Data recommendation method and device, electronic equipment and storage medium

Abstract

The invention relates to a data recommendation method and device, electronic equipment and a storage medium, and the method comprises the steps: responding to target retrieval information received in a target retrieval domain, and determining a target convolutional network corresponding to the target retrieval domain from a plurality of specific convolutional networks, the plurality of specific convolutional networks are generated by training a plurality of personalized convolutional networks of a plurality of retrieval domains according to common sample data, one personalized convolutional network is generated by performing element-level interaction on a shared convolutional network corresponding to the plurality of retrieval domains and an initial convolutional network, target retrieval information is analyzed based on a target convolutional network, and the target retrieval information is obtained. And generating push data. Therefore, the retrieval model is trained by learning the generality and the difference between different retrieval domains, and meanwhile, the model prediction performance of the multiple retrieval domains is improved, so that the determined push data can be close to the real retrieval appeal of the user.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data analysis, and in particular to a data recommendation method, device, electronic device and storage medium. Background Art

[0002] The data recommendation system includes multiple domain scenarios. For example, in the software recommendation application, the user enters the software name or software type to be retrieved in the corresponding search dialog box, and the application software background recommends software to the user based on the input content. Among them, the domain scenarios can include game center and application store scenarios, Douban movie rating and book rating scenarios. There are certain similarities between the various scenarios, and shared information can be used to break the problems of cold start and information cocoon, and achieve better recommendation results. Summary of the invention

[0003] In order to overcome the problems existing in the related art, the present disclosure provides a data recommendation method, device, electronic device and storage medium.

[0004] According to a first aspect of an embodiment of the present disclosure, a data recommendation method is provided, including:

[0005] In response to the target retrieval information received in the target retrieval domain, a target convolutional network corresponding to the target retrieval domain is determined from a plurality of specific convolutional networks, wherein the plurality of specific convolutional networks are generated by training a plurality of individual convolutional networks of a plurality of retrieval domains according to the shared sample data, wherein a individual convolutional network is generated by element-level interaction between the shared convolutional network corresponding to the plurality of retrieval domains and the initial convolutional network;

[0006] The target retrieval information is analyzed based on the target convolutional network to generate push data.

[0007] Optionally, the method comprises:

[0008] generating multi-domain common features according to user behaviors in the multiple search domains;

[0009] Based on a plurality of learnable parameters corresponding one to one to the plurality of search domains, partition normalization processing is performed on the common features of the plurality of domains to generate the common sample data.

[0010] Optionally, the partitioning and normalizing of the multi-domain common features based on the multiple learnable parameters corresponding to the multiple search domains to generate common sample data includes:

[0011] Comparing the multiple initial convolutional networks corresponding to the multiple search domains with the shared convolutional network to generate the multiple learnable parameters, wherein the multiple learnable parameters include multiple scaling parameters and multiple offset parameters;

[0012] Performing standard normalization processing on the multi-domain common features to generate a standard normalization result;

[0013] Batch normalization is performed on the multiple scaling parameters, the multiple offset parameters, and the standard normalization result to generate the common sample data.

[0014] Optionally, analyzing the target retrieval information based on the target convolutional network to generate push data includes:

[0015] Analyzing the target retrieval information based on the target convolutional network to generate a target retrieval intent;

[0016] According to the target search intention, determining a plurality of degrees of fit corresponding to the plurality of pushed data one by one;

[0017] The plurality of push data are sorted in descending order of the plurality of fits to generate a data push list, wherein the data push list includes the sorted plurality of push data.

[0018] Optionally, the method further comprises:

[0019] Generating the shared convolutional network according to the common features of the multiple initial convolutional networks corresponding to the multiple search domains;

[0020] For any retrieval domain corresponding to the multiple retrieval domains, comparing the shared convolutional network with an initial convolutional network corresponding to the retrieval domain to generate an offset parameter of the retrieval domain;

[0021] Performing element-level interaction between the shared convolutional network and the corresponding initial convolutional network to generate a fused convolutional network;

[0022] The personalized convolutional network is generated according to the offset parameters of the fused convolutional network and the retrieval domain.

[0023] Optionally, the batch normalizing the multiple scaling parameters, the multiple offset parameters and the standard normalization result to generate the common sample data includes:

[0024] Obtaining shared scaling parameters and shared offset parameters corresponding to the shared convolutional network;

[0025] For any scaling parameter, offset parameter and the standard normalization result corresponding to any search domain, generating partition sample data corresponding to the search domain in a preset manner;

[0026] Traversing the multiple search domains to generate multiple partition sample data;

[0027] The plurality of partitioned sample data are concatenated to generate the common sample data.

[0028] Optionally, analyzing the target retrieval information based on the target convolutional network to generate a data push list includes:

[0029] Acquire identity feature information of the target search domain;

[0030] Based on the identity feature information and the common features among the multiple search domains, a shared auxiliary network is trained to generate a target auxiliary network;

[0031] The target retrieval information is analyzed by the target auxiliary network and the target convolutional network to generate the data push list.

[0032] According to a second aspect of an embodiment of the present disclosure, a data recommendation device is provided, the device comprising:

[0033] A determination module is configured to determine, in response to target retrieval information received in a target retrieval domain, a target convolutional network corresponding to the target retrieval domain from a plurality of specific convolutional networks, wherein the plurality of specific convolutional networks are generated by training a plurality of individual convolutional networks of a plurality of retrieval domains according to shared sample data, wherein a specific convolutional network is generated by element-level interaction between a shared convolutional network and an initial convolutional network corresponding to the plurality of retrieval domains;

[0034] An execution module is configured to analyze the target retrieval information based on the target convolutional network to generate push data.

