Content recommendation method, device, electronic device and storage medium

Abstract

The present disclosure relates to a content recommendation method, device, electronic device, and storage medium, and belongs to the field of computers. The content recommendation method includes: obtaining historical content sequence data of a target account and content attribute data of content data to be recommended; determining the total similarity between the historical content sequence data and the content attribute data; based on the historical content sequence data, the content attribute data, and the total similarity, obtaining the predicted results of the target account's operation events on each content data in the historical content sequence data; and according to the predicted results, performing content recommendations based on the content data to be recommended to the target account. The present disclosure realizes information interaction between historical data and target content, so that the calculation of the prediction result can integrate the total similarity information, achieving the effect of using the total similarity to correct the prediction result, which is conducive to improving the accuracy of the prediction result.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a content recommendation method, device, electronic device, and storage medium. Background Art

[0002] With the explosive growth of information data, personalized recommendations have become a standard feature of recommendation systems, and the quality of recommendation systems directly affects the user experience. A good recommendation system can recommend potential content of interest (such as photo albums, videos, and live broadcasts).

[0003] Existing recommendation systems usually make recommendations based on historical events (such as historical viewing content), or by calculating the similarity between target accounts or the similarity between content, and then recommending content liked by other target accounts similar to the current target account to the current target account. This is not a good way to recommend content that the target account cares about or pays attention to. Summary of the Invention

[0004] The present application provides a content recommendation method, device, electronic device, and storage medium to at least solve the problem of accuracy of content recommendation in related technologies.

[0005] According to a first aspect of the embodiments disclosed in this application, a content recommendation method is provided, comprising:

[0006] Obtaining historical content sequence data of the target account and content attribute data of the content data to be recommended;

[0007] determining the total similarity between the historical content sequence data and the content attribute data;

[0008] Obtaining, based on the historical content sequence data, the content attribute data, and the total similarity, a prediction result of the target account's operation event on each content data in the historical content sequence data;

[0009] According to the prediction result, content recommendation based on the to-be-recommended content data is performed for the target account.

[0010] As a possible implementation method, determining the total similarity between the historical content sequence data and the content attribute data includes: using a first tower model to obtain a feature average vector of the historical content sequence data; using a second tower model to obtain a content attribute feature vector of the content attribute data; and using the feature average vector and the content attribute feature vector to determine the total similarity between the historical content sequence data and the content attribute data.

[0011] As a possible implementation manner, the use of the first tower model to obtain the feature average vector of the historical content sequence data includes: using the first neural network model to project each of the content data in the historical content sequence data to obtain the first historical content feature vector of each of the content data, the first historical content feature vector having a first dimension, and using the first tower model to obtain the feature average vector of the first historical content feature vectors of all the content data in the historical content sequence data.

[0012] As a possible implementation, the obtaining of the content attribute feature vector of the content attribute data by using the second tower model includes: projecting the content attribute data by using the second tower model to obtain a content attribute feature vector having a first dimension.

[0013] As a possible implementation method, the use of the feature average vector and the content attribute feature vector to determine the total similarity between the historical content sequence data and the content attribute data includes: obtaining a first dot product result of the feature average vector and the content feature vector, and the first dot product result is used to represent the total similarity.

[0014] As a possible implementation manner, the use of the feature average vector and the content attribute feature vector to determine the total similarity between the historical content sequence data and the content attribute data includes: obtaining the account attribute data of the target account or the account attribute data and its embedded feature data, and extracting the sequence feature information of the historical content sequence data, inputting the sequence feature information into the first tower model to obtain a first output result with a second dimension, the first output result is used to characterize the event operation characteristics of the target account, inputting the content attribute data into the second tower model to obtain a second output result with a second dimension, the second output result is used to characterize the content attribute feature vector, merging the first output result and the feature average vector in the vector dimension to obtain a first merged result, merging the second output result and the content attribute feature vector in the vector dimension to obtain a second merged result, and determining a second dot product result of the first merged result and the second merged result, the second dot product result including a dot integral component representing the total similarity.

[0015] As a possible implementation manner, the prediction result of the target account's operation event on each content data in the historical content sequence data based on the historical content sequence data, the content attribute data and the total similarity includes: normalizing the second dot product result to obtain the prediction result.

[0016] As a possible implementation manner, the prediction result of the target account's operation event on each content data in the historical content sequence data based on the historical content sequence data, the content attribute data and the total similarity includes: obtaining the account attribute data of the target account or the account attribute data and its embedded feature data, and extracting the sequence feature information of the historical content sequence data, inputting the sequence feature information into the first tower model to obtain a first output result with a second dimension, and the first output result is used to characterize the event operation characteristics of the target account, inputting the content attribute data into the second tower model to obtain a second output result with a second dimension, and the second output result is used to characterize the content attribute feature vector, determining a third dot product result of the first output result and the second output result, and normalizing the sum of the third dot product result and the first dot product result to obtain the prediction result.

[0017] As a possible implementation manner, the extracting of sequence feature information of the historical content sequence data includes: using a second neural network model to project the account attribute data or the account attribute data and its embedded feature data to obtain an account attribute feature vector having a second dimension, using a third neural network model to project each content data in the historical content sequence data to obtain a second historical content feature vector for each content data, the second historical content feature vector having the second dimension, determining a first similarity between the account attribute feature vector and the second historical content feature vector for each content data, determining a first attention weight of the account attribute feature vector for each content data based on the first similarity, using a fourth neural network model to project each content data in the historical content sequence data to obtain a third historical content feature vector for each content data, the third historical content feature vector having the second dimension, determining a fourth dot product result of the first attention weight of each content data and the third historical content feature vector, accumulating the fourth dot product results of all the content data to obtain the sequence feature information.

