Material recommendation method, material recommendation device and electronic equipment

Abstract

The present application discloses a material recommendation method, a material recommendation device and an electronic device, which belongs to the field of computer technology. The method comprises: the electronic device obtains the first attribute information of the target cold start user; the electronic device inputs the first attribute information into the first cold start recall model to obtain the first vector corresponding to the target cold start user, the first cold start recall model is a model obtained based on the second cold start recall model, the second cold start recall model is a model obtained by the server based on the user data set and trained using a meta-learning method, the first vector is used to characterize the preference of the target cold start user; the electronic device obtains the first material matching the first vector and displays the first material.

Description

Technical Field

[0001] The present application belongs to the field of computer technology, and specifically relates to a material recommendation method, a material recommendation device and an electronic device. Background Art

[0002] Recommendation systems capture users' interests by learning massive interactions between users and items, and recommend items of interest to users. They are widely used in e-commerce, video, or music applications. However, for cold-start users who have no or few clicks on items, it is difficult for recommendation systems to capture their true preferences.

[0003] The problem that related technologies can only capture the interests of active users but cannot accurately recommend materials of interest to cold-start users needs to be solved urgently. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a material recommendation method, a material recommendation device and an electronic device, which can solve the problem in the related art that the interests of active users can only be captured but materials of interest cannot be recommended to cold-start users.

[0005] In a first aspect, an embodiment of the present application provides a material recommendation method, the method comprising:

[0006] The electronic device obtains first attribute information of a target cold start user;

[0007] The electronic device inputs the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on a second cold start recall model, the second cold start recall model is a model obtained by a server based on a user data set and trained using a meta-learning method, and the first vector is used to characterize the preference of the target cold start user;

[0008] The electronic device obtains a first material matching the first vector and displays the first material.

[0009] In a second aspect, an embodiment of the present application provides a material recommendation method, the method comprising:

[0010] The server receives a first request message sent by the electronic device, where the first request message carries a first vector corresponding to a target cold start user, where the first vector is used to represent a preference of the target cold start user;

[0011] The server obtains a first material based on the first vector and a second cold start recall model local to the server;

[0012] The server sends a first response message to the electronic device, wherein the first response message carries the first material;

[0013] The second cold start recall model is a model trained by the server based on the user data set using a meta-learning method.

[0014] In a third aspect, an embodiment of the present application provides a material recommendation device, the device comprising:

[0015] An acquisition unit, used for acquiring first attribute information of a target cold start user;

[0016] a first output unit, configured to input the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on a second cold start recall model, wherein the second cold start recall model is a model obtained by a server based on a user data set and trained using a meta-learning method, and the first vector is used to characterize a preference of the target cold start user;

[0017] The recommendation unit is configured to obtain a first material matching the first vector and display the first material.

[0018] In a fourth aspect, an embodiment of the present application provides a material recommendation device, the device comprising:

[0019] A second receiving unit, configured to receive a first request message sent by an electronic device, wherein the first request message carries a first vector corresponding to a target cold start user, and the first vector is used to represent a preference of the target cold start user;

[0020] A second output unit, configured to obtain a first material based on the first vector and a second local cold start recall model of the server;

[0021] A first sending unit, configured to send a first response message to the electronic device, wherein the first response message carries the first material;

[0022] The second cold start recall model is a model trained by the server based on the user data set using a meta-learning method.

[0023] In a fifth aspect, an embodiment of the present application provides an electronic device, which includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method described in the first aspect or the second aspect.

[0024] In a sixth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method described in the first aspect or the second aspect are implemented.

[0025] In the seventh aspect, an embodiment of the present application provides a chip, which includes a processor and a communication interface, the communication interface is coupled to the processor, and the processor is used to run a program or instruction to implement the method described in the first aspect or the second aspect.

[0026] In an embodiment of the present application, the first attribute information of the target cold start user is input into the first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on the second cold start recall model, and the second cold start recall model is a model obtained based on the user data set and trained using a meta-learning method. Therefore, the first vector can characterize the preference of the target cold start user, and then by obtaining the first material matching the first vector, it is possible to quickly capture the interest of the cold start user and improve the accuracy and real-time performance of personalized recommendations for cold start users. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a structural schematic diagram of a double tower model provided by the relevant technology;

[0028] Figure 2 It is a schematic diagram of the interactive process of the material recommendation method provided in the embodiment of the present application;

[0029] Figure 3 This is one of the flow charts of the material recommendation method provided in the embodiment of the present application;

[0030] Figure 4 It is a structural schematic diagram of a first cold start recall model or a second cold start recall model or a second benchmark model or a meta-learning model provided in an embodiment of the present application;

[0031] Figure 5 This is the second flow chart of the material recommendation method provided in the embodiment of the present application;

[0032] Figure 6 It is a structural schematic diagram of an improved double-tower model or a first reference model provided in an embodiment of the present application;

[0033] Figure 7 is a schematic diagram of a process of training to obtain a first benchmark model provided in an embodiment of the present application;

[0034] Figure 8 It is a flowchart of training a meta-learning model provided in an embodiment of the present application;

[0035] Fig. 9 This is one of the structural schematic diagrams of the material recommendation device provided in the embodiment of the present application;

[0036] Fig.10 This is the second structural schematic diagram of the material recommendation device provided in the embodiment of the present application;

[0037] Fig.11 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0038] Fig.12 It is a hardware schematic diagram of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.

[0040] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0041] The material recommendation method, material recommendation device, electronic device and readable storage medium provided in the embodiments of the present application are described in detail below with reference to specific embodiments and their application scenarios in conjunction with the accompanying drawings.

[0042] In order to achieve the recommendation effect for cold start users, the traditional dual-tower model is used in related technologies, such as Figure 1 As shown, the method in the related art mainly includes: inputting the user's attribute information and the material's attribute information into the embedding layer 101 of the dual-tower model respectively to obtain the user attribute embedding vector and the material attribute embedding vector; inputting the user attribute embedding vector into the user tower 102 of the dual-tower model to obtain the user attribute vector, and inputting the material attribute embedding vector into the material tower 103 of the dual-tower model to obtain the material attribute vector; then performing inner product operation on the user attribute vector and the material attribute vector to obtain the similarity calculation result between the user and the material, and performing material recommendation based on the similarity calculation result.

[0043] It should be noted that the traditional dual-tower model utilizes cold-start user information to a certain extent to alleviate the ability to capture the interests of cold-start users, but this method does not make improvements for cold-start users. In essence, it still captures the interests of active users and cannot effectively improve the accuracy of recommendation results for cold-start users.

[0044] The present application embodiment provides a material recommendation method, such as Figure 2 As shown, the material recommendation method includes the following execution steps:

[0045] Step 200: The server trains to obtain a second cold start recall model;

[0046] Step 201: The server deploys the second cold start recall model to the electronic device;

[0047] Step 202: The electronic device obtains a first cold start recall model based on the second cold start recall model, obtains first attribute information of a target cold start user, inputs the first attribute information into the first cold start recall model, and obtains a first vector corresponding to the target cold start user;

[0048] Step 203: The electronic device sends a first request message to the server, where the first request message carries a first vector;

[0049] Step 204: The server receives the first request message sent by the electronic device, and obtains the first material based on the first vector and the second cold start recall model of the server.

[0050] Step 205: The server sends a first response message to the electronic device, where the first response message carries the first material;

[0051] Step 206: The electronic device receives the first response message sent by the server, obtains the first material, and displays the first material.

[0052] The material recommendation method provided in the embodiment of the present application can be executed by an electronic device or a functional module or functional entity in the electronic device that can implement the material recommendation method. The electronic devices mentioned in the embodiment of the present application include but are not limited to mobile phones, tablet computers, computers, wearable devices, etc. The material recommendation method provided in the embodiment of the present application is described below using the electronic device as an example of the execution subject. It should be noted that the above execution subject does not constitute a limitation on the present application.

[0053] like Figure 3 As shown, the material recommendation method includes: step 300, step 301 and step 302.

[0054] Step 300: The electronic device obtains first attribute information of a target cold start user.

[0055] It should be noted that this embodiment can be applied to application scenarios of application software such as e-commerce, video or music, but does not constitute a limitation to the present application.

[0056] In the embodiment of the present application, the target cold start user is a new user who has just registered the application software, or a user who has no behavior data or less behavior data in the application software.