[0035] Optionally, the device further includes a generating module, the generating module including:

[0036] A first generating submodule is configured to generate multi-domain common features according to user behaviors in the multiple search domains;

[0037] The second generating submodule is configured to perform partition normalization processing on the common features of the multiple domains based on a plurality of learnable parameters corresponding one to one to the multiple retrieval domains, so as to generate the common sample data.

[0038] Optionally, the second generation submodule includes:

[0039] A first generating unit is configured to compare a plurality of initial convolutional networks corresponding to a plurality of search domains with a shared convolutional network to generate a plurality of learnable parameters, wherein the plurality of learnable parameters include a plurality of scaling parameters and a plurality of offset parameters;

[0040] The second generating unit is configured to perform standard normalization processing on the common features of multiple domains to generate a standard normalization result;

[0041] The third generating unit is configured to perform batch normalization on multiple scaling parameters, multiple offset parameters and standard normalization results to generate common sample data.

[0042] Optionally, the third generating unit is configured to:

[0043] Get the shared scaling parameters and shared offset parameters corresponding to the shared convolutional network;

[0044] For any scaling parameter, offset parameter and standard normalization result corresponding to a search domain, generate partition sample data corresponding to the search domain in a preset manner;

[0045] Traverse multiple search domains to generate multiple partition sample data;

[0046] Multiple partition sample data are concatenated to generate common sample data.

[0047] Optionally, the device further includes a fourth generating module, and the fourth generating module is configured to:

[0048] Generate a shared convolutional network based on the common features of multiple initial convolutional networks corresponding to multiple retrieval domains;

[0049] For any retrieval domain corresponding to the multiple retrieval domains, comparing the shared convolutional network with a first initial convolutional network corresponding to the retrieval domain to generate an offset parameter of the retrieval domain;

[0050] Perform element-level interaction between the shared convolutional network and the corresponding initial convolutional network to generate a fused convolutional network.

[0051] According to the offset parameters of the fused convolutional network and the retrieval domain, a personalized convolutional network is generated.

[0052] Optionally, the execution module is configured to:

[0053] Analyze the target retrieval information based on the target convolutional network and generate the target retrieval intent;

[0054] According to the target search intent, multiple degrees of fit corresponding to multiple push data are determined;

[0055] The plurality of push data are sorted in descending order according to the plurality of fits to generate a data push list, wherein the data push list includes the sorted plurality of push data.

[0056] Optionally, the execution module is configured to:

[0057] Obtain identity feature information of the target search domain;

[0058] Based on identity feature information and common features between multiple retrieval domains, a shared auxiliary network is trained to generate a target auxiliary network.

[0059] The target retrieval information is analyzed through the target auxiliary network and the target convolutional network to generate a data push list.

[0060] According to a third aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0061] a memory having a computer program stored thereon;

[0062] A processor is used to execute the computer program in the memory to implement the steps of any one of the methods described in the first aspect of the present disclosure.

[0063] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored. When the program instructions are executed by a processor, the steps of any one of the methods described in the first aspect of the present disclosure are implemented.

[0064] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects:

[0065] In the above manner, in response to the target search information received in the target search domain, the target convolution network corresponding to the target search domain is determined from multiple specific convolution networks, wherein the multiple specific convolution networks are generated by training multiple individual convolution networks of multiple search domains based on the common sample data, wherein a specific convolution network is generated by element-level interaction between the shared convolution network and the initial convolution network corresponding to the multiple search domains, and the target search information is analyzed based on the target convolution network to generate push data. Thus, the retrieval model is trained by learning the commonalities and differences between different search domains, and the model prediction performance of multiple search domains is improved at the same time, so that the determined push data can be close to the user's real search demands.

[0066] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0068] Figure 1 It is a schematic diagram showing a cross-domain recommendation method according to an exemplary embodiment.

[0069] Figure 2 is a schematic diagram of a multi-task learning structure according to an exemplary embodiment.

[0070] Figure 3 The figure is a flow chart of a data recommendation method according to an exemplary embodiment.

[0071] Figure 4 It is a flowchart of a method for generating a personalized convolutional network according to an exemplary embodiment.

[0072] Figure 5 It is a schematic diagram of a method for generating a target convolutional network according to an exemplary embodiment.

[0073] Figure 6 The figure is a flowchart of a method for generating common sample data according to an exemplary embodiment.

[0074] Figure 7 The figure is a flowchart of a method for generating common sample data according to an exemplary embodiment.

[0075] Figure 8 It is a flowchart of a model training method according to an exemplary embodiment.

[0076] Fig. 9 The figure is a block diagram of a data pushing device according to an exemplary embodiment.

[0077] Fig.10 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0078] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0079] It should be noted that all actions of acquiring signals, information or data in the present disclosure are carried out in compliance with the relevant data protection laws and policies of the country where the device is located and with the authorization given by the owner of the corresponding device.

[0080] In existing recommendation systems, there is a long-standing problem: how to use multi-domain information to improve search recommendation performance, because commercial platforms often include multiple application scenarios such as search advertising recommendations or search content recommendations. In related technologies, one method is to use auxiliary domain information to help the target domain solve the problem of insufficient information, so that the recommendation model can score correctly. This method mainly uses various auxiliary information to improve the recommendation performance of the target domain, such as:

[0081] 1. Use user attribute information (identity, gender, age, etc.) to solve the user's cold start problem;

[0082] 2. Use the attribute information of the retrieved project to help solve the project cold start problem;

[0083] 3. Use auxiliary domain information (friends of the target user, corresponding to the relevant information of the friend domain) to help solve the cold start problem in the target domain, which is also known as CDR (Cross-Domain Recommendation) technology.

[0084] Generally speaking, among the above three methods, the third method directly utilizes the information of a domain, which is richer than the auxiliary information utilized by the first three methods, and therefore can achieve better recommendation effects.