[0018] As a possible implementation manner, the extracting of sequence feature information of the historical content sequence data includes: using a fifth neural network model to project each of the content data in the historical content sequence data to obtain a fourth historical content feature vector for each of the content data, using a sixth neural network model to project each of the content data in the historical content sequence data to obtain a fifth historical content feature vector for each of the content data, using a seventh neural network model to project each of the content data in the historical content sequence data to obtain a sixth historical content feature vector for each of the content data, determining a second similarity between the fourth historical content feature vector and the fifth historical content feature vector for each of the content data, determining a second attention weight of the fifth historical content feature vector of each of the content data to the fourth historical content feature vector based on the second similarity, and determining a fifth dot product result of the second attention weight of each of the content data and the sixth historical content feature vector, wherein the fifth dot product result is used to characterize the sequence feature information, wherein the fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

[0019] According to a second aspect of the embodiments disclosed in this application, a method for training a content recommendation model is provided, comprising:

[0020] Obtaining sample content sequence data and content attribute sample data of the sample content data of the sample account;

[0021] Inputting the sample content sequence data and the content attribute sample data into a content recommendation model, the content recommendation model being configured to determine a total similarity between the sample content sequence data and the content attribute sample data, and obtaining, based on the sample content sequence data, the content attribute sample data, and the total similarity, a prediction result of an operation event of the sample account for each sample content data in the sample content sequence data;

[0022] Determining a target loss function value of the content recommendation model based on the prediction result output by the content recommendation model,

[0023] According to the target loss function value, the model parameters of the content recommendation model are updated until the training end condition is met.

[0024] As a possible implementation manner, determining the total similarity between the historical content sequence sample data and the content attribute sample data includes: using a first tower model to obtain a feature average vector of the historical content sequence sample data, and using a second tower model to obtain a content attribute feature vector of the content attribute sample data; and using the feature average vector and the content attribute feature vector to determine the total similarity between the sample content sequence data and the content attribute sample data.

[0025] As a possible implementation manner, the use of the first tower model to obtain the feature average vector of the sample content sequence data includes: using a first neural network model to project each sample content data in the sample content sequence data to obtain a first historical content feature vector of each sample content data, the first historical content feature vector having a first dimension; and using the first tower model to obtain the feature average vector of the first historical content feature vectors of all the sample content data in the sample content sequence data.

[0026] As a possible implementation, the obtaining of the content attribute feature vector of the content attribute sample data by using the second tower model includes: projecting the content attribute sample data by using the second tower model to obtain a content attribute feature vector having a first dimension.

[0027] As a possible implementation manner, the use of the feature average vector and the content attribute feature vector to determine the total similarity between the sample content sequence data and the content attribute sample data includes: obtaining a first dot product result of the feature average vector and the content feature vector, the first dot product result being used to characterize the total similarity.

[0028] As a possible implementation manner, the use of the feature average vector and the content attribute feature vector to determine the total similarity between the sample content sequence data and the content attribute sample data includes: obtaining the account attribute sample data of the sample account or the account attribute sample data and its embedded feature data, and extracting the sequence feature information of the sample content sequence data, inputting the sequence feature information into the first tower model to obtain a first output result with a second dimension, the first output result is used to characterize the event operation characteristics of the sample account, inputting the content attribute sample data into the second tower model to obtain a second output result with a second dimension, the second output result is used to characterize the content attribute feature vector, merging the first output result and the feature average vector in the vector dimension to obtain a first merged result, merging the second output result and the content attribute feature vector in the vector dimension to obtain a second merged result, and determining a second dot product result of the first merged result and the second merged result, the second dot product result including a dot integral component representing the total similarity.

[0029] As a possible implementation manner, obtaining the prediction result of the sample account's operation event on each sample content data in the sample content sequence data based on the sample content sequence data, the content attribute sample data and the total similarity includes: normalizing the second dot product result to obtain the prediction result.

[0030] As a possible implementation manner, the method of obtaining a prediction result of the operation event of the sample account on each sample content data in the sample content sequence data based on the sample content sequence data, the content attribute sample data and the total similarity includes: obtaining account attribute sample data or the account attribute sample data and its embedded feature data, and extracting sequence feature information of the historical content sequence sample data, inputting the sequence feature information into the first tower model to obtain a first output result with a second dimension, and the first output result is used to characterize the event operation feature of the sample account, inputting the content attribute sample data into the second tower model to obtain a second output result with a second dimension, and the second output result is used to characterize the content attribute feature vector, determining a third dot product result of the first output result and the second output result, and normalizing the sum of the third dot product result and the first dot product result to obtain the prediction result.

[0031] As a possible implementation manner, the extracting sequence feature information of the sample content sequence data includes: using a second neural network model to project the account attribute sample data or the account attribute sample data and its embedded feature sample data to obtain an account attribute feature vector having a second dimension; using a third neural network model to project each sample content data in the sample content sequence data to obtain a second historical content feature vector for each sample content data, the second historical content feature vector having the second dimension; determining a first similarity between the account attribute feature vector and the second historical content feature vector of each sample content data; determining a first attention weight of the account attribute feature vector for each sample content data based on the first similarity; using a fourth neural network model to project each sample content data in the sample content sequence data to obtain a third historical content feature vector for each sample content data, the third historical content feature vector having the second dimension; determining a fourth dot product result between the first attention weight of each sample content data and the third historical content feature vector; and accumulating the fourth dot product results of all the content data to obtain the sequence feature information.

[0032] As a possible implementation manner, the extracting sequence feature information of the sample content sequence data includes: using a fifth neural network model to project each of the sample content data in the sample content sequence data to obtain a fourth historical content feature vector for each of the content data, using a sixth neural network model to project each of the sample content data in the sample content sequence data to obtain a fifth historical content feature vector for each of the content data, using a seventh neural network model to project each of the sample content data in the sample content sequence data to obtain a sixth historical content feature vector for each of the content data, determining a second similarity between the fourth historical content feature vector and the fifth historical content feature vector for each of the sample content data, determining a second attention weight of the fifth historical content feature vector of each of the sample content data to the fourth historical content feature vector based on the second similarity, and determining a fifth dot product result of the second attention weight of each of the sample content data and the sixth historical content feature vector, wherein the fifth dot product result is used to characterize the sequence feature information, wherein the fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

[0033] According to a third aspect of the embodiments disclosed in this application, a content recommendation device is provided, including:

[0034] An acquisition unit is configured to acquire the historical content sequence data of the target account and the content attribute data of the content data to be recommended,

[0035] a determining unit configured to determine the total similarity between the historical content sequence data and the content attribute data,

[0036] a prediction unit configured to obtain a prediction result of the target account's operation event on each content data in the historical content sequence data based on the historical content sequence data, the content attribute data, and the total similarity;

[0037] The output unit is configured to perform content recommendation based on the content data to be recommended to the target account.