[0057] Optionally, the first attribute information is user attribute information, which is used to characterize relevant attributes of the target cold start user, such as age, gender or region.

[0058] It is understandable that, when the target cold start user uses the application software in the electronic device, the electronic device obtains the first attribute information of the target cold start user, such as age, gender or region.

[0059] Step 301: The electronic device inputs the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user.

[0060] The first cold start recall model is a model obtained based on the second cold start recall model, and the second cold start recall model is a model trained by the server based on the user data set and using a meta-learning method.

[0061] It can be understood that the server deploys the second cold start recall model to the electronic device, and the electronic device obtains the first cold start recall model according to the second cold start recall model.

[0062] The second cold start recall model is trained locally on the server based on the user dataset using a meta-learning method.

[0063] The user data set includes: active user data set and cold start user data set.

[0064] Meta-learning is a method of learning to learn, which aims to use the information that has been learned to quickly adapt to new tasks that have not been learned. Meta-learning can quickly capture user interests and learn users' real-time preferences more quickly.

[0065] Furthermore, the second cold start recall model is obtained by training the improved dual-tower model locally on the server based on the user data set and using a meta-learning method. The improved dual-tower model adds a user ID vector generator on the basis of the dual-tower model, which can utilize the basic attributes of the user and improve the model expression ability of the second cold start recall model.

[0066] The embodiment of the present application obtains a second cold start recall model by training using a meta-learning method based on a user data set, and obtains a first cold start recall model based on the second cold start recall model, so that the first cold start recall model can capture the interests of active users and cold start users at the same time. Compared with related technologies, it can improve the accuracy of recommending materials to cold start users.

[0067] It should be noted that the second cold start recall model is used to determine the corresponding to-be-recommended material according to the user attribute information. The first cold start recall model and the second cold start recall model have the same function.

[0068] Optionally, determining the corresponding material to be recommended according to the user attribute information includes: generating a user vector according to the user attribute information, performing similarity calculation between the user vector and a plurality of material vectors generated based on material attributes, and determining the corresponding material recommendation result according to the similarity calculation result.

[0069] In this embodiment, the first attribute information is input into the first cold start recall model to obtain a first vector corresponding to the target cold start user, and the first vector can reflect the preference of the target cold start user.

[0070] Step 302: The electronic device obtains a first material matching the first vector, and displays the first material.

[0071] Optionally, the electronic device obtains the first material that matches the first vector from the service. If a large number of material attributes are stored locally in the electronic device, a large amount of memory space in the electronic device will be occupied. Therefore, the electronic device in the embodiment of the present application inputs the first attribute information into the first cold start recall model to obtain the first vector corresponding to the target cold start user, and then obtains the first material that matches the first vector from the server and displays the first material. In this way, the electronic device does not need to store a large number of material attributes locally, which can save the memory space occupied by the electronic device and improve resource utilization.

[0072] Optionally, the electronic device acquires a first material matching the first vector, including:

[0073] The electronic device sends a first request message to the server, where the first request message carries a first vector;

[0074] The electronic device receives the first response message sent by the server and obtains the first material.

[0075] It can be understood that the first request message is used to request the server to return the material recommendation result corresponding to the target cold start user based on the first vector. After receiving the first request message sent by the electronic device, the server obtains the first material based on the first vector and the second cold start recall model of the server, and sends a first response message to the electronic device, and the first response message carries the first material.

[0076] After the electronic device acquires the first material, it displays the first material, thereby completing the recommendation of the material to the target cold start user.

[0077] In an embodiment of the present application, the first attribute information of the target cold start user is input into the first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on the second cold start recall model, and the second cold start recall model is a model obtained based on the user data set and trained using a meta-learning method. Therefore, the first vector can characterize the preference of the target cold start user, and then by obtaining the first material matching the first vector, it is possible to quickly capture the interest of the cold start user and improve the accuracy and real-time performance of personalized recommendations for cold start users.

[0078] In some embodiments, the first cold start recall model is a model obtained based on a second cold start recall model deployed by the server on the electronic device, including one of the following:

[0079] The first cold start recall model is the same as the second cold start recall model;

[0080] The first cold start recall model is a model obtained by the electronic device training the second cold start recall model based on the real-time behavior data of the target cold start user;

[0081] The first cold start recall model is a model obtained by the electronic device updating the second cold start recall model based on the update parameters sent by the server.

[0082] It can be understood that there are three ways to obtain the first cold start recall model.

[0083] Method 1: The server deploys the second cold start recall model on the electronic device as the first cold start recall model.

[0084] Method 2: In order to effectively improve the accuracy and real-time performance of cold start user recommendations, the electronic device obtains the real-time behavior data of the target cold start user, builds a training data set based on the real-time behavior data, and uses the training data set to train the second cold start recall model. After the training is completed, the first cold start recall model is obtained.

[0085] Method three, in order to further improve the accuracy and real-time performance of cold start user recommendations, the server obtains parameters of multiple cold start recall models by training the second cold start recall model based on the real-time behavior data of the target cold start user using multiple electronic devices through federated learning, and processes these parameters (for example, taking the average) to obtain updated parameters, and then sends the updated parameters to the electronic device, which updates the second cold start recall model based on the updated parameters to obtain the first cold start recall model.

[0086] In an embodiment of the present application, three different methods are provided for an electronic device to obtain a first cold start recall model based on a second cold start recall model deployed on the electronic device by a server. The first cold start recall model has a strong ability to capture the interests of cold start users, thereby improving the accuracy of personalized recommendations for cold start users.

[0087] In some embodiments, the method further comprises:

[0088] The electronic device obtains real-time behavior data corresponding to the target cold start user, and constructs a training data set based on the real-time behavior data;

[0089] The electronic device trains the second cold start recall model using the training data set to obtain the first cold start recall model.

[0090] It can be understood that the real-time behavior data of the target cold start user includes the click behavior data of the target cold start user on at least one material.

[0091] Optionally, constructing a training data set based on the real-time behavior data includes:

[0092] For target cold start users in different states, the electronic device constructs positive samples and negative samples respectively based on the real-time behavior data of the target cold start users, and merges the positive samples and negative samples to obtain the sample set of the target cold start users, and obtains a training data set for each sample of the target cold start user in the form of "user attribute information-first category material attribute information-second category material attribute information".

[0093] The first material attribute information is different from the second material attribute information.

[0094] The present application provides the following methods for constructing a training data set:

[0095] In one implementation, for a target cold start user with less real-time behavior data, when the target cold start user generates real-time behavior data, the real-time behavior data of the target cold start user is used as a positive sample; the exposure behavior data of the target cold start user and materials similar to the materials corresponding to the exposure behavior data are randomly sampled to obtain negative samples.

[0096] In one implementation, when the target cold start user does not generate real-time behavior data, the attribute information of the material is clustered according to the first attribute information of the target cold start user, and based on the attribute information of the clustered material, the material with higher click behavior data is randomly sampled to obtain a positive sample; the exposure behavior data of the target cold start user and the material similar to the material corresponding to the exposure behavior data are randomly sampled to obtain a negative sample.

[0097] Among them, the exposure behavior data represents data that the target cold start user did not generate real-time behavior on some of the materials recommended by the application software on the electronic device. For example, if the target cold start user did not perform a click operation on some of the materials recommended by the application software, then these materials will be used as exposure behavior data.

[0098] In another implementation, for target cold start users without real-time behavior data, the attribute information of the materials is clustered according to the first attribute information of the target cold start users, and based on the attribute information of the clustered materials, materials with higher click behavior data are randomly sampled to obtain positive samples; based on the attribute information of the clustered materials, materials with lower click behavior data are randomly sampled to obtain negative samples.

[0099] In an embodiment of the present application, a training data set is constructed based on the real-time behavior data of the target cold start user, and the second cold start recall model is trained using the training data set to obtain the first cold start recall model. The first vector is obtained using the first attribute information and the first cold start recall model, which can further improve the accuracy and real-time performance of personalized recommendations for cold start users.

[0100] In some embodiments, the method further comprises:

[0101] The electronic device receives the update parameters sent by the server;

[0102] The electronic device updates the second cold start recall model based on the update parameter to obtain the first cold start recall model;

[0103] The update parameters are obtained by the server through federated learning to obtain the parameters of the first cold start recall model obtained by training multiple electronic devices.