[0085] Among them, EMCDR (Cross-Domain Recommendation: An Embedding and Mapping Approach) uses cross-domain recommendation technology to solve the cold start problem of the target domain. Its recommendation goal is to target users with little information in the target domain or to retrieve "items" (including: retrieving target information, retrieving target data, etc.). For example, for a new user or new item, no embedding vector (low-dimensional vector) that can represent it can be found in the target domain, but it is more active in the source domain and can find an embedding vector that can represent it. At this time, the embedding vector in the source domain is converted through the bridge function (mapping function) between the target domain and the source domain, and the embedding vector of the target domain can be obtained. Then, recommendations are made to users based on the embedding vector of the target domain. Among them, mapping can be considered as a mapping from the embedding vector of the source domain to the embedding vector of the target domain. Therefore, the embedding vector can only represent the intersection of the source domain and the target domain. For example, Figure 1 is a schematic diagram of a cross-domain recommendation method according to an exemplary embodiment. Figure 1 As shown, the method includes the following steps.

[0086] (1) Generate the first embedding vector of the source domain and the second embedding vector of the target domain based on the score matrix. For example, matrix decomposition and other techniques can be used to learn from the source domain model and the target domain model, and then use a bridge function for mapping. The bridge function (mapping function) here can be mapped by linear mapping, that is, multiplying a parameter matrix, or by MLP (Multilayer Perceptron, artificial neural network).

[0087] (2) Map the first embedding vector through the bridge function to obtain the target embedding vector, which is used as the target embedding vector of the cold-start user or cold-start item in the target domain.

[0088] (3) The optimization goal is that the distance between the target embedding vector obtained by mapping and the second embedding vector obtained by separation in the target domain should be as small as possible, that is, the closer the target embedding vector is to the second embedding vector, the better. For example, the loss function in the target domain can be:

[0089]

[0090] Among them, L is the loss function, f(u) is the bridge function, and u i is the target embedding vector, is the user embedding vector of the target domain.

[0091] There are two problems in the above EMCDR technical solution: (1) The model is not end-to-end and must be trained in segments. This means that there is no guarantee that the model can achieve the best prediction performance. The best effect achieved in a certain segment, for example, after a good prediction effect is achieved in the source domain, it cannot be used in the target domain for prediction, so that the entire system cannot achieve the best performance; (2) Such a method can only work for the target domain, and is not helpful for model training in the source domain. It cannot improve the model prediction performance of both domains at the same time.

[0092] MMoE (Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts) is based on multi-objective learning to simultaneously learn information from two fields. For example, Figure 2 is a schematic diagram of a multi-task learning structure according to an exemplary embodiment. Figure 2As shown in the figure, in the MMoE structure, the input is the combined information of the two domains. The expert network 0-2 in the middle learns the common information of the two domains, while the gate network A and gate network B control the information of the two domains respectively. Finally, they are combined and input into tower A and tower B respectively. The two output models of the tower are used as the models of the two domains AB respectively. Therefore, the model can only learn the commonalities between domains, but cannot express the differences between domains. Moreover, each domain needs to train a model separately, which greatly increases the computing resources and human maintenance costs. The MMoE technical solution also has three problems: (1) It requires strong correlation between multiple domains, such as cross-users or items, or similar item labels. If the correlation between multiple domains is very low, the effect of the model will be very poor; (2) This type of model can only learn the commonalities between domains, but cannot model the differences between each domain; (3) Each domain requires a separate model for separate training, which has high computing resources and human maintenance costs.

[0093] Based on the problems existing in the above related technologies, this application provides a data recommendation method. Figure 3 is a flow chart of a data recommendation method according to an exemplary embodiment. Figure 3 As shown, the method is used in a terminal and includes the following steps.

[0094] In step S11, in response to target retrieval information received in a target retrieval domain, a target convolutional network corresponding to the target retrieval domain is determined from multiple specific convolutional networks, wherein the multiple specific convolutional networks are generated by training multiple individual convolutional networks of multiple retrieval domains based on common sample data, wherein a specific convolutional network is generated by element-level interaction between a shared convolutional network and an initial convolutional network corresponding to the multiple retrieval domains.

[0095] It should be noted that the present embodiment is applied to a terminal, and the terminal is loaded with an application program, and the application program has a search dialog box. The user can enter the target item category, target item ID, target item name and other information to be searched in the search dialog box, and search for target items similar to the search content in the application program. Among them, the searched items can be application software, target goods, takeaway stores, etc. For example, when the application software is a shopping software, the user enters the name of the target goods in the search dialog box, and the shopping software will recommend goods to the user based on the input name. When the application software is a software store, the user enters the target software name or target software type in the search dialog box, and the software store will recommend software to the user based on the name or type. Generally, in order to improve the accuracy of the application software search results and make the recommended application data close to the user's real search needs, a model for prediction will be loaded in the application software. The model can analyze the user's user behavior in the application software, analyze the user's preferences, interests, preferences and other feature information, generate a user portrait, and recommend items to the user based on the user portrait and the search terms entered by the user. Among them, when the application software recommends items, it can recommend multiple items to the user, and generate an item recommendation list in descending order of the degree of fit between the items and the user, and recommend multiple items to the user.

[0096] For example, based on the opacity of user data between various application software, this embodiment can be applied to the same application software of the terminal, wherein the same application software is divided into multiple search domains, and different search domains are loaded with sub-models for prediction. The search domains in the application software can be divided according to the distribution position of the search dialog box in the application software.

[0097] For example, when the application software is a shopping software, a search dialog box is usually set in the homepage of the shopping software. The user can enter the store name in the search dialog box to search the store list. At this time, the search domain is the target store search domain, and the corresponding item in the target store search domain is the store; when the user clicks to enter the selected target store, in order to facilitate the user to search for the target product in the target store, a search dialog box is set in the product search domain corresponding to the target store. The user enters the target product to be searched in the search dialog box to obtain the corresponding target product information. At this time, in this application scenario, the search domain is the product search domain, and the corresponding item is the product in the target store. When the application software is a software store, referring to the setting method of the above shopping software, a search dialog box is set in the homepage of the software store. The user can enter the application software name in the search dialog box to search for the application software. At this time, the search domain is the software search domain of all software in the application store. The items retrieved by the user through the search term are all types of software, including life software, game software, tool software, etc.; a search dialog box is set in the game center of the software store. When the user searches in the search dialog box, the items obtained are various types of game software.