[0038] As a possible implementation, the determination unit includes: a vector mean acquisition module, configured to use a first tower model to obtain a feature average vector of the historical content sequence data; an attribute feature acquisition module, configured to use a second tower model to obtain a content attribute feature vector of the content attribute data; and a total similarity determination module, configured to use the feature average vector and the content attribute feature vector to determine the total similarity between the historical content sequence data and the content attribute data.

[0039] As a possible implementation, the vector mean acquisition module is specifically configured to: use the first neural network model to project each content data in the historical content sequence data to obtain a first historical content feature vector for each content data, wherein the first historical content feature vector has a first dimension; and use the first tower model to obtain the feature average vector of the first historical content feature vectors of all the content data in the historical content sequence data.

[0040] As a possible implementation, the attribute feature acquisition module is specifically configured to: project the content attribute data using a second tower model to obtain a content attribute feature vector having a first dimension.

[0041] As a possible implementation manner, the total similarity determination module is specifically configured to: obtain a first dot product result of the feature average vector and the content feature vector to obtain the total similarity.

[0042] As a possible implementation manner, the total similarity determination module is specifically configured to: obtain the account attribute data of the target account or the account attribute data and its embedded feature data, and extract the sequence feature information of the historical content sequence data, input the sequence feature information into the first tower model to obtain a first output result with a second dimension, and the first output result is used to characterize the event operation characteristics of the target account, input the content attribute data into the second tower model to obtain a second output result with a second dimension, and the second output result is used to characterize the content attribute feature vector, merge the first output result and the feature average vector in the vector dimension to obtain a first merged result, merge the second output result and the content attribute feature vector in the vector dimension to obtain a second merged result, determine a second dot product result of the first merged result and the second merged result, and the second dot product result includes a dot integral component representing the total similarity.

[0043] As a possible implementation, the prediction unit is specifically configured to: normalize the second dot product result to obtain the prediction result.

[0044] As a possible implementation manner, the prediction unit is specifically configured to: obtain the account attribute data of the target account or the account attribute data and its embedded feature data, and extract the sequence feature information of the historical content sequence data, input the sequence feature information into the first tower model to obtain a first output result with a second dimension, and the first output result is used to characterize the event operation characteristics of the target account, input the content attribute data into the second tower model to obtain a second output result with a second dimension, and the second output result is used to characterize the content attribute feature vector, determine the third dot product result of the first output result and the second output result, normalize the sum of the third dot product result and the first dot product result, and obtain the prediction result.

[0045] As a possible implementation manner, the prediction unit is specifically configured to: use a second neural network model to project the account attribute data or the account attribute data and its embedded feature data to obtain an account attribute feature vector with a second dimension; use a third neural network model to project each content data in the historical content sequence data to obtain a second historical content feature vector for each content data, the second historical content feature vector has the second dimension; determine a first similarity between the account attribute feature vector and the second historical content feature vector for each content data; determine a first attention weight of the account attribute feature vector for each content data based on the first similarity; use a fourth neural network model to project each content data in the historical content sequence data to obtain a third historical content feature vector for each content data, the third historical content feature vector has the second dimension; determine a fourth dot product result of the first attention weight of each content data and the third historical content feature vector; accumulate the fourth dot product results of all the content data to obtain the sequence feature information.

[0046] As a possible implementation manner, the prediction unit is specifically configured to: use the fifth neural network model to project each of the content data in the historical content sequence data to obtain the fourth historical content feature vector of each of the content data, use the sixth neural network model to project each of the content data in the historical content sequence data to obtain the fifth historical content feature vector of each of the content data, use the seventh neural network model to project each of the content data in the historical content sequence data to obtain the sixth historical content feature vector of each of the content data, determine the second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each of the content data, determine the second attention weight of the fifth historical content feature vector of each of the content data to the fourth historical content feature vector based on the second similarity, determine the fifth dot product result of the second attention weight of each of the content data and the sixth historical content feature vector, and the fifth dot product result is used to characterize the sequence feature information, wherein the fourth historical content feature vector, the fifth historical content feature vector and the sixth historical content feature vector all have a third dimension.

[0047] According to a fourth aspect of the embodiments disclosed in this application, a training device for a content recommendation model is provided, comprising:

[0048] an acquisition unit configured to acquire sample content sequence data and content attribute sample data of the sample content data of the sample account,

[0049] The training unit is configured as follows:

[0050] The sample content sequence data and the content attribute sample data are input into a content recommendation model, wherein the content recommendation model is used to determine the total similarity between the sample content sequence data and the content attribute sample data, and based on the sample content sequence data, the content attribute sample data, and the total similarity, obtain a prediction result of the operation event of the sample account for each sample content data in the sample content sequence data.

[0051] Determining a target loss function value of the content recommendation model based on the prediction result output by the content recommendation model,

[0052] According to the target loss function value, the model parameters of the content recommendation model are updated until the training end condition is met.

[0053] As a possible implementation, the content recommendation model is configured to: utilize a first tower model to obtain a feature average vector of the historical content sequence sample data, and utilize a second tower model to obtain a content attribute feature vector of the content attribute sample data; and utilize the feature average vector and the content attribute feature vector to determine the total similarity between the sample content sequence data and the content attribute sample data.

[0054] As a possible implementation, the content recommendation model is specifically configured as follows: using a first neural network model to project each sample content data in the sample content sequence data to obtain a first historical content feature vector for each sample content data, wherein the first historical content feature vector has a first dimension; and using the first tower model to obtain the feature average vector of the first historical content feature vectors of all the sample content data in the sample content sequence data.