[0104] In an embodiment of the present application, in order to further improve the accuracy and real-time performance of cold start user recommendations, the server obtains update parameters by means of federated learning, and the electronic device updates the second cold start recall model based on the update parameters to obtain the first cold start recall model.

[0105] It can be understood that multiple electronic devices train the second cold start recall model based on the real-time behavior data of the target cold start user to obtain multiple first cold start recall models; the server obtains the parameters of the multiple first cold start recall models through federated learning, and processes the parameters of the multiple first cold start recall models (for example, mean operation) to obtain updated parameters.

[0106] In an embodiment of the present application, the second cold start recall model is updated by the update parameters sent by the server to obtain the first cold start recall model, and then the first cold start recall model is used to process the target cold start user to obtain the first vector, thereby further improving the accuracy and real-time performance of the cold start user recommendation.

[0107] like Figure 4 As shown, the structure of the first cold start recall model includes an embedding layer 401, a user ID vector generator 402, a user tower 403 and a material tower 404.

[0108] The embedding layer 401 is used to convert user attributes and material attributes into user attribute embedding vectors and material attribute embedding vectors respectively.

[0109] The user ID vector generator 402 is used to generate a user ID embedding vector based on the user attribute embedding vector, and then concatenate the user ID embedding vector with the user attribute embedding vector to obtain a user embedding vector.

[0110] User tower 403 is used to generate a user vector based on the user embedding vector.

[0111] The material tower 404 is used to generate a material vector based on the material attribute embedding vector.

[0112] In some embodiments, the electronic device inputs the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user, including:

[0113] Inputting the first attribute information into the embedding layer 401 of the first cold start recall model to obtain a first embedding vector;

[0114] Inputting the first embedding vector into the user ID vector generator 402 of the first cold start recall model to obtain a second embedding vector;

[0115] Concatenate the first embedding vector and the second embedding vector to obtain a third embedding vector;

[0116] The third embedding vector is input into the user tower 403 of the first cold start recall model to obtain the first vector.

[0117] In this step, when the first attribute information of the target cold start user is input into the first cold start recall model, the first vector can be obtained through the embedding layer, the user ID vector generator and the user tower.

[0118] It should be noted that in the embodiment of the present application, the user tower 403 and the material tower 404 are arbitrary network structures, for example, the network structure is a multi-layer perceptron, wherein the output result of the user tower and the output result of the material tower have the same dimension.

[0119] The first embedding vector is used to represent the first attribute information of the target cold start user.

[0120] The second embedding vector is a user ID vector corresponding to the first attribute information.

[0121] The third embedding vector is a high-dimensional vector obtained by concatenating the first embedding vector and the second embedding vector.

[0122] In an embodiment of the present application, by inputting the first attribute information of the target cold start user into the first cold start recall model to obtain the first vector, it is possible to quickly capture the interest of the cold start user and improve the accuracy and real-time performance of personalized recommendations for the cold start user.

[0123] The material recommendation method provided in the embodiment of the present application can also be executed by a server. The material recommendation method provided in the embodiment of the present application is described below by taking the server as the execution subject. It should be noted that the above execution subject does not constitute a limitation to the present application.

[0124] like Figure 5 As shown, the material recommendation method includes: step 500, step 501 and step 502.

[0125] Step 500: The server receives a first request message sent by an electronic device, where the first request message carries a first vector corresponding to a target cold start user.

[0126] In this step, the first request message is used to request the server to return the material recommendation result corresponding to the target cold start user based on the first vector.

[0127] It can be understood that the material recommendation result is the first material.

[0128] Step 501: The server obtains a first material based on the first vector and a second cold start recall model local to the server.

[0129] Optionally, the first material is a material recommended by application software in the electronic device to the target cold start user, such as a commodity, video or music.

[0130] It can be understood that after receiving the first request message sent by the electronic device, the server obtains the first vector, inputs the first vector into the second cold start recall model of the server locally, and obtains the first material.

[0131] The second cold start recall model is a model trained by the server based on the user data set using a meta-learning method.

[0132] The user data set includes: active user data set and cold start user data set.

[0133] Meta-learning is a method of learning to learn, which aims to use the information that has been learned to quickly adapt to new tasks that have not been learned. Meta-learning can quickly capture user interests and learn users' real-time preferences more quickly.

[0134] Furthermore, the second cold start recall model is obtained by training the improved dual-tower model locally on the server based on the user data set and using a meta-learning method. The improved dual-tower model adds a user ID vector generator on the basis of the dual-tower model, which can utilize the basic attributes of the user and improve the model expression ability of the second cold start recall model.

[0135] The embodiment of the present application obtains a second cold start recall model by training based on a user data set using a meta-learning method. The second cold start recall model can capture the interests of active users and cold start users at the same time. Compared with related technologies, it can improve the accuracy of recommending materials to cold start users.

[0136] It should be noted that, in this step, the second cold start recall model is used to determine the corresponding materials to be recommended according to the user attribute information.

[0137] Optionally, determining the corresponding material to be recommended according to the user attribute information includes: generating a user vector according to the user attribute information, performing similarity calculation between the user vector and a plurality of material vectors generated based on material attributes, and determining the corresponding material recommendation result according to the similarity calculation result.

[0138] Step 502: The server sends a first response message to the electronic device, wherein the first response message carries the first material.

[0139] In this step, the first response message is a response message of the server to the first request message, and is used to return a material recommendation result matching the first vector to the electronic device.

[0140] It can be understood that after the server obtains the first material based on the first vector and the local second cold start recall model, it sends a first response message to the electronic device, wherein the first response message carries the first material.

[0141] In an embodiment of the present application, the server obtains a first vector through a first request message, obtains a first material based on the first vector and a local second cold start recall model, and returns the first material to the electronic device, thereby quickly capturing the interests of cold start users and improving the accuracy and real-time performance of personalized recommendations for cold start users.

[0142] The structure of the second cold start recall model includes: an embedding layer, a user ID vector generator, a user tower, and a material tower. The structure of the second cold start recall model is the same as that of the first cold start recall model. Figure 4 , the various layers included in the second cold start recall model will not be described in detail here.

[0143] In some embodiments, obtaining the first material based on the first vector and the second cold start recall model locally on the server includes:

[0144] Obtaining N material vectors output by the material tower of the second cold start recall model;

[0145] Perform inner product calculations on the first vector and the N material vectors respectively to obtain N calculation results, wherein the calculation results are used to characterize the similarity between the target cold start user and the material corresponding to the first vector;

[0146] Sorting the N calculation results in descending order, obtaining materials corresponding to the top M calculation results, and obtaining the first material;

[0147] Wherein, M and N are both positive integers greater than 1.

[0148] In an embodiment of the present application, the server obtains attribute information of multiple stored materials, inputs the attribute information of the multiple materials into the embedding layer of the second cold start recall model to obtain multiple material attribute embedding vectors, inputs the multiple material attribute embedding vectors into the material tower of the second cold start recall model to obtain N material vectors, and then performs inner product calculations on the first vector and the N material vectors respectively to obtain N similarities, sorts the N similarities in descending order, and takes the materials corresponding to the top M similarities as the first materials.

[0149] In an embodiment of the present application, multiple material vectors are obtained through the second cold start recall model, the similarity between the target cold start user and the material is obtained based on the first vector and the multiple material vectors, and then the first material is obtained based on the similarity, thereby quickly capturing the interest of the cold start user and improving the accuracy and real-time performance of personalized recommendations for cold start users.

[0150] In some embodiments, the method further comprises:

[0151] The second cold start recall model is obtained by training, and the second cold start recall model is deployed to the electronic device.

[0152] It should be noted that the second cold start recall model is obtained by training with the cold start user data set as training samples. The second cold start recall model is used to recommend materials that the target cold start user is interested in to the target cold start user.

[0153] Furthermore, the server deploys the second cold start recall model to the electronic device, so that the electronic device obtains the first vector according to the first attribute information of the target cold start user, thereby recommending materials according to the real-time behavior data of the target cold start user.

[0154] In an embodiment of the present application, by deploying the trained second cold start recall model to an electronic device, it is possible to quickly capture the interests of cold start users and improve the accuracy and real-time performance of personalized recommendations for cold start users.