[0098] For example, if the user data between various application software is open and transparent, this embodiment can also be applied to multiple application software in the same terminal. In this case, one application software corresponds to one search domain, and the items in the search domain are items presented to the user based on the functions of each application software. For example, in the shopping software domain, the corresponding items are commodities; in the application software store domain, the corresponding items are application software; in the food delivery platform domain, the corresponding items are food. The definition of the search domain is not limited in this embodiment, and the scope of different search domains in the same terminal can be set according to actual needs and the transparency of user behavior in the terminal.

[0099] By way of example, in this embodiment, a specific convolutional network is generated by training a personalized convolutional network. When the target retrieval information input by the user in the corresponding search box is received in the target retrieval domain, the target retrieval domain corresponding to the target retrieval information is determined, and the target convolutional network corresponding to the target retrieval domain is determined from multiple specific convolutional networks. By way of example, multiple retrieval domains can be numbered, and the number of the corresponding retrieval domain can be determined according to the input position of the target retrieval information, and the target convolutional network of the target retrieval domain can be determined according to the number.

[0100] It should be noted that in this embodiment, one search domain corresponds to one personalized convolutional network. In response to the search information entered by the user in the search domain, the search information is analyzed by the personalized convolutional network, and target items that meet the user's needs are recommended to the user. Among them, the personalized convolutional network is generated by element-level interaction between the shared convolutional network and the initial convolutional network corresponding to multiple search domains. The personalized convolutional networks corresponding to each search domain are trained through the common sample data in each search domain to generate specific convolutional networks corresponding to each search domain. During the application process, when the terminal receives the target search information, it determines the target search domain corresponding to the target search information, determines the corresponding target convolutional network based on the target search domain, and identifies the target search information based on the target convolutional network, thereby determining the user's search intention in the target search domain.

[0101] In step S12, the target retrieval information is analyzed based on the target convolutional network to generate push data.

[0102] For example, in this embodiment, the target retrieval information input by the user is analyzed based on the target convolutional network to determine the user intention that meets the user's personalized needs, retrieve push data that meets the user's intention from the database corresponding to the target retrieval domain, and display the push data in the current page displayed on the terminal.

[0103] Optionally, in another implementation, the above step S12 includes:

[0104] Analyze the target retrieval information based on the target convolutional network and generate the target retrieval intent;

[0105] According to the target search intent, multiple degrees of fit corresponding to multiple push data are determined;

[0106] The plurality of push data are sorted in descending order according to the plurality of fits to generate a data push list, wherein the data push list includes the sorted plurality of push data.

[0107] It is worth mentioning that, under normal circumstances, the identification of the target retrieval intent is relative. It can be used to measure the fit between the pushed data and the user's intent, but it cannot fully represent the user's true intent. Therefore, when pushing data, multiple data will be pushed to the user, allowing the user to make a choice based on the multiple data, thereby increasing the possibility of the user choosing the pushed data.

[0108] For example, in this embodiment, based on the trained target convolution network, the target search information input by the user in the target search domain is analyzed, so as to generate the user's target search intention based on the user's personalized characteristics, and according to the target search intention, multiple push data that meet the target search intention are retrieved from the database corresponding to the target search domain, and multiple fits of the multiple push data are determined based on the target search intention. The multiple push data are sorted in order of fit from large to small, and a data push list is generated, wherein the data push list includes the sorted multiple push data.

[0109] Optionally, in another implementation, the above step S12 includes:

[0110] Obtain identity feature information of the target search domain;

[0111] Based on identity feature information and common features between multiple retrieval domains, a shared auxiliary network is trained to generate a target auxiliary network.

[0112] The target retrieval information is analyzed through the target auxiliary network and the target convolutional network to generate a data push list.

[0113] It is worth mentioning that the differences between domains are not considered in the process of training the personalized convolutional network. Therefore, in order to enhance the model's ability to distinguish between various retrieval domains, an auxiliary network is introduced. For example, based on identity feature information and common features between multiple retrieval domains, the shared auxiliary network is trained to generate a target auxiliary network. On the basis of the target convolutional network generated in the above steps, the target auxiliary network is superimposed on the target convolutional network. The target retrieval information is analyzed through the target auxiliary network and the target convolutional network to generate a data push list. For example, each domain is represented as an ID feature. The feature vector of the ID feature is learned through the shared auxiliary network, and the function is mapped to the feature vector to generate a target auxiliary network. The target auxiliary network is then superimposed on the target convolutional network, and the target retrieval information is analyzed through the superimposed network to generate a data push list.

[0114] In the above manner, in response to the target search information received in the target search domain, the target convolution network corresponding to the target search domain is determined from multiple specific convolution networks, wherein the multiple specific convolution networks are generated by training multiple individual convolution networks of multiple search domains based on the shared sample data, wherein a specific convolution network is generated by element-level interaction between the shared convolution network and the initial convolution network corresponding to the multiple search domains, and the target search information is analyzed based on the target convolution network to generate push data. Thus, the retrieval model is trained by learning the commonalities and differences between different search domains, and the model prediction performance of multiple search domains is improved at the same time, so that the determined push data can be close to the user's real search demands.

[0115] Figure 4 is a flow chart of a method for generating a personalized convolutional network according to an exemplary embodiment. Figure 4 As shown, applied to a terminal, the method further includes the following steps.