[0055] As a possible implementation, the content recommendation model is specifically configured to: project the content attribute sample data using a second tower model to obtain a content attribute feature vector having a first dimension.

[0056] As a possible implementation manner, the content recommendation model is specifically configured to: obtain a first dot product result of the feature average vector and the content feature vector, where the first dot product result is used to represent the total similarity.

[0057] As a possible implementation manner, the content recommendation model is specifically configured as follows: obtaining account attribute sample data of the sample account or the account attribute sample data and its embedded feature data, and extracting sequence feature information of the sample content sequence data, inputting the sequence feature information into the first tower model, obtaining a first output result having a second dimension, and the first output result is used to characterize the event operation characteristics of the sample account, inputting the content attribute sample data into the second tower model, obtaining a second output result having a second dimension, and the second output result is used to characterize the content attribute feature vector, merging the first output result and the feature average vector in the vector dimension to obtain a first merged result, merging the second output result and the content attribute feature vector in the vector dimension to obtain a second merged result, determining a second dot product result of the first merged result and the second merged result, and the second dot product result includes a dot integral component representing the total similarity.

[0058] As a possible implementation, the content recommendation model is specifically configured to: normalize the second dot product result to obtain the prediction result.

[0059] As a possible implementation manner, the content recommendation model is specifically configured as follows: obtaining account attribute sample data or the account attribute sample data and its embedded feature data, and extracting sequence feature information of the historical content sequence sample data, inputting the sequence feature information into the first tower model, obtaining a first output result with a second dimension, and the first output result is used to characterize the event operation characteristics of the sample account, inputting the content attribute sample data into the second tower model, obtaining a second output result with a second dimension, and the second output result is used to characterize the content attribute feature vector, determining a third dot product result of the first output result and the second output result, normalizing the sum of the third dot product result and the first dot product result, and obtaining the prediction result.

[0060] As a possible implementation manner, the content recommendation model is specifically configured as follows: using a second neural network model to project the account attribute sample data or the account attribute sample data and its embedded feature sample data to obtain an account attribute feature vector having a second dimension; using a third neural network model to project each sample content data in the sample content sequence data to obtain a second historical content feature vector for each sample content data, the second historical content feature vector having the second dimension; determining a first similarity between the account attribute feature vector and the second historical content feature vector of each sample content data; determining a first attention weight of the account attribute feature vector for each sample content data based on the first similarity; using a fourth neural network model to project each sample content data in the sample content sequence data to obtain a third historical content feature vector for each sample content data, the third historical content feature vector having the second dimension; determining a fourth dot product result between the first attention weight of each sample content data and the third historical content feature vector; accumulating the fourth dot product results of all the content data to obtain the sequence feature information.

[0061] As a possible implementation manner, the content recommendation model is specifically configured as follows: using a fifth neural network model to project each of the sample content data in the sample content sequence data to obtain a fourth historical content feature vector for each of the content data; using a sixth neural network model to project each of the sample content data in the sample content sequence data to obtain a fifth historical content feature vector for each of the content data; using a seventh neural network model to project each of the sample content data in the sample content sequence data to obtain a sixth historical content feature vector for each of the content data; determining a second similarity between the fourth historical content feature vector and the fifth historical content feature vector for each of the sample content data; determining a second attention weight of the fifth historical content feature vector of each of the sample content data to the fourth historical content feature vector based on the second similarity; determining a fifth dot product result of the second attention weight of each of the sample content data and the sixth historical content feature vector; the fifth dot product result is used to characterize the sequence feature information, wherein the fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

[0062] According to a fifth aspect of the embodiments disclosed in this application, there is provided an electronic device, including:

[0063] processor;

[0064] a memory for storing executable instructions for the processor;

[0065] The processor is configured to execute the executable instructions to implement any of the content recommendation methods and / or content model training methods.

[0066] According to the sixth aspect of the embodiments disclosed in the present application, a computer-readable storage medium is provided. When at least one instruction in the computer-readable storage medium is executed by a processor of an electronic device, the electronic device is enabled to implement the content recommendation method and / or content model training method described in any of the above-mentioned embodiments.

[0067] According to a seventh aspect of the embodiments disclosed in the present application, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the above-mentioned content recommendation method and / or content model training method.

[0068] The technical solutions provided by the embodiments disclosed in this application bring at least the following beneficial effects:

[0069] The content recommendation method provided in the embodiment of the present application realizes information interaction between historical data and target content through the total similarity between the historical content sequence data and the content attribute data, so that the calculation of the prediction result can integrate the total similarity information, thereby achieving the effect of correcting the prediction result by using the total similarity, which is conducive to improving the accuracy of the prediction result, thereby providing more accurate content recommendations.

[0070] It should 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

[0071] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the disclosure of this application, and together with the specification are used to explain the principles disclosed in this application, and do not constitute an improper limitation on the disclosure of this application.

[0072] Figure 1 is a flowchart of a recommendation method according to an exemplary embodiment;

[0073] Figure 2 is a flowchart of a recommendation method according to an exemplary embodiment;

[0074] Figure 3 This is a flowchart of a method for training a content recommendation model according to an exemplary embodiment;

[0075] Figure 4 is a schematic diagram of a content recommendation device according to an exemplary embodiment;

[0076] Figure 5is a schematic diagram of a content recommendation model training device according to an exemplary embodiment;

[0077] Figure 6 The figure is a schematic diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0078] In order to enable ordinary persons in the art to better understand the technical solutions disclosed in this application, the technical solutions in the embodiments disclosed in this application will be clearly and completely described below with reference to the accompanying drawings.

[0079] It should be noted that the terms "first," "second," and the like in the specification and claims disclosed herein and in the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments disclosed herein described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the disclosure herein. Instead, they are merely examples of apparatus and methods consistent with certain aspects of the disclosure herein, as detailed in the appended claims.

[0080] In a recommendation system, different content needs to be recommended for different target accounts. For example, different multimedia resources or different products can be used as content. The more accurate the recommended content, the better the user experience.

[0081] The recommendation process includes recall, rough sorting, fine sorting and re-sorting.