[0155] The step of training to obtain the second cold start recall model includes:

[0156] Constructing a user data set, and using the user data set to train the improved dual-tower model to obtain a first benchmark model;

[0157] Improving the first benchmark model to obtain a second benchmark model;

[0158] constructing an active user data set, and using the active user data set to train the second benchmark model to obtain a meta-learning model;

[0159] A cold start user data set is constructed, and the meta-learning model is trained using the cold start user data set to obtain the second cold start recall model.

[0160] It should be noted that since the user behavior data corresponding to the user data set is sufficient, first using the user data set to train the first benchmark model can produce better training results. Since the cold start user data set corresponds to less behavior data of cold start users, directly using the cold start user data set to train the first benchmark model cannot achieve the accuracy of cold start user personalized recommendation. Therefore, in the embodiment of the present application, the active user data set is first used to train the meta-learning model to improve the training effect, and then the cold start user data set is used to train the meta-learning model to obtain the second cold start recall model, which can effectively improve the accuracy and real-time performance of cold start user personalized recommendation.

[0161] The improved twin-tower model includes the improvement of input data and the improvement of the operation process in the twin-tower model.

[0162] The improvement of input data is that the input end of the embedding layer includes user ID, user attribute information and material attribute information.

[0163] The improvement of the operation process in the dual-tower model includes adding an operation of concatenating the user ID and the user attribute information between the output end of the embedding layer and the input end of the user tower.

[0164] The embodiment of the present application improves the structure of the first benchmark model by adding a hierarchical structure of a user ID vector generator between the output end of the embedding layer and the input end of the user tower, and retains the other layers of structures and the parameters corresponding to each layer of structure, such as the embedding matrix of user attributes and material attributes, and the parameters of the mapping functions of the user tower and the material tower, to obtain the second benchmark model.

[0165] like Figure 6 As shown, the structure of the improved dual-tower model or the first benchmark model includes: an embedding layer 601, a user tower 602 and a material tower 603, wherein the embedding layer 601 is used to convert the user ID, user attributes and material attributes into a user ID embedding vector, a user attribute embedding vector and a material attribute embedding vector respectively; the user ID embedding vector and the user attribute embedding vector are concatenated to obtain a user embedding vector, the user tower 602 is used to generate a user vector based on the user embedding vector, and the material tower 603 is used to generate a material vector based on the material attribute embedding vector.

[0166] The structure of the second baseline model or the meta-learning model or the second cold-start recall model is Figure 4The structure of the first cold start recall model shown is the same, including: an embedding layer, a user ID vector generator, a user tower and a material tower, wherein the embedding layer is used to convert user attributes and material attributes into user attribute embedding vectors and material attribute embedding vectors respectively, and the user ID vector generator is used to generate a user ID embedding vector based on the user attribute embedding vector; then the user ID embedding vector and the user attribute embedding vector are concatenated to obtain a user embedding vector; the user tower is used to generate a user vector based on the user embedding vector, and the material tower is used to generate a material vector based on the material attribute embedding vector.

[0167] It should be noted that the user tower and the material tower can be any network structure, for example, the network structure is a multi-layer perceptron, wherein the output result of the user tower and the output result of the material tower have the same dimension.

[0168] In an embodiment of the present application, a first baseline model is obtained by training with a user data set, the first baseline model is improved to obtain a second baseline model, the second baseline model is trained with an active user data set to obtain a meta-learning model, and finally, the meta-learning model is trained with a cold-start user data set to obtain a second cold-start recall model. This realizes a model that captures the interests of cold-start users and performs personalized recommendations, thereby effectively improving the accuracy and real-time performance of personalized recommendations for cold-start users.

[0169] The present application embodiment also provides a flow chart of training to obtain a first benchmark model, such as Figure 7 As shown, the user data set is constructed, and the improved dual-tower model is trained using the user data set to obtain a first benchmark model, including:

[0170] Step 700: Take the user's click behavior data as a positive sample, randomly sample the materials that have not been clicked by each user among the users, obtain negative samples, and obtain the user data set based on the positive samples and negative samples, wherein each sample in the user data set includes user ID, user attribute information and material attribute information.

[0171] Optionally, the users are all users registered in the application software on the electronic device.

[0172] Optionally, the label corresponding to the positive sample is represented as 1, and the label corresponding to the negative sample is represented as 0. The positive sample and the negative sample are merged, and each merged sample is constructed according to the user ID, user attribute information and material attribute information to obtain a user data set.

[0173] In some embodiments, the user attribute information represents the statistical attribute information of the user, such as age, gender or region, and the material attribute information includes basic attributes such as the category or keywords of the material.

[0174] Step 701: Take any sample in the user data set as the current sample, input the current sample into the embedding layer of the improved twin-tower model, and obtain the first user ID embedding vector, the first user attribute embedding vector and the first material attribute embedding vector corresponding to the current sample.

[0175] Step 702: Concatenate the first user ID embedding vector and the first user attribute embedding vector to obtain a first user embedding vector.

[0176] Step 703: input the first user embedding vector into the user tower of the improved twin-tower model to obtain a first user vector, and input the material attribute embedding vector into the material tower of the improved twin-tower model to obtain a first material vector.

[0177] Step 704: perform an inner product operation on the first user vector and the first material vector to obtain a first similarity calculation result, and calculate the loss using a cross entropy loss function based on the first similarity calculation result and the label of the current sample to obtain a first loss value.

[0178] The first similarity calculation result is used to characterize the similarity between the user and the material in the current sample.

[0179] In some embodiments, the cross entropy loss function is:

[0180] loss=-(y*log(sigmoid(logits))+(1-y)*log(1-sigmoid(logits)) (1)

[0181] Among them, loss represents the loss value, y represents the label, and logits represents the similarity.

[0182] Step 705: Use the first loss value to update the parameters of the improved twin-tower model.

[0183] Step 706: input other samples in the user data set one by one into the improved dual-tower model after parameter update to continue training until the training end condition is met, end the training, save the parameters of the improved dual-tower model at the end of the training, and obtain the first benchmark model.

[0184] In an embodiment of the present application, a first benchmark model is obtained by sequentially inputting the constructed user data set into the embedding layer, user tower and material tower of the improved dual-tower model for training, and a second cold-start recall model is obtained based on the first benchmark model, thereby achieving rapid capture of the interests of cold-start users and improving the accuracy and real-time performance of personalized recommendations for cold-start users.

[0185] The present application embodiment also provides a flow chart of training a meta-learning model, such as Figure 8 As shown, the active user data set is constructed, and the second benchmark model is trained using the active user data set to obtain a meta-learning model, including:

[0186] Step 800: Screen the user data set according to the user's activity level to obtain an active user data set, wherein each sample in the active user data set is constructed according to user attribute information, first-category material attribute information, and second-category material attribute information.

[0187] Optionally, the user's activity level can be obtained based on the user's usage of application software on the electronic device within a preset time period.

[0188] For example, based on the user's usage within one month after registering the application software, if the user does not generate behavioral data of clicking on materials, then the user is a cold start user; if the user generates a large amount of behavioral data of clicking on materials, then the user is an active user.

[0189] Optionally, the user's activity level is determined based on the frequency of the user's use of the application software on the electronic device within a preset time period. For example, if the user's use frequency of the application software on the electronic device within the preset time period is greater than or equal to a preset threshold, the user is determined to be an active user, and if the user's use frequency of the application software on the electronic device within the preset time period is less than the preset threshold, the user is determined to be a cold start user.

[0190] In this step, the filtered active user data set is aggregated with material attribute information according to user attribute information to obtain first-category material attribute information and second-category material attribute information respectively, and each sample in the active user data set is constructed in the form of user attribute information, first-category material attribute information and second-category material attribute information.

[0191] Among them, the first category of material attribute information is different from the second category of material attribute information.

[0192] In this embodiment, samples in the reference meta-learning paradigm include a support set and a query set, and the first type of material attribute information is used as the support set, and the second type of material attribute information is used as the query set.

[0193] Step 801: Take any sample in the active user data set as the current sample, input the user attribute information, first-category material attribute information, and second-category material attribute information of the current sample into the embedding layer of the second benchmark model, and obtain the second user attribute embedding vector, first-category material attribute embedding vector, and second-category material attribute embedding vector corresponding to the current sample.

[0194] The structure of the second benchmark model refers to Figure 4 .