[0116] In step S101, a shared convolutional network is generated according to the common features of multiple initial convolutional networks corresponding to multiple search domains;

[0117] In step S102, for any search domain corresponding to the multiple search domains, the shared convolutional network is compared with the first initial convolutional network corresponding to the search domain to generate an offset parameter of the search domain;

[0118] In step S103, the shared convolutional network and the corresponding initial convolutional network are interacted at the element level to generate a fused convolutional network;

[0119] In step S104, a personalized convolutional network is generated according to the offset parameters of the fused convolutional network and the retrieval domain.

[0120] In the present embodiment, in the initial state, each search domain is provided with an initial convolutional network, which is used to respond to the search content input by the user in the search domain, and to perform model training on the preset convolutional network based on the unique characteristics of each search domain, so as to generate an initial convolutional network unique to each search domain. Among them, based on different search purposes and search requirements, the preset convolutional networks set in each search domain are different. Therefore, the initial convolutional networks between different search domains are different. The preset convolutional network will be based on the sample data collected in the corresponding search domain for learning and training, so as to more accurately identify the user's search intention based on the search terms, so that the initial convolutional network can be close to the actual search requirements of the corresponding search domain. The convolution parameters between multiple initial convolutional networks are compared to determine the common characteristic parameters in each initial convolutional network, and a shared convolutional network is generated based on the common characteristic parameters of multiple initial convolutional networks. Then, the shared convolutional network is interacted with each initial convolutional network at the element level to generate a personalized convolutional network corresponding to each search domain.

[0121] For example, for the personalized convolutional network in any retrieval domain, the sample data of each personalized convolutional network is compared to determine the common sample data of the personalized convolutional network, and then the common sample data is input into the personalized convolutional network, and the personalized convolutional network is trained based on the common sample data to generate a specific convolutional network corresponding to the retrieval domain. It is worth mentioning that the multiple personalized convolutional networks corresponding to the retrieval domain include a first personalized convolutional network that has completed learning and training, and may also include a second personalized convolutional network that has not undergone learning and training. The first personalized convolutional network can improve the first personalized convolutional network to generate a first unique convolutional network by learning the common sample data extracted from multiple retrieval domains, so that the first personalized convolution trains different features in multiple retrieval domains; the second personalized convolutional network is in a cold start state before training because it has not undergone training with multiple sample data. By learning the common sample data, the cold start barrier is broken, and the second personalized convolutional network is improved to generate a second specific convolutional network. For example, Figure 5 is a schematic diagram of a method for generating a target convolutional network according to an exemplary embodiment, such as Figure 5 As shown in the figure, the shared sample data after the input layer will be used as the input of the convolutional network to train the convolutional network. The model parameters of the target convolutional network are obtained by element-wise multiplication of the shared convolutional network and the initial convolutional network corresponding to each retrieval domain, and the bias vector is obtained by adding the offset parameters corresponding to each retrieval domain. For example,

[0122]

[0123]

[0124] in, is the initial specific convolutional network, W i is the initial convolutional network corresponding to the retrieval domain i, W is the shared convolutional network, is the bias vector, b i is the offset parameter corresponding to the retrieval domain i, and b is the shared offset parameter.

[0125] It should be noted that in the process of element-level interaction between the shared convolutional network and the corresponding initial convolutional network, assuming that the dimension of the corresponding latent vector of each convolutional network is 3-dimensional, if two features (the corresponding vectors are (a1, b1, c1) and (a2, b2, c2)) interact in a form similar to f(w1*a1*a2, w2*b1*b2, w3*c1*c2), then it is determined that the interaction between the two features occurs at the element level. The inherent characteristics and personalized features of the two feature vectors are retained through element-level interaction, so that the obtained personalized convolutional network can more accurately identify user intentions.

[0126] In the above manner, the shared convolutional network is generated by using the common features between convolutions, the shared convolutional network is compared with the initial convolutional network, an offset parameter is generated, and the fused convolutional network is adjusted by using the offset parameter to generate a personalized convolutional network. Thus, the personalized convolutional network retains the common features of each convolutional network and has personalized features to meet the corresponding application requirements.

[0127] Figure 6 is a flow chart of a method for generating common sample data according to an exemplary embodiment. Figure 6 As shown, in one embodiment, before the above step S11, the method further includes:

[0128] In step S201, multi-domain common features are generated according to user behaviors in multiple search domains.

[0129] For example, in this embodiment, multiple search domains under the same application software in the same terminal are taken as an example, and the search domain in the application software is used as the data collection scope to collect the user behavior of the user in the page corresponding to the search domain. The user behavior includes the browsing behavior, subscription behavior, search behavior, selection behavior, etc. of the user based on the terminal. The user behavior is summarized and processed with the search domain as the summary unit, and then the summary data of multiple search domains are summarized to generate multi-domain common features. For example, in this embodiment, the multi-domain common features can be divided into three dimensions: users, items, and search terms in the search domain. The user behavior is summarized and processed based on the three dimensions, and then the data information of the three dimensions is summarized to generate multi-domain common features. Among them, when collecting user behaviors in multiple search domains, data collection is performed for the category of user behavior, without paying attention to whether there is commonality in user data under the same category. Therefore, when collecting user features, the common features corresponding to each search domain include both the common features of each search domain and the unique features of each search domain. It is ensured that there is commonality in the multi-domain common features corresponding to multiple search domains, while also retaining the differences between each search domain.

[0130] In step S202, based on a plurality of learnable parameters corresponding one to one to a plurality of search domains, a partition normalization process is performed on the common features of the multiple domains to generate common sample data.

[0131] It is worth mentioning that sub-models are set in different search domains. After analyzing the user behavior in the search domain through the sub-model, the items that meet the user's needs are identified from the item library based on the search terms entered by the user and the user analysis results. In this embodiment, the initial sub-models set in each search domain are the same in the initial state. By adding training samples to each search domain, the initial sub-model is trained to generate a sub-model specific to each search domain. When the trained sub-model is compared with the initial sub-model, the unique personalized parameters in the trained sub-model are the learnable parameters of the corresponding search domain. The trained sub-models corresponding to each search domain are compared with the preset model to determine multiple learnable parameters specific to each search domain. Among them, there is a one-to-one correspondence between the search domain-trained sub-model-learnable parameters.