[0082] Recall processing is to select thousands of contents from the content library based on the characteristics of the target account and content, and obtain the recall processing results;

[0083] Rough sorting is to select hundreds of contents from the recall processing results (thousands of contents) to obtain the rough sorting results;

[0084] Fine ranking processing is to obtain the prediction results of different events of the rough ranking processing results (hundreds of contents). The events involved in the rough ranking processing may include clicks, views, comments, forwarding, etc. The prediction results can be represented by estimated values.

[0085] The re-ranking process is to sort the prediction results according to different events, and use the top-ranked content in the sorted content as one of the recommended results.

[0086] In the recall processing, rough ranking processing, and fine ranking processing, processing is performed by estimating the predicted results of various events of the content.

[0087] In the coarse ranking process, since each target account needs to calculate the prediction results of thousands of content and different events, considering the performance issues, a dual-tower structure is often used, including a target account (user) tower for processing the target account side feature data, and an object (item) tower for processing the content side feature data.

[0088] The first tower model is based on the characteristics of the target account side. A fixed-dimensional vector output is obtained through a neural network, which is defined as the target account feature and recorded as Uout. For example, the first tower model in the embodiment of the present application can be called the user tower.

[0089] The second tower model is based on the features on the content side, and obtains a vector output with the same dimension as the Uout vector through another neural network, which is defined as the content feature and recorded as Iout. For example, the second tower model in the embodiment of the present application can be called the item tower.

[0090] Calculate the dot product of Uout and Iout. After passing the dot product through the normalized exponential (Sigmoid) function, we get the prediction results of each event of the target account feature for the content feature. The expression is expressed as:

[0091] pred=sigmoid(Uout·Iout)

[0092] Among them, pred represents the prediction result.

[0093] Compared to the normal non-dual-tower approach, when calculating the estimated result of a target account for 1,000 pieces of content, it needs to be executed one-to-one 1,000 times. However, with this dual-tower approach, the final estimated result can be obtained by executing the first tower model (such as the user tower) once, executing the second tower model (such as the item tower) 1,000 times (generally using the cache Iout mechanism, the actual number of executions is far less than 1,000 times), and performing the dot product operation once. This method of splitting the target account and content into two towers for separate calculations greatly speeds up the running speed of the coarse sorting process, allowing the "from thousands to hundreds" screening to be executed normally.

[0094] Applicants have discovered that in a dual-tower architecture, the final output, Uout, is currently obtained by extracting information from the target account's features on the target account side. The target account also extracts information from internal content features to obtain the final output, Iout. There is no information fusion between the target account's features and the content's features. However, historical content sequences on the target account side, such as video playback history, browsing history, and location history, are highly correlated with content features on the content side. Providing direct feature fusion between these two would be beneficial to the accuracy of the coarse-grained ranking results.

[0095] In view of this, the present application provides a content recommendation method that fuses the feature information on both sides of the dual towers to utilize the correlation between the target account features and the content features to improve the accuracy of the rough sorting results, thereby improving the accuracy of the recommendation method.

[0096] See also Figure 1 As shown, Figure 1 FIG1 is a flow chart of a recommendation method according to an exemplary embodiment. The method includes:

[0097] Step 101: Obtain historical content sequence data of a target account and content attribute data of content data to be recommended;

[0098] Step 102: determining the total similarity between the acquired historical content sequence data and the content attribute data;

[0099] Step 103: Based on the acquired historical content sequence data and content attribute data, and the determined total similarity between the historical content sequence data and the content attribute data, obtain a prediction result of the target account's operation event for each content data in the historical content sequence data;

[0100] Step 104 : Recommend content to the target account based on the content data to be recommended according to the determined prediction result.

[0101] This embodiment allows the historical content sequence data on the target account side and the attribute data on the content side to directly interact with each other in features, thereby improving the estimation effect through this enhanced information interaction method.

[0102] For ease of understanding, the following explanation will be given using video recommendation as an example. It should be understood that this application is not limited to video recommendations, but can also be recommendations for multimedia resources such as pictures, articles, music, etc., or recommendations for any objects such as goods and electronic resources. These objects are collectively referred to as content.

[0103] For ease of description, the historical content sequence data on the target account side is recorded as: ActionList=[a1,a2,…,a N ], where a i Indicates the video data that has been effectively watched, N indicates the total number of videos that have been effectively watched; account attribute data is represented by uid; and the video attribute data on the target account side is represented by pid. i , uid, pid, as well as the basic features of the target account side, the basic features of the target account side video attributes, and the context features used are all obtained from the query vector table to obtain the vector corresponding to each feature, which is called embedding.

[0104] See also Figure 2 As shown, Figure 2FIG1 is a flow chart of a recommendation method according to an exemplary embodiment. The method includes:

[0105] Step 201: Obtain historical content sequence data of a target account and content attribute data of content data to be recommended.

[0106] Step 202: Obtain the feature average vector of the historical content sequence data.

[0107] As an example, step 202 may utilize the first neural network model to project each content data in the historical content sequence data to obtain a first historical content feature vector of each content data, where the first historical content feature vector has a first dimension.

[0108] For example, assuming that the content data includes video data, the N video data in the historical content sequence are projected into a k-dimensional first historical content feature vector, and a vector set B containing N k-dimensional first historical content feature vectors is obtained, where N is a positive integer greater than 1, the first dimension is k, and k is a positive integer greater than 1.

[0109] Among them, any video data a i The projected k-dimensional first historical content feature vector may be a video content vector and may be denoted as b i , and b i ∈R k ,

[0110] Furthermore, the first neural network model may be, for example, a multi-layer fully connected layer (MLP) neural network model. The first neural network model may be a first function, which may be expressed as the following expression (1):

[0111] funcA(x)=δ(w1x+bias1),

[0112] Wherein, w1 is the first weight parameter of the first neural network model, bias1 is the first bias parameter of the first neural network model, δ(x) is the activation function, and the activation function can adopt the softplus function, that is, δ(x)=log(1+e x ), x represents the input quantity input to the first neural network model, which is content data such as video data; thus, N k-dimensional first historical content feature vectors can be obtained. As mentioned above, the vector set of N first historical content feature vectors can be recorded as B = [b1, b2, ..., b N ].