[0195] Step 802: Input the second user attribute embedding vector into the user ID vector generator of the second benchmark model to obtain the second user ID embedding vector corresponding to the current sample.

[0196] Step 803: Concatenate the second user ID embedding vector and the second user attribute embedding vector to obtain a second user embedding vector.

[0197] Step 804: input the second user embedding vector into the user tower of the second benchmark model to obtain a second user vector, input the first type of material attribute embedding vector into the material tower of the second benchmark model to obtain a second material vector, and input the second type of material attribute embedding vector into the material tower of the second benchmark model to obtain a third material vector.

[0198] Step 805: perform an inner product operation on the second user vector and the second material vector to obtain a second similarity calculation result, and based on the second similarity calculation result and the label of the current sample, calculate the loss using a cross entropy loss function to obtain a second loss value.

[0199] The second similarity calculation result is used to characterize the similarity between the active users in the current sample and the first category of materials.

[0200] Step 806: Use the second loss value to update the second user ID embedding vector, and concatenate the updated second user ID embedding vector with the second user attribute embedding vector to obtain a third user embedding vector.

[0201] In some embodiments, updating the second user ID embedding vector using the second loss value includes:

[0202] Based on formula (2), the second user ID embedding vector is updated using the second loss value:

[0203]

[0204] in, represents the updated embedding vector of the second user ID, represents the second user ID embedding vector, loss A Represents the second loss value, δ>0.

[0205] Step 807: Input the third user embedding vector into the user tower of the second benchmark model to obtain a third user vector.

[0206] Step 808: perform an inner product operation on the third user vector and the third material vector to obtain a third similarity calculation result, and based on the third similarity calculation result and the label of the current sample, calculate the loss using the cross entropy loss function to obtain a third loss value.

[0207] Step 809: merge the second loss value and the third loss value to obtain a fourth loss value, and use the fourth loss value to update the parameters of the user ID vector generator of the second benchmark model.

[0208] In some embodiments, the second loss value and the third loss value are merged to obtain a fourth loss value, including:

[0209] The second loss value and the third loss value are combined using formula (3) to obtain the fourth loss value:

[0210] loss meta =α*loss A +(1-α)*loss B (3)

[0211] Among them, loss meta Represents the fourth loss value, loss A Represents the second loss value, loss B represents the third loss value, and α represents the weighting coefficient.

[0212] Step 810: Input other samples in the active user data set one by one into the second benchmark model with updated parameters to continue training until the training end condition is met, end the training, save the parameters of the second benchmark model at the end of the training, and obtain a meta-learning model.

[0213] In an embodiment of the present application, a meta-learning model is obtained by inputting each sample of the constructed active user data set into the embedding layer, user ID vector generator, user tower and material tower of the second benchmark model for training in turn, and then a second cold start recall model is obtained based on the meta-learning model, thereby quickly capturing the interests of cold start users and improving the accuracy and real-time performance of personalized recommendations for cold start users.

[0214] In some embodiments, the step of constructing a cold start user data set includes:

[0215] The user data set is screened according to the activity level of the user to obtain the cold start user data set, wherein each sample in the cold start user data set is constructed according to user attribute information, first type material attribute information and second type material attribute information.

[0216] It should be noted that the user's activity level can be obtained based on the user's use of application software on the electronic device.

[0217] For example, based on the user's usage within one month after registering the application software, if the user does not generate behavioral data of clicking on materials, the user is a cold start user.

[0218] In this step, the screened cold start user data set is aggregated according to the user attribute information to obtain the first category material attribute information and the second category material attribute information, respectively, and each sample in the active user data set is constructed in the form of user attribute information, first category material attribute information and second category material attribute information, wherein the first category material attribute information is different from the second category material attribute information.

[0219] In an embodiment of the present application, the process of training the meta-learning model using the cold-start user dataset is similar to the process of training the second benchmark model using the active user dataset in the aforementioned embodiment, and will not be repeated here.

[0220] In an embodiment of the present application, by inputting each sample of the constructed cold start user data set into the embedding layer, user ID vector generator, user tower and material tower of the meta-learning model for training in turn, a second cold start recall model is obtained, and then the first cold start recall model is obtained based on the second cold start recall model, thereby quickly capturing the interests of cold start users and improving the accuracy and real-time performance of personalized recommendations for cold start users.

[0221] It should be noted that the material recommendation method provided in the embodiment of the present application can be executed by a material recommendation device, or a control module in the material recommendation device for executing the material recommendation method. In the embodiment of the present application, the material recommendation device provided in the embodiment of the present application is described by taking the material recommendation method executed by the material recommendation device as an example.

[0222] The embodiment of the present application also provides a material recommendation device.

[0223] like Fig. 9 As shown, the material recommendation device 900 includes: an acquisition unit 910, a first output unit 920 and a recommendation unit 930.

[0224] An acquisition unit 910 is used to acquire first attribute information of a target cold start user;

[0225] A first output unit 920 is used to input the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on a second cold start recall model, wherein the second cold start recall model is a model obtained by a server based on a user data set and trained using a meta-learning method, and the first vector is used to characterize the preference of the target cold start user;

[0226] The recommendation unit 930 is configured to obtain a first material matching the first vector and display the first material.

[0227] In an embodiment of the present application, the first attribute information of the target cold start user is input into the first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on the second cold start recall model, and the second cold start recall model is a model trained based on the user data set and using a meta-learning method. Therefore, the first vector can characterize the preference of the target cold start user, and then by obtaining the first material matching the first vector, it is possible to quickly capture the interest of the cold start user, thereby improving the accuracy and real-time performance of personalized recommendations for the cold start user, and after obtaining the first vector, obtaining the first material matching the first vector from the server can save memory space occupied by the electronic device and improve resource utilization.

[0228] In some embodiments, the first cold start recall model is a model obtained based on a second cold start recall model deployed by the server on the electronic device, including one of the following:

[0229] The first cold start recall model is the same as the second cold start recall model;

[0230] The first cold start recall model is a model obtained by the electronic device training the second cold start recall model based on the real-time behavior data of the target cold start user;

[0231] The first cold start recall model is a model obtained by the electronic device updating the second cold start recall model based on the update parameters sent by the server.

[0232] In an embodiment of the present application, three different methods are provided for obtaining a first cold start recall model based on a second cold start recall model deployed by a server on an electronic device. The first cold start recall model has a strong ability to capture the interests of cold start users, thereby improving the accuracy of personalized recommendations for cold start users.

[0233] In some embodiments, the structure of the first cold start recall model includes: an embedding layer, a user ID vector generator, a user tower, and a material tower;

[0234] The first output unit 920 is used to:

[0235] Inputting the first attribute information into the embedding layer to obtain a first embedding vector;

[0236] Inputting the first embedding vector into the user ID vector generator to obtain a second embedding vector;

[0237] Concatenate the first embedding vector and the second embedding vector to obtain a third embedding vector;

[0238] The third embedding vector is input into the user tower to obtain the first vector.

[0239] In an embodiment of the present application, by inputting the first attribute information of the target cold start user into the first cold start recall model to obtain the first vector, it is possible to quickly capture the interest of the cold start user and improve the accuracy and real-time performance of personalized recommendations for the cold start user.

[0240] In some embodiments, the apparatus further comprises:

[0241] A training set construction unit, used to obtain real-time behavior data corresponding to the target cold start user, and construct a training data set based on the real-time behavior data;

[0242] The first training unit is used to train the second cold start recall model using the training data set to obtain the first cold start recall model.

[0243] In an embodiment of the present application, a training data set is constructed based on the real-time behavior data of the target cold start user, and the second cold start recall model is trained using the training data set to obtain the first cold start recall model. The first vector is obtained using the first attribute information and the first cold start recall model, which can further improve the accuracy and real-time performance of personalized recommendations for cold start users.

[0244] In some embodiments, the apparatus further comprises:

[0245] A first receiving unit, configured to receive update parameters sent by a server;

[0246] a model updating unit, configured to update the second cold start recall model based on the update parameter to obtain the first cold start recall model;

[0247] The update parameters are obtained by the server acquiring the parameters of the first cold start recall model obtained by training multiple electronic devices through federated learning.

[0248] In an embodiment of the present application, the second cold start recall model is updated by the update parameters sent by the server to obtain the first cold start recall model, and then the first cold start recall model is used to process the target cold start user to obtain the first vector, thereby further improving the accuracy and real-time performance of the cold start user recommendation.