[0132] In order to speed up the convergence of the model, the training samples of the model need to be normalized. In this embodiment, according to multiple learnable parameters, the common features of multiple domains are partitioned and normalized, and the learnable parameters of the sub-models corresponding to each retrieval domain are used to normalize the common features of multiple domains, generate partitioned normalized data corresponding to each retrieval domain, and then splice multiple partitioned normalized data to generate common sample data. For example, in this embodiment, the BN (Batch Normalization) method can be used to perform normalization using the shared statistics of the common features of multiple domains, and then splice the processed results to generate common sample data. It is worth mentioning that in this embodiment, in the process of partitioning and normalizing the common features of multiple domains based on learnable parameters, the scene difference information of different retrieval domains is retained, thereby facilitating subsequent fine model building.

[0133] Figure 7 is a flow chart of a method for generating common sample data according to an exemplary embodiment. Figure 7 As shown, the above step S202 includes the following steps.

[0134] In step S221, multiple initial convolutional networks corresponding to multiple retrieval domains are compared with the shared convolutional network to generate multiple learnable parameters, wherein the multiple learnable parameters include multiple scaling parameters and multiple offset parameters.

[0135] For example, in this embodiment, the shared convolutional network is generated based on the common features between the initial convolutional networks. The initial convolutional networks corresponding to each retrieval domain are compared with the shared convolutional network to determine the differences between the initial convolutional networks and the shared convolutional network. Based on the differences, multiple learnable parameters of each initial convolutional network are generated, wherein the learnable parameters include multiple scaling parameters and multiple offset parameters.

[0136] In step S222, standard normalization processing is performed on the common features of multiple domains to generate standard normalization results.

[0137] For example, in this embodiment, the common features of multiple domains are subjected to standard normalization. At this time, a unified standard of the common features is normalized without partitioning the standard to generate a standard normalization result.

[0138] In step S223, batch normalization is performed on multiple scaling parameters, multiple offset parameters and standard normalization results to generate common sample data.

[0139] For example, in this embodiment, batch normalization is performed on the standard normalization results based on the scaling parameters and offset parameters corresponding to each retrieval domain, and partition normalization is performed on the standard normalization results according to the partition characteristics of each partition retrieval domain. The standard normalization results are first scaled based on the scaling parameters, and then the scaled standard normalization results are offset based on the offset parameters to generate common sample data.

[0140] Optionally, in one implementation, the above step S223 includes the following steps:

[0141] Get the shared scaling parameters and shared offset parameters corresponding to the shared convolutional network;

[0142] For any scaling parameter, offset parameter and standard normalization result corresponding to a search domain, generate partition sample data corresponding to the search domain in a preset manner;

[0143] Traverse multiple search domains to generate multiple partition sample data;

[0144] Multiple partition sample data are concatenated to generate common sample data.

[0145] For example, in this embodiment, the shared convolutional network is continuously trained to determine the shared scaling parameters and shared offset parameters corresponding to the shared convolutional network. For the scaling parameters and offset parameters corresponding to any retrieval domain and the standard normalization result, the partitioned sample data corresponding to the retrieval domain is generated in a preset manner based on the shared scaling parameters and shared offset parameters of the shared convolutional network. For multiple retrieval domains, multiple partitioned sample data are generated, and the generated multiple partitioned sample data are spliced ​​to generate shared sample data.

[0146] Optionally, the preset mode is:

[0147] z=(r*r i )*A+(b+b i )

[0148] Among them, z is the partition sample data, r is the shared scaling parameter, and r i is the scaling parameter, A is the standard normalization result, b is the shared offset parameter, and b i is the offset parameter. After scaling the scaling parameters corresponding to each search domain based on the shared scaling data and the shared offset data, the scaling results are offset to generate the partitioned sample data corresponding to each search domain. The partitioned sample data are spliced ​​to generate the shared sample data.

[0149] Through the above method, the characteristics of each partition are used to partition and normalize the common features of multiple domains, retaining the scene difference information of different domains, which is convenient for subsequent refined model building.

[0150] Figure 8 is a flow chart of a model training method according to an exemplary embodiment. Figure 8 As shown, the method is applied to a terminal and includes the following steps.

[0151] (1) obtaining user behavior sequences, user profiles, target project data, and domain indication information of multiple domains, wherein the domain indication information is domain ID information;

[0152] (2) inputting the above information into the shared embedding layer of the model, performing partition normalization processing on the information based on the partition characteristics of each partition, and concatenating the partition data after the partition normalization processing to generate normalized sample data;

[0153] (3) inputting the sample data into a convolutional network layer for model training, wherein the convolutional network layer includes a star convolutional network with a star topology structure, and the model parameters of the star convolutional network are obtained by element-wise multiplication of a shared convolutional network and independent convolutional networks in each domain;

[0154] (4) Based on the sample data, the star-shaped convolutional network corresponding to each domain is trained to generate a target convolutional network for each domain. The target convolutional network is used to identify the user's intention based on the search content input by the user and the user's behavior data, and retrieve the push data that meets the user's intention from the database corresponding to the domain based on the user's intention.

[0155] Through the above method, the convolutional network model of each domain is trained based on the common features of multiple domains, which avoids the cold start problem while retaining the characteristics of each domain, so that the trained convolutional network can push data based on the real needs of users, thereby improving the user experience.

[0156] Fig. 9 is a block diagram of a data push device according to an exemplary embodiment. Fig. 9 The device 100 includes: a determination module 110 and an execution module 120.