[0113] Through the above method, the first historical content feature vector of all content data can be extracted from the historical content sequence data.

[0114] As an example, the first tower model (eg, user tower) in the dual-tower model used for processing target account-side feature data may be used to obtain a feature average vector of historical content sequence data.

[0115] Step 203: Obtain a content attribute feature vector of the content attribute data.

[0116] As an example, step 203 may project the content attribute data using the second tower model (eg, item tower) in the dual-tower model for content-side feature data processing to obtain a content attribute feature vector having a first dimension (eg, k dimension).

[0117] For example, assuming that the content attribute data includes video attribute data, each video attribute data pid is projected into a k-dimensional vector through the second tower model (such as the item tower), and a k-dimensional video attribute feature vector can be obtained, wherein the video attribute data may include video type, production time, video length, etc. The k-dimensional video attribute feature vector can be used as the content attribute feature vector and can be recorded as z, and z∈R k ;

[0118] The second tower model (e.g., item tower) can be an MLP neural network, which can implement the function of the second function, which can be expressed as the following expression (2):

[0119] funcI(x)=δ(w2x+bias2),

[0120] Wherein, w2 is the second weight parameter of the second tower model (for example, the item tower), bias2 is the second bias parameter of the second tower model (for example, the item tower), δ(x) is the activation function, which can adopt the softplus function, that is, δ(x)=log(1+e x ), x represents the input quantity, which is the content attribute data here.

[0121] Through this step, the content attribute feature vector is extracted from the content attribute data.

[0122] Step 204 : Determine the total similarity between the historical content sequence data and the content attribute data using the feature average vector of the historical content sequence data and the content attribute feature vector of the content attribute data.

[0123] As a first example of determining the total similarity, step 204 may perform a dot product between the feature average vector obtained by the first tower model (e.g., the user tower) and the content attribute feature vector obtained by the second tower model (e.g., the item tower) to obtain a first dot product result, which represents the total similarity between all content data of the historical content sequence and the content attribute data of the content data to be recommended.

[0124] For example, the total similarity can be expressed as follows:

[0125]

[0126] Since there may be an insufficient number of vectors in the vector set B of the first historical content feature vector, when the number of vectors is less than N, the number of vectors in the vector set b can be padded to N by adding 0 vectors. This means that the actual sequence length of the historical content sequence (that is, the actual number of content data in the historical content sequence) is not necessarily N. Therefore, size is used to represent the actual sequence length of the historical content sequence (that is, the actual number of content data in the historical content sequence). is the feature average vector, denoted as simU.

[0127] The calculation method for the above total similarity is:

[0128] 1. Calculate the content attribute feature vector on the content side and each content data a in the historical content sequence on the target account side i The similarity s i ∈R 1 , that is, by converting any content data (such as video data) a i The projected k-dimensional first historical content feature vector b i The dot product of the content attribute feature vector z can be used to obtain each content data (such as video data) a i Similarity s with content attribute feature vector z i , the similarity s i It can be expressed by the following expression (4):

[0129] s i =b i ·z

[0130] In this way, the similarity S between each video data and the video attribute data in the historical content sequence is obtained, which can be expressed as the following expression (5):

[0131] S=[s1,s2,…,s N ]

[0132] 2. Calculate the total similarity between the historical content sequence on the target account side and the content attribute data on the content side. The above expression (3) for expressing the total similarity can also be transformed into the following expression (6):

[0133]

[0134] As a second example of determining the total similarity, step 204 may specifically include:

[0135] a) Obtaining the target account’s account attribute data or the account attribute data and its embedded feature data;

[0136] b) Extracting sequence feature information of the acquired historical content sequence data.

[0137] c) Inputting the extracted sequence feature information into the first tower model to obtain a first output result having a second dimension, where the first output result is used to characterize the event operation characteristics of the target account; as an example, inputting the extracted sequence feature information into the first tower model (e.g., the user tower) to obtain an account attribute feature vector Uout of a second dimension (e.g., M dimension).

[0138] d) Inputting the acquired content attribute data into the second tower model to obtain a second output result having a second dimension, where the second output result is used to represent the content attribute feature vector, but the dimension is different from the content attribute feature vector acquired in step 203; as an example, passing the video attribute data on the content side through the second tower model (e.g., the item tower) to obtain a content attribute feature vector Iout of the second dimension (e.g., M dimension).

[0139] e) Merging the first output result and the feature average vector obtained in step 202 in a vector dimension to obtain a first merged result.

[0140] That is: merge the first output result Uout and the feature average vector simU in vector dimension to obtain the first merged result [Uout, simU], [Uout, simU]∈R M+K vector.

[0141] f) Merging the second output result and the content attribute feature vector obtained in step 203 in terms of vector dimension to obtain a second merged result.

[0142] That is, the second output result Iout and the content attribute feature vector z are merged in vector dimension to obtain the second merged result [Iout, z], [Iout, z]∈R M+K vector.

[0143] g) Determine a second dot product result of the first merging result and the second merging result, wherein the dot product component included in the second dot product result may represent the total similarity between all content data of the historical content sequence and content attribute data of the content data to be recommended.

[0144] Among them, the acquired historical content sequence data is subjected to an attention method or a self-attention method to extract sequence feature information.

[0145] The process of extracting sequence feature information using the attention method can specifically include:

[0146] (b1) Projecting the account attribute data of the target account or the account attribute data and its embedded feature data using the second neural network model to obtain an account attribute feature vector having a second dimension.

[0147] As an example, the account attribute data uid or other embedded feature data on the target account side is first projected into an account attribute feature vector with a second dimension through a third function funcq(x) to obtain an account attribute feature vector, for example, a second dimension (for example, M dimension) account attribute feature vector q. The function of the third function can be implemented by the second neural network model.

[0148] (b2) Projecting each content data in the historical content sequence data using the third neural network model to obtain a second historical content feature vector for each content data, where the second historical content feature vector has a second dimension.