[0249] The material recommendation device in the embodiment of the present application can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, a wearable device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc., which is not specifically limited in the embodiment of the present application.

[0250] The material recommendation device in the embodiment of the present application may be a device having an operating system. The operating system may be an Android operating system, an IOS operating system, or other possible operating systems, which are not specifically limited in the embodiment of the present application.

[0251] The material recommendation device provided in the embodiment of the present application can achieve Figure 3 to Figure 4 To avoid repetition, the various processes implemented by the method embodiment are not described here.

[0252] It should be noted that the material recommendation method provided in the embodiment of the present application can be executed by a material recommendation device, or a control module in the material recommendation device for executing the material recommendation method. In the embodiment of the present application, the material recommendation device provided in the embodiment of the present application is described by taking the material recommendation method executed by the material recommendation device as an example.

[0253] The embodiment of the present application also provides a material recommendation device.

[0254] like Fig.10 As shown, the material recommendation device 1000 includes: a second receiving unit 1010, a second output unit 1020 and a first sending unit 1030.

[0255] The second receiving unit 1010 is configured to receive a first request message sent by an electronic device, where the first request message carries a first vector corresponding to a target cold start user;

[0256] A second output unit 1020 is configured to obtain a first material based on the first vector and a second local cold start recall model of the server;

[0257] The first sending unit 1030 is configured to send a first response message to the electronic device, wherein the first response message carries the first material.

[0258] In an embodiment of the present application, a first vector is obtained through a first request message, a first material is obtained based on the first vector and a local second cold start recall model, and the first material is returned to the electronic device, so as to quickly capture the interests of cold start users and improve the accuracy and real-time performance of personalized recommendations for cold start users.

[0259] In some embodiments, the structure of the second cold start recall model includes: an embedding layer, a user ID vector generator, a user tower, and a material tower;

[0260] The second output unit 1020 is used to:

[0261] Obtaining N material vectors output by the material tower of the second cold start recall model;

[0262] Perform inner product calculations on the first vector and the N material vectors respectively to obtain N calculation results, wherein the calculation results are used to characterize the similarity between the target cold start user and the material corresponding to the first vector;

[0263] Sorting the N calculation results in descending order, obtaining materials corresponding to the top M calculation results, and obtaining the first material;

[0264] Wherein, M and N are both positive integers greater than 1.

[0265] In an embodiment of the present application, multiple material vectors are obtained through the second cold start recall model, the similarity between the target cold start user and the material is obtained based on the first vector and the multiple material vectors, and then the first material is obtained based on the similarity, thereby quickly capturing the interest of the cold start user and improving the accuracy and real-time performance of personalized recommendations for cold start users.

[0266] In some embodiments, the apparatus further comprises:

[0267] The second training unit is used to train the second cold start recall model and deploy the second cold start recall model to the electronic device.

[0268] In some embodiments, the second training unit is used to:

[0269] Constructing a user data set, and using the user data set to train the improved dual-tower model to obtain a first benchmark model;

[0270] Improving the first benchmark model to obtain a second benchmark model;

[0271] constructing an active user data set, and using the active user data set to train the second benchmark model to obtain a meta-learning model;

[0272] Constructing a cold start user data set, and using the cold start user data set to train the meta-learning model to obtain the second cold start recall model;

[0273] Wherein, the structure of the improved double-tower model or the first benchmark model includes: an embedding layer, a user tower and a material tower;

[0274] The structure of the second benchmark model or the meta-learning model or the second cold start recall model includes: an embedding layer, a user ID vector generator, a user tower and a material tower.

[0275] In an embodiment of the present application, a first baseline model is obtained by training with a user data set, the first baseline model is improved to obtain a second baseline model, the second baseline model is trained with an active user data set to obtain a meta-learning model, and finally, the meta-learning model is trained with a cold-start user data set to obtain a second cold-start recall model. This realizes a model that captures the interests of cold-start users and performs personalized recommendations, thereby effectively improving the accuracy and real-time performance of personalized recommendations for cold-start users.

[0276] In some embodiments, the step of constructing a user data set and using the user data set to train an improved dual-tower model to obtain a first benchmark model includes:

[0277] Taking the user's click behavior data as positive samples, randomly sampling the materials that each user has not clicked on to obtain negative samples, and obtaining the user data set based on the positive samples and negative samples, wherein each sample in the user data set includes a user ID, user attribute information, and material attribute information;

[0278] Taking any sample in the user data set as the current sample, inputting the current sample into the embedding layer of the improved twin-tower model, and obtaining a first user ID embedding vector, a first user attribute embedding vector, and a first material attribute embedding vector corresponding to the current sample;

[0279] Concatenate the first user ID embedding vector with the first user attribute embedding vector to obtain a first user embedding vector;

[0280] Inputting the first user embedding vector into the user tower of the improved twin-tower model to obtain a first user vector, and inputting the material attribute embedding vector into the material tower of the improved twin-tower model to obtain a first material vector;

[0281] Performing an inner product operation on the first user vector and the first material vector to obtain a first similarity calculation result, and calculating a loss using a cross entropy loss function based on the first similarity calculation result and the label of the current sample to obtain a first loss value;

[0282] updating the parameters of the improved twin-tower model using the first loss value;

[0283] The other samples in the user data set are input one by one into the improved dual-tower model after parameter update to continue training until the training end condition is met, the training is terminated, the parameters of the improved dual-tower model at the end of the training are saved, and the first benchmark model is obtained.

[0284] In an embodiment of the present application, a first benchmark model is obtained by sequentially inputting the constructed user data set into the embedding layer, user tower and material tower of the improved dual-tower model for training, and a second cold-start recall model is obtained based on the first benchmark model, thereby achieving rapid capture of the interests of cold-start users and improving the accuracy and real-time performance of personalized recommendations for cold-start users.

[0285] In some embodiments, the step of constructing an active user dataset and using the active user dataset to train the second benchmark model to obtain a meta-learning model includes:

[0286] According to the activity level of the user, the user data set is screened to obtain an active user data set, wherein each sample in the active user data set is constructed according to user attribute information, first-category material attribute information, and second-category material attribute information;

[0287] Taking any sample in the active user data set as the current sample, inputting the user attribute information, the first type of material attribute information, and the second type of material attribute information of the current sample into the embedding layer of the second benchmark model, and obtaining the second user attribute embedding vector, the first type of material attribute embedding vector, and the second type of material attribute embedding vector corresponding to the current sample;

[0288] Inputting the second user attribute embedding vector into the user ID vector generator of the second benchmark model to obtain a second user ID embedding vector corresponding to the current sample;

[0289] Concatenate the second user ID embedding vector with the second user attribute embedding vector to obtain a second user embedding vector;

[0290] Input the second user embedding vector into the user tower of the second benchmark model to obtain a second user vector, input the first type of material attribute embedding vector into the material tower of the second benchmark model to obtain a second material vector, and input the second type of material attribute embedding vector into the material tower of the second benchmark model to obtain a third material vector;

[0291] Performing an inner product operation on the second user vector and the second material vector to obtain a second similarity calculation result, and calculating the loss using a cross entropy loss function based on the second similarity calculation result and the label of the current sample to obtain a second loss value;

[0292] Using the second loss value to update the second user ID embedding vector, concatenating the updated second user ID embedding vector with the second user attribute embedding vector to obtain a third user embedding vector;

[0293] Inputting the third user embedding vector into the user tower of the second benchmark model to obtain a third user vector;

[0294] Performing an inner product operation on the third user vector and the third material vector to obtain a third similarity calculation result, and calculating a loss using the cross entropy loss function based on the third similarity calculation result and the label of the current sample to obtain a third loss value;

[0295] The second loss value and the third loss value are merged to obtain a fourth loss value, and the fourth loss value is used to update the parameters of the user ID vector generator of the second benchmark model;

[0296] Input other samples in the active user data set one by one into the second benchmark model with updated parameters to continue training until the training end condition is met, end the training, save the parameters of the second benchmark model at the end of the training, and obtain a meta-learning model.

[0297] In an embodiment of the present application, a meta-learning model is obtained by inputting each sample of the constructed active user data set into the embedding layer, user ID vector generator, user tower and material tower of the second benchmark model for training in turn, and then a second cold start recall model is obtained based on the meta-learning model, thereby quickly capturing the interests of cold start users and improving the accuracy and real-time performance of personalized recommendations for cold start users.