[0157] The determination module 110 is configured to determine, in response to the target retrieval information received in the target retrieval domain, a target convolutional network corresponding to the target retrieval domain from a plurality of specific convolutional networks, wherein the plurality of specific convolutional networks are generated by training a plurality of individual convolutional networks of a plurality of retrieval domains according to the common sample data, wherein a individual convolutional network is generated by element-level interaction between a shared convolutional network and an initial convolutional network corresponding to the plurality of retrieval domains;

[0158] The execution module 120 is configured to analyze the target retrieval information based on the target convolutional network and generate push data.

[0159] Optionally, the apparatus 100 further includes a generating module, and the generating module includes:

[0160] A first generating submodule is configured to generate multi-domain common features according to user behaviors in the multiple search domains;

[0161] The second generating submodule is configured to perform partition normalization processing on the common features of the multiple domains based on a plurality of learnable parameters corresponding one to one to the multiple retrieval domains, so as to generate the common sample data.

[0162] Optionally, the second generation submodule includes:

[0163] A first generating unit is configured to compare a plurality of initial convolutional networks corresponding to a plurality of search domains with a shared convolutional network to generate a plurality of learnable parameters, wherein the plurality of learnable parameters include a plurality of scaling parameters and a plurality of offset parameters;

[0164] The second generating unit is configured to perform standard normalization processing on the common features of multiple domains to generate a standard normalization result;

[0165] The third generating unit is configured to perform batch normalization on multiple scaling parameters, multiple offset parameters and standard normalization results to generate common sample data.

[0166] Optionally, the third generating unit is configured to:

[0167] Get the shared scaling parameters and shared offset parameters corresponding to the shared convolutional network;

[0168] For any scaling parameter, offset parameter and standard normalization result corresponding to a search domain, generate partition sample data corresponding to the search domain in a preset manner;

[0169] Traverse multiple search domains to generate multiple partition sample data;

[0170] Multiple partition sample data are concatenated to generate common sample data.

[0171] Optionally, the apparatus 100 further includes a fourth generating module, and the fourth generating module is configured to:

[0172] Generate a shared convolutional network based on the common features of multiple initial convolutional networks corresponding to multiple retrieval domains;

[0173] For any retrieval domain corresponding to the multiple retrieval domains, comparing the shared convolutional network with a first initial convolutional network corresponding to the retrieval domain to generate an offset parameter of the retrieval domain;

[0174] Perform element-level interaction between the shared convolutional network and the corresponding initial convolutional network to generate a fused convolutional network.

[0175] According to the offset parameters of the fused convolutional network and the retrieval domain, a personalized convolutional network is generated.

[0176] Optionally, the execution module 120 is configured to:

[0177] Analyze the target retrieval information based on the target convolutional network and generate the target retrieval intent;

[0178] According to the target search intent, multiple degrees of fit corresponding to multiple push data are determined;

[0179] The plurality of push data are sorted in descending order according to the plurality of fits to generate a data push list, wherein the data push list includes the sorted plurality of push data.

[0180] Optionally, the execution module 150 is configured to:

[0181] Obtain identity feature information of the target search domain;

[0182] Based on identity feature information and common features between multiple retrieval domains, a shared auxiliary network is trained to generate a target auxiliary network.

[0183] The target retrieval information is analyzed through the target auxiliary network and the target convolutional network to generate a data push list.

[0184] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0185] The present disclosure also provides a computer-readable storage medium having computer program instructions stored thereon, which implement the steps of the data recommendation method provided by the present disclosure when the program instructions are executed by a processor.

[0186] Fig.10 1 is a block diagram of an electronic device 1000 according to an exemplary embodiment. For example, the electronic device 1000 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0187] Reference Fig.10 , the electronic device 1000 may include one or more of the following components: a processing component 1002 , a memory 1004 , a power component 1006 , a multimedia component 1009 , an audio component 1010 , an input / output interface 1012 , a sensor component 1014 , and a communication component 1016 .

[0188] The processing component 1002 generally controls the overall operation of the electronic device 1000, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 1002 may include one or more processors 1020 to execute instructions to complete all or part of the steps of the above-mentioned data recommendation method. In addition, the processing component 1002 may include one or more modules to facilitate the interaction between the processing component 1002 and other components. For example, the processing component 1002 may include a multimedia module to facilitate the interaction between the multimedia component 1009 and the processing component 1002.

[0189] The memory 1004 is configured to store various types of data to support operations on the electronic device 1000. Examples of such data include instructions for any application or method operating on the electronic device 1000, contact data, phone book data, messages, pictures, videos, etc. The memory 1004 may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0190] The power supply component 1006 provides power to various components of the electronic device 1000. The power supply component 1006 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the electronic device 1000.

[0191] The multimedia component 1009 includes a screen that provides an output interface between the electronic device 1000 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundaries of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 1009 includes a front camera and / or a rear camera. When the electronic device 1000 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and the rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.

[0192] The audio component 1010 is configured to output and / or input audio signals. For example, the audio component 1010 includes a microphone (MIC), and when the electronic device 1000 is in an operation mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in the memory 1004 or sent via the communication component 1016. In some embodiments, the audio component 1010 also includes a speaker for outputting audio signals.

[0193] The input / output interface 1012 provides an interface between the processing component 1002 and the peripheral interface modules, which may be keyboards, click wheels, buttons, etc. These buttons may include but are not limited to: a home button, a volume button, a start button, and a lock button.

