobile phone number, landline phone number, email address, home address, QQ number, MSN number, and various network account numbers of the user; the user's certificate information includes bank card, ID card, and academic certificate; A user identification module is used to determine the category of the user according to the user's personal information, wherein the category of the user includes product domain users, content domain users, and public domain users; extract individual attribute information of content domain users to form a content domain user vector, and extract individual attribute information of public domain users to form a public domain user vector; The content domain recommendation module is used to input the content domain user vector, the public domain user vector, the historically subscribed reading content information, the historically subscribed user access information, and the set target subscription content embedding vector into the trained cross-domain recommendation model, and output the probability of the user subscribing to each target subscription content; the target subscription content embedding vector is composed of each target subscription content; A second ranking and recommendation module, configured to recommend target subscription content to the user according to the probability ranking of the user subscribing to each target subscription content; The cross-domain recommendation model adopts a multi-task learning network structure, including a second task recommendation model, which includes a second attention layer, a second pooling layer, a second gating network layer and a second fully connected layer connected in sequence; The second attention layer is used to input the target subscription content embedding vector, the content domain user vector, the historical subscription content vector composed of the reading content information of the historical subscription, and the second context vector composed of the user access information of the historical subscription; the second attention layer is used to output the weighted value of each historical subscription content; The second pooling layer is used to perform weighted summation on the historical subscription content vectors according to the weighted values ​​of the various historical subscription contents to obtain a second sum pooling vector; the second gating network layer is used to update the content domain user vector according to the second sum pooling vector to obtain an updated content domain user vector; The second fully connected layer is used to input the target subscription content embedding vector, the second context vector, the second sum pooling vector, the updated content domain user vector and the public domain user vector, and the second fully connected layer is used to output the probability of the user subscribing to each target subscription content.

7. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, characterized in that: When the computer-readable instructions are executed by the processor, the cross-domain recommendation method based on multi-task learning is implemented as described in any one of claims 1 to 4.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the cross-domain recommendation method based on multi-task learning is implemented.

Images