[0298] In some embodiments, the step of constructing a cold start user data set includes:

[0299] The user data set is screened according to the activity level of the user to obtain the cold start user data set, wherein each sample in the cold start user data set is constructed according to user attribute information, first type material attribute information and second type material attribute information.

[0300] In an embodiment of the present application, by inputting each sample of the constructed cold start user data set into the embedding layer, user ID vector generator, user tower and material tower of the meta-learning model for training in turn, a second cold start recall model is obtained, and then the first cold start recall model is obtained based on the second cold start recall model, thereby quickly capturing the interests of cold start users and improving the accuracy and real-time performance of personalized recommendations for cold start users.

[0301] The material recommendation device in the embodiment of the present application can be a device, or a component, integrated circuit, or chip in the device. The device can be a non-mobile electronic device. Exemplarily, the non-mobile electronic device can be a server, a network attached storage (Network Attached Storage, NAS), a personal computer (personal computer, PC), a television (television, TV), a teller machine or a self-service machine, etc., which is not specifically limited in the embodiment of the present application.

[0302] The material recommendation device provided in the embodiment of the present application can achieve Figures 4 to 8 To avoid repetition, the various processes implemented by the method embodiment are not described here.

[0303] Alternatively, if Fig.11 As shown, an embodiment of the present application also provides an electronic device 1100, including a processor 1101, a memory 1102, and a program or instruction stored in the memory 1102 and executable on the processor 1101. When the program or instruction is executed by the processor 1101, each process of the above-mentioned material recommendation method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0304] It should be noted that the electronic devices in the embodiments of the present application include the mobile electronic devices and non-mobile electronic devices mentioned above.

[0305] Fig.12 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of the present application.

[0306] The electronic device 1200 includes but is not limited to: a radio frequency unit 1201, a network module 1202, an audio output unit 1203, an input unit 1204, a sensor 1205, a display unit 1206, a user input unit 1207, an interface unit 1208, a memory 1209, a processor 1210 and other components.

[0307] Those skilled in the art will appreciate that the electronic device 1200 may also include a power source (such as a battery) for supplying power to each component, and the power source may be logically connected to the processor 1210 through a power management system, thereby implementing functions such as managing charging, discharging, and power consumption management through the power management system. Fig.12 The electronic device structure shown in the figure does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently, which will not be described in detail here.

[0308] The processor 1210 is configured to:

[0309] Obtaining the first attribute information of the target cold start user;

[0310] Inputting the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on a second cold start recall model, wherein the second cold start recall model is a model obtained by a server based on a user data set and trained using a meta-learning method, and the first vector is used to characterize the preference of the target cold start user;

[0311] A first material matching the first vector is obtained, and the first material is displayed.

[0312] In an embodiment of the present application, the first attribute information of the target cold start user is input into the first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on the second cold start recall model, and the second cold start recall model is a model trained based on the user data set and using a meta-learning method. Therefore, the first vector can characterize the preference of the target cold start user, and then by obtaining the first material matching the first vector, it is possible to quickly capture the interest of the cold start user, thereby improving the accuracy and real-time performance of personalized recommendations for the cold start user, and after obtaining the first vector, obtaining the first material matching the first vector from the server can save memory space occupied by the electronic device and improve resource utilization.

[0313] Optionally, the first cold start recall model is a model obtained based on a second cold start recall model deployed by the server on the electronic device, and includes one of the following:

[0314] The first cold start recall model is the same as the second cold start recall model;

[0315] The first cold start recall model is a model obtained by the electronic device training the second cold start recall model based on the real-time behavior data of the target cold start user;

[0316] The first cold start recall model is a model obtained by the electronic device updating the second cold start recall model based on the update parameters sent by the server.

[0317] In an embodiment of the present application, three different methods are provided for obtaining a first cold start recall model based on a second cold start recall model deployed by a server on an electronic device. The first cold start recall model has a strong ability to capture the interests of cold start users, thereby improving the accuracy of personalized recommendations for cold start users.

[0318] Optionally, the structure of the first cold start recall model includes: an embedding layer, a user ID vector generator, a user tower and a material tower;

[0319] The step of inputting the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user includes:

[0320] Inputting the first attribute information into the embedding layer to obtain a first embedding vector;

[0321] Inputting the first embedding vector into the user ID vector generator to obtain a second embedding vector;

[0322] Concatenate the first embedding vector and the second embedding vector to obtain a third embedding vector;

[0323] The third embedding vector is input into the user tower to obtain the first vector.

[0324] In an embodiment of the present application, by inputting the first attribute information of the target cold start user into the first cold start recall model to obtain the first vector, it is possible to quickly capture the interest of the cold start user and improve the accuracy and real-time performance of personalized recommendations for the cold start user.

[0325] Optionally, the processor 1210 is further configured to:

[0326] Acquire real-time behavior data corresponding to the target cold start user, and construct a training data set based on the real-time behavior data;

[0327] The second cold start recall model is trained using the training data set to obtain the first cold start recall model.

[0328] In an embodiment of the present application, a training data set is constructed based on the real-time behavior data of the target cold start user, and the second cold start recall model is trained using the training data set to obtain the first cold start recall model. The first vector is obtained using the first attribute information and the first cold start recall model, which can further improve the accuracy and real-time performance of personalized recommendations for cold start users.

[0329] Optionally, the processor 1210 is further configured to:

[0330] Receive update parameters sent by the server;

[0331] updating the second cold start recall model based on the update parameter to obtain the first cold start recall model;

[0332] The update parameters are obtained by the server through federated learning to obtain the parameters of the first cold start recall model obtained by training multiple electronic devices.

[0333] In an embodiment of the present application, the second cold start recall model is updated by the update parameters sent by the server to obtain the first cold start recall model, and then the first cold start recall model is used to process the target cold start user to obtain the first vector, thereby further improving the accuracy and real-time performance of the cold start user recommendation.

[0334] It should be understood that in the embodiment of the present application, the input unit 1204 may include a graphics processor (Graphics Processing Unit, GPU) 12041 and a microphone 12042, and the graphics processor 12041 processes the image data of the static picture or video obtained by the image capture device (such as a camera) in the video capture mode or the image capture mode. The display unit 1206 may include a display panel 12061, and the display panel 12061 may be configured in the form of a liquid crystal display, an organic light emitting diode, etc. The user input unit 1207 includes a touch panel 12071 and other input devices 12072. The touch panel 12071 is also called a touch screen. The touch panel 12071 may include two parts: a touch detection device and a touch controller. Other input devices 12072 may include, but are not limited to, a physical keyboard, a function key (such as a volume control button, a switch button, etc.), a trackball, a mouse, and a joystick, which will not be repeated here. The memory 1209 can be used to store software programs and various data, including but not limited to applications and operating systems. The processor 1210 may integrate an application processor and a modem processor, wherein the application processor mainly processes an operating system, a user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the modem processor may not be integrated into the processor 1210.

[0335] An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the above-mentioned material recommendation method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0336] The processor is a processor in the electronic device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0337] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned material recommendation method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0338] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0339] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0340] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0341] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A material recommendation method, characterized in that: include: The electronic device obtains first attribute information of a target cold start user; The electronic device inputs the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on a second cold start recall model, the second cold start recall model is a model obtained by a server based on a user data set and trained using a meta-learning method, and the first vector is used to characterize the preference of the target cold start user; The electronic device acquires a first material matching the first vector and displays the first material; The structure of the first cold start recall model includes: an embedding layer, a user ID vector generator, a user tower and a material tower; The electronic device inputs the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user, including: Inputting the first attribute information into the embedding layer to obtain a first embedding vector; Inputting the first embedding vector into the user ID vector generator to obtain a second embedding vector; Concatenate the first embedding vector and the second embedding vector to obtain a third embedding vector; The third embedding vector is input into the user tower to obtain the first vector.

2. The material recommendation method according to claim 1, characterized in that: The first cold start recall model is a model obtained based on the second cold start recall model, including one of the following: The first cold start recall model is the same as the second cold start recall model; The first cold start recall model is a model obtained by the electronic device training the second cold start recall model based on the real-time behavior data of the target cold start user; The first cold start recall model is a model obtained by the electronic device updating the second cold start recall model based on the update parameters sent by the server.