[0194] The sensor assembly 1014 includes one or more sensors for providing various aspects of status assessment for the electronic device 1000. For example, the sensor assembly 1014 can detect the open / closed state of the electronic device 1000, the relative positioning of components, such as the display and keypad of the electronic device 1000, and the sensor assembly 1014 can also detect the position change of the electronic device 1000 or a component of the electronic device 1000, the presence or absence of user contact with the electronic device 1000, the orientation or acceleration / deceleration of the electronic device 1000, and the temperature change of the electronic device 1000. The sensor assembly 1014 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor assembly 1014 may also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 1014 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0195] The communication component 1016 is configured to facilitate wired or wireless communication between the electronic device 1000 and other devices. The electronic device 1000 can access a wireless network based on a communication standard, such as WiFi, 2G or 3G, or a combination thereof. In an exemplary embodiment, the communication component 1016 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 1016 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0196] In an exemplary embodiment, the electronic device 1000 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above-mentioned data recommendation method.

[0197] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 1004 including instructions, and the instructions can be executed by the processor 1020 of the electronic device 1000 to complete the above data recommendation method. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0198] In addition to being an independent electronic device, the above-mentioned device can also be a part of an independent electronic device. For example, in one embodiment, the device can be an integrated circuit (IC) or a chip, wherein the integrated circuit can be an IC or a collection of multiple ICs; the chip can include but is not limited to the following types: GPU (Graphics Processing Unit), CPU (Central Processing Unit), FPGA (Field Programmable Gate Array), DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), SOC (System on Chip, SoC), etc. The above-mentioned integrated circuit or chip can be used to execute executable instructions (or codes) to implement the above-mentioned data recommendation method. The executable instructions can be stored in the integrated circuit or chip, or can be obtained from other devices or equipment, such as the integrated circuit or chip including a processor, a memory, and an interface for communicating with other devices. The executable instruction may be stored in the memory, and when the executable instruction is executed by the processor, the above-mentioned data recommendation method may be implemented; alternatively, the integrated circuit or chip may receive the executable instruction through the interface and transmit it to the processor for execution, so as to implement the above-mentioned data recommendation method.

[0199] In another exemplary embodiment, a computer program product is further provided. The computer program product includes a computer program executable by a programmable device, and the computer program has a code portion for executing the above data recommendation method when executed by the programmable device.

[0200] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the present disclosure. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0201] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A data recommendation method, characterized in that: include: In response to the target retrieval information received in the target retrieval domain, a target convolutional network corresponding to the target retrieval domain is determined from a plurality of specific convolutional networks, wherein the plurality of specific convolutional networks are generated by training a plurality of individual convolutional networks of a plurality of retrieval domains according to the shared sample data, wherein a individual convolutional network is generated by element-level interaction between the shared convolutional network corresponding to the plurality of retrieval domains and the initial convolutional network; The target retrieval information is analyzed based on the target convolutional network to generate push data.

2. The method according to claim 1, characterized in that The method comprises: generating multi-domain common features according to user behaviors in the multiple search domains; Based on a plurality of learnable parameters corresponding one to one to the plurality of search domains, partition normalization processing is performed on the common features of the plurality of domains to generate the common sample data.

3. The method according to claim 2, characterized in that The method of performing partition normalization processing on the common features of the multiple domains based on the multiple learnable parameters corresponding to the multiple search domains one by one to generate common sample data includes: Comparing the multiple initial convolutional networks corresponding to the multiple search domains with the shared convolutional network to generate the multiple learnable parameters, wherein the multiple learnable parameters include multiple scaling parameters and multiple offset parameters; Performing standard normalization processing on the multi-domain common features to generate a standard normalization result; Batch normalization is performed on the multiple scaling parameters, the multiple offset parameters, and the standard normalization result to generate the common sample data.

4. The method according to claim 3, characterized in that The batch normalizing the multiple scaling parameters, the multiple offset parameters and the standard normalization result to generate the common sample data comprises: Obtaining shared scaling parameters and shared offset parameters corresponding to the shared convolutional network; For any scaling parameter, offset parameter and the standard normalization result corresponding to any search domain, generating partition sample data corresponding to the search domain in a preset manner; Traversing the multiple search domains to generate multiple partition sample data; The plurality of partitioned sample data are concatenated to generate the common sample data.

5. The method according to claim 1, characterized in that The method further comprises: Generating the shared convolutional network according to the common features of the multiple initial convolutional networks corresponding to the multiple search domains; For any retrieval domain corresponding to the multiple retrieval domains, comparing the shared convolutional network with an initial convolutional network corresponding to the retrieval domain to generate an offset parameter of the retrieval domain; Performing element-level interaction between the shared convolutional network and the corresponding initial convolutional network to generate a fused convolutional network; The personalized convolutional network is generated according to the offset parameters of the fused convolutional network and the retrieval domain.

6. The method according to any one of claims 1 to 5, characterized in that The step of analyzing the target retrieval information based on the target convolutional network to generate push data includes: Analyzing the target retrieval information based on the target convolutional network to generate a target retrieval intent; According to the target search intention, determining a plurality of degrees of fit corresponding to the plurality of pushed data one by one; The plurality of push data are sorted in descending order of the plurality of fits to generate a data push list, wherein the data push list includes the sorted plurality of push data.

7. The method according to any one of claims 1 to 5, characterized in that The step of analyzing the target retrieval information based on the target convolutional network to generate push data includes: Acquire identity feature information of the target search domain; Based on the identity feature information and the common features among the multiple search domains, a shared auxiliary network is trained to generate a target auxiliary network; The target retrieval information is analyzed by the target auxiliary network and the target convolutional network to generate the data push list.

8. A data recommendation device, characterized in that: The device comprises: A determination module is configured to determine, in response to target retrieval information received in a target retrieval domain, a target convolutional network corresponding to the target retrieval domain from a plurality of specific convolutional networks, wherein the plurality of specific convolutional networks are generated by training a plurality of individual convolutional networks of a plurality of retrieval domains according to shared sample data, wherein a specific convolutional network is generated by element-level interaction between a shared convolutional network and an initial convolutional network corresponding to the plurality of retrieval domains; An execution module is configured to analyze the target retrieval information based on the target convolutional network to generate push data.

9. An electronic device, characterized in that: include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

Images