[0149] For example, each content data (eg video data) in the historical content sequence a i All of them are projected into a second historical video feature vector with a second dimension by the fourth function funck(x), and N second historical video feature vectors of the second dimension (for example, M dimension) are obtained. The vector set of the second historical video feature vector can be recorded as: keys = [k1, k2, ..., k N ], the function of the fourth function can be realized by the third neural network model.

[0150] (b3) Determining a first similarity between the account attribute feature vector and a second historical content feature vector of each content data in the historical content sequence.

[0151] For example, by dot product calculation, the account attribute feature vector q and the second historical video feature k of each content data in the vector set keys are calculated. i The similarity s i , the similarity s i It can be expressed as the following expression (7):

[0152] s $ =qk i T

[0153] Among them, the symbol T represents the video feature vector k i The transpose of .

[0154] (b4) Determining a first attention weight of the account attribute feature vector for each content data in the historical content sequence based on a first similarity between the account attribute feature vector and a second historical content feature vector of each content data in the historical content sequence.

[0155] For example, calculate each historical video content feature vector k i For the first attention weight of q, the weight vector of the first attention weight of the account attribute feature vector to all content data in the historical content sequence can be expressed by the following expression (8):

[0156]

[0157] Among them, w i is the first attention weight of the account attribute feature vector to the i-th content data in the historical content sequence.

[0158] (b5) Projecting each content data in the historical content sequence data using the fourth neural network model to obtain a third historical content feature vector for each content data, where the third historical content feature vector has a second dimension.

[0159] For example, each content data (eg video data) in the historical content sequence data a i All of them are projected into the third historical content feature vector of the second dimension (for example, M dimension) by the fifth function funcv(x). The obtained set of N M-dimensional third historical content vectors can be recorded as: values ​​= [v1, v2, ..., v N ], the function of the fifth function can be realized by the fourth neural network model.

[0160] (b6) Determine the fourth dot product result of the first attention weight of each content data and the content data in the third historical content feature vector, and accumulate all the fourth dot product results to obtain sequence feature information.

[0161] For example, the sequence feature information is obtained by the following expression (9):

[0162]

[0163] The process of extracting sequence feature information using the self-attention method can specifically include:

[0164] (b1') Using the fifth neural network model to project each content data in the historical content sequence data, obtain the fourth historical content feature vector of each content data; using the sixth neural network model to project each content data in the historical content sequence data, obtain the fifth historical content feature vector of each content data; and using the seventh neural network model to project each content data in the historical content sequence data, obtain the sixth historical content feature vector of each content data, wherein the fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

[0165] For example, each content data (such as video data) in the ActionList is respectively processed by the fifth neural network model of the sixth function funcq(x), the sixth neural network model of the seventh function funck(x), and the seventh neural network model of the eighth function funcv(x). i Project N M-dimensional vectors, where:

[0166] The vector set containing N fourth historical content feature vectors can be recorded as Q = [q1, q2, ..., q N ];

[0167] The vector set containing N fifth historical content feature vectors can be recorded as K = [k1, k2, ..., k N ];

[0168] The vector set containing N sixth historical content feature vectors can be recorded as V = [v1, v2, ..., v N ].

[0169] (b2′) Determine the second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each content data. The similarity vector of the second similarity between the fourth historical content feature vector and the fifth historical content feature vector of all content data in the historical content sequence can be expressed by the following expression (10):

[0170] S=QK T

[0171] (b3′) Determine the second attention weight of the fifth historical content feature vector of each content data with respect to the fourth historical content feature vector based on the second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each content data. The weight vector of the second attention weights of the fifth historical content feature vector of all content data in the historical content sequence with respect to the fourth historical content feature vector can be expressed by the following expression (11):

[0172]

[0173] (b4′) Determine the fifth dot product of the second attention weight of each content data and the sixth historical content feature vector to obtain sequence feature information. The sequence feature information can be expressed as follows:

[0174]

[0175] The two methods generally adopt a multi-head approach, that is, to obtain multiple attention out Combine them together to get the combined attention out .

[0176] Step 205 : Based on the acquired historical content sequence data and content attribute data, and the determined total similarity between the historical content sequence data and the content attribute data, obtain the target account's prediction result of the operation event for each content data in the historical content sequence data.

[0177] For the case of the second example of determining the total similarity in step 204, step 205 normalizes the sum of the third dot product result and the first dot product result to obtain a prediction result.

[0178] For the case where step 204 adopts the first example of determining the total similarity, step 205 may specifically include:

[0179] Obtain the target account's account attribute data or the account attribute data and its embedded feature data, and extract sequence feature information from the historical content sequence data. The sequence feature information extraction here can use the aforementioned attention method or self-attention method, which will not be described in detail here.

[0180] Inputting the sequence feature information into the first tower model to obtain a first output result having a second dimension, where the first output result is used to characterize event operation features of the target account;

[0181] Inputting the content attribute data into the second tower model to obtain a second output result having a second dimension, wherein the second output result is used to represent a content attribute feature vector;

[0182] A third dot product result of the first output result and the second output result is determined.

[0183] The second dot product result [Uout,simU]·[Iout,z] can be considered to be equivalent to the sum of the third dot product result Uout·Iout of the first output result and the second output result and the total similarity simU·z.

[0184] This equivalent relationship can be expressed by the following expression (13):

[0185] pred=sigmoid([Uout, simU]·[Iout, z])=sigmoid(Uout·Iout+simU·z)

[0186] Among them, [Uout, simU] · [Iout, z] is the second dot product result, Uout · Iout is the third dot product result, and simU · z is the first dot product result.

[0187] As can be seen from the above expression (13), the prediction result is contributed by the third dot product result Uout·Iout and the first dot product result (total similarity) simU·z. Among them, sim=simU·z is the total similarity between the content attribute characteristics on the target account side and the historical content characteristics on the target account side, realizing the interaction and fusion of historical content sequences and video attributes.

[0188] The above process realizes obtaining the prediction result of the operation event of each content data in the historical content sequence data based on the historical content sequence data, the content attribute data and the total similarity.