3. The material recommendation method according to claim 2, characterized in that: The method further comprises: The electronic device obtains real-time behavior data corresponding to the target cold start user, and constructs a training data set based on the real-time behavior data; The electronic device trains the second cold start recall model using the training data set to obtain the first cold start recall model.

4. The material recommendation method according to claim 2, characterized in that: The method further comprises: The electronic device receives the update parameters sent by the server; The electronic device updates the second cold start recall model based on the update parameter to obtain the first cold start recall model; The update parameters are obtained by the server acquiring the parameters of the first cold start recall model obtained by training multiple electronic devices through federated learning.

5. A material recommendation method, characterized in that: include: The server receives a first request message sent by the electronic device, where the first request message carries a first vector corresponding to a target cold start user, where the first vector is used to represent a preference of the target cold start user; The server obtains a first material based on the first vector and a second cold start recall model local to the server; The server sends a first response message to the electronic device, wherein the first response message carries the first material; The second cold start recall model is a model trained by the server based on the user data set using a meta-learning method; Wherein, the method further comprises: Training to obtain the second cold start recall model, and deploying the second cold start recall model to the electronic device; The training to obtain the second cold start recall model includes: Constructing a user data set, and using the user data set to train the improved dual-tower model to obtain a first benchmark model; Improving the first benchmark model to obtain a second benchmark model; constructing an active user data set, and using the active user data set to train the second benchmark model to obtain a meta-learning model; Constructing a cold start user data set, and using the cold start user data set to train the meta-learning model to obtain the second cold start recall model; Wherein, the structure of the improved double-tower model or the first benchmark model includes: an embedding layer, a user tower and a material tower; The structure of the second benchmark model or meta-learning model includes: an embedding layer, a user ID vector generator, a user tower and a material tower.

6. The material recommendation method according to claim 5, characterized in that: The structure of the second cold start recall model includes: an embedding layer, a user ID vector generator, a user tower and a material tower; The obtaining of the first material based on the first vector and the second cold start recall model of the server includes: Obtaining N material vectors output by the material tower of the second cold start recall model; Perform inner product calculations on the first vector and the N material vectors respectively to obtain N calculation results, wherein the calculation results are used to characterize the similarity between the target cold start user and the material corresponding to the first vector; Sorting the N calculation results in descending order, obtaining materials corresponding to the top M calculation results, and obtaining the first material; Wherein, M and N are both positive integers greater than 1.

7. The material recommendation method according to claim 5, characterized in that: The step of constructing a user data set and using the user data set to train an improved dual-tower model to obtain a first benchmark model includes: Taking the user's click behavior data as positive samples, randomly sampling the materials that each user has not clicked on to obtain negative samples, and obtaining the user data set based on the positive samples and negative samples, wherein each sample in the user data set includes a user ID, user attribute information, and material attribute information; Taking any sample in the user data set as the current sample, inputting the current sample into the embedding layer of the improved twin-tower model, and obtaining a first user ID embedding vector, a first user attribute embedding vector, and a first material attribute embedding vector corresponding to the current sample; Concatenate the first user ID embedding vector with the first user attribute embedding vector to obtain a first user embedding vector; Inputting the first user embedding vector into the user tower of the improved twin-tower model to obtain a first user vector, and inputting the material attribute embedding vector into the material tower of the improved twin-tower model to obtain a first material vector; Performing an inner product operation on the first user vector and the first material vector to obtain a first similarity calculation result, and calculating a loss using a cross entropy loss function based on the first similarity calculation result and the label of the current sample to obtain a first loss value; updating the parameters of the improved twin-tower model using the first loss value; The other samples in the user data set are input one by one into the improved dual-tower model after parameter update to continue training until the training end condition is met, the training is terminated, the parameters of the improved dual-tower model at the end of the training are saved, and the first benchmark model is obtained.

8. The material recommendation method according to claim 6, characterized in that: The step of constructing an active user data set and using the active user data set to train the second benchmark model to obtain a meta-learning model includes: According to the activity level of the user, the user data set is screened to obtain an active user data set, wherein each sample in the active user data set is constructed according to user attribute information, first-category material attribute information, and second-category material attribute information; Taking any sample in the active user data set as the current sample, inputting the user attribute information, the first type of material attribute information, and the second type of material attribute information of the current sample into the embedding layer of the second benchmark model, and obtaining the second user attribute embedding vector, the first type of material attribute embedding vector, and the second type of material attribute embedding vector corresponding to the current sample; Inputting the second user attribute embedding vector into the user ID vector generator of the second benchmark model to obtain a second user ID embedding vector corresponding to the current sample; Concatenate the second user ID embedding vector with the second user attribute embedding vector to obtain a second user embedding vector; Input the second user embedding vector into the user tower of the second benchmark model to obtain a second user vector, input the first type of material attribute embedding vector into the material tower of the second benchmark model to obtain a second material vector, and input the second type of material attribute embedding vector into the material tower of the second benchmark model to obtain a third material vector; Performing an inner product operation on the second user vector and the second material vector to obtain a second similarity calculation result, and calculating the loss using a cross entropy loss function based on the second similarity calculation result and the label of the current sample to obtain a second loss value; Using the second loss value to update the second user ID embedding vector, concatenating the updated second user ID embedding vector with the second user attribute embedding vector to obtain a third user embedding vector; Inputting the third user embedding vector into the user tower of the second benchmark model to obtain a third user vector; Performing an inner product operation on the third user vector and the third material vector to obtain a third similarity calculation result, and calculating a loss using the cross entropy loss function based on the third similarity calculation result and the label of the current sample to obtain a third loss value; The second loss value and the third loss value are merged to obtain a fourth loss value, and the fourth loss value is used to update the parameters of the user ID vector generator of the second benchmark model; Input other samples in the active user data set one by one into the second benchmark model with updated parameters to continue training until the training end condition is met, end the training, save the parameters of the second benchmark model at the end of the training, and obtain a meta-learning model.

9. A material recommendation device, characterized in that: include: An acquisition unit, used for acquiring first attribute information of a target cold start user; a first output unit, configured to input the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user, wherein the first cold start recall model is a model obtained based on a second cold start recall model, wherein the second cold start recall model is a model obtained by a server based on a user data set and trained using a meta-learning method, and the first vector is used to characterize a preference of the target cold start user; A recommendation unit, configured to obtain a first material matching the first vector and display the first material; The structure of the first cold start recall model includes: an embedding layer, a user ID vector generator, a user tower and a material tower; The electronic device inputs the first attribute information into a first cold start recall model to obtain a first vector corresponding to the target cold start user, including: Inputting the first attribute information into the embedding layer to obtain a first embedding vector; Inputting the first embedding vector into the user ID vector generator to obtain a second embedding vector; Concatenate the first embedding vector and the second embedding vector to obtain a third embedding vector; The third embedding vector is input into the user tower to obtain the first vector.

10. A material recommendation device, characterized in that: include: A second receiving unit, configured to receive a first request message sent by an electronic device, where the first request message carries a first vector corresponding to a target cold start user, where the first vector is used to represent a preference of the target cold start user; A second output unit, configured to obtain a first material based on the first vector and a second local cold start recall model of the server; A first sending unit, configured to send a first response message to the electronic device, wherein the first response message carries the first material; The second cold start recall model is a model trained by the server based on the user data set using a meta-learning method; Wherein, the device is also used for: Training to obtain the second cold start recall model, and deploying the second cold start recall model to the electronic device; The training to obtain the second cold start recall model includes: Constructing a user data set, and using the user data set to train the improved dual-tower model to obtain a first benchmark model; Improving the first benchmark model to obtain a second benchmark model; constructing an active user data set, and using the active user data set to train the second benchmark model to obtain a meta-learning model; Constructing a cold start user data set, and using the cold start user data set to train the meta-learning model to obtain the second cold start recall model; Wherein, the structure of the improved double-tower model or the first benchmark model includes: an embedding layer, a user tower and a material tower; The structure of the second benchmark model or meta-learning model includes: an embedding layer, a user ID vector generator, a user tower and a material tower.

11. An electronic device, characterized in that: It includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the material recommendation method as described in any one of claims 1 to 4, or implements the steps of the material recommendation method as described in any one of claims 5 to 8.

Images