[0189] Step 206: Recommend content to the target account based on the content data to be recommended according to the determined prediction result.

[0190] In this step, the prediction results can be sorted, and the content with a prediction result greater than a set prediction threshold can be recommended or screened out for refined sorting.

[0191] In the above embodiment, the calculation of the feature average vector and the acquisition of the first output result Uout can be performed simultaneously, that is, the first historical content feature vector and the sequence feature information are simultaneously input into the first tower model (for example, the user tower), and the first output result Uout with the second dimension M and the feature average vector simU with the first dimension k are obtained from the output result of the first tower model (for example, the user tower).

[0192] Similarly, the acquisition of the content attribute feature vector z and the second output result Iout with the second dimension M can be carried out simultaneously, that is, the content attribute data is input into the second tower model (for example, the item tower), and the content attribute feature vector z with the first dimension k and the second output result Iout with the second dimension M are obtained from the output result of the second tower model (for example, the item tower).

[0193] The content recommendation method of the embodiment of the present application breaks the restriction of the conventional coarse-sorting dual-tower structure on the intersection of item features on the target account side and the content side. Under the condition of satisfying the dual-tower model, the attribute features on the target account side and the historical content flow on the target account side are directly interactive through total similarity, which is conducive to capturing the correlation between historical content and current content attributes, so that the prediction result can be calculated based on the interaction result, that is, the total similarity information, which can achieve the effect of correcting the prediction result, thereby improving the accuracy of the prediction result. The embodiment of the present application provides a new paradigm of the dual-tower model, which makes the dual-tower model have greater potential to explore the relationship between historical interests and content, which is conducive to improving the recommendation effect of coarse sorting.

[0194] See also Figure 3 As shown, Figure 3 FIG. 1 is a flow chart illustrating a method for training a content recommendation model according to an exemplary embodiment. The training method includes:

[0195] Step 301: Obtain sample content sequence data and content attribute sample data of the sample content data of the sample account;

[0196] Step 302: Input the sample content sequence data and the content attribute sample data into the content recommendation model;

[0197] The content recommendation model is used to determine the total similarity between the sample content sequence data and the content attribute sample data, and, based on the sample content sequence data, the content attribute sample data, and the total similarity, obtain a prediction result of an operation event of the sample account on each sample content data in the sample content sequence data;

[0198] Step 303: determining a target loss function value of the content recommendation model based on the prediction result output by the content recommendation model;

[0199] Step 304: Update the model parameters of the content recommendation model according to the target loss function value.

[0200] The above process can be executed repeatedly until the training end condition is met.

[0201] That is, the working principle of the content recommendation model trained by the above training method can be compared with Figure 1 Steps 102 and 103 in the illustrated flow are essentially the same, but the input data differs. Specifically, the historical content sequence data and content attribute data in steps 102 and 103 are replaced by sample content sequence data and sample content attribute data of the sample account, respectively, to represent the inference process of the content recommendation model during training.

[0202] Specifically, the following processing is performed in the content recommendation model.

[0203] Determine the overall similarity between historical content sequence sample data and content attribute sample data:

[0204] Using the first tower model, we can obtain the feature average vector of the historical content sequence sample data.

[0205] For example:

[0206] Projecting each sample content data in the sample content sequence data using the first neural network model to obtain a first historical content feature vector of each sample content data, wherein the first historical content feature vector has a first dimension.

[0207] The first tower model is used to obtain a feature average vector of the first historical content feature vectors of all sample content data in the sample content sequence data.

[0208] Using the second tower model, obtaining the content attribute feature vector of the content attribute sample data;

[0209] For example, the content attribute sample data is projected using the second tower model to obtain a content attribute feature vector with a first dimension.

[0210] Using the feature average vector and content attribute feature vector, the total similarity between the sample content sequence data and the content attribute sample data is determined.

[0211] In one implementation manner, a first dot product result of the feature average vector and the content feature vector is obtained, and the first dot product result is used to represent the total similarity.

[0212] In a second embodiment, the account attribute sample data of the sample account or the account attribute sample data and its embedded feature data are obtained, and the sequence feature information of the sample content sequence data is extracted.

[0213] Input the sequence feature information into the first tower model to obtain a first output result with a second dimension. The first output result is used to characterize the event operation characteristics of the sample account.

[0214] Input the content attribute sample data into the second tower model to obtain a second output result with a second dimension, and the second output result is used to represent the content attribute feature vector.

[0215] Merge the first output result and the feature average vector in vector dimension to obtain the first merged result.

[0216] Merge the second output result and the content attribute feature vector in vector dimension to obtain a second merged result.

[0217] A second dot product result of the first merging result and the second merging result is determined, where the second dot product result includes a dot product component representing the total similarity.

[0218] The prediction process is handled as follows:

[0219] In one embodiment, when the total similarity is one of the embodiments, account attribute sample data or account attribute sample data and its embedded feature data are obtained, and sequence feature information of historical content sequence sample data is extracted.

[0220] Input the sequence feature information into the first tower model to obtain a first output result with a second dimension. The first output result is used to characterize the event operation characteristics of the sample account.

[0221] Input the content attribute sample data into the second tower model to obtain a second output result with a second dimension, and the second output result is used to represent the content attribute feature vector.

[0222] Determine a third dot product result of the first output result and the second output result,

[0223] Normalize the sum of the third dot product result and the first dot product result to get the prediction result.

[0224] In the second implementation mode, when the total similarity is the second implementation mode, the second dot product result is normalized to obtain a prediction result.

[0225] The sequence feature information of the above-mentioned extracted sample content sequence data is specifically:

[0226] One of the implementation methods,

[0227] Projecting the account attribute sample data or the account attribute sample data and its embedded feature sample data using the second neural network model to obtain an account attribute feature vector having a second dimension,

[0228] Projecting each sample content data in the sample content sequence data using the third neural network model to obtain a second historical content feature vector of each sample content data, the second historical content feature vector having a second dimension,

[0229] Determine a first similarity between the account attribute feature vector and the second historical content feature vector of each sample content data,

[0230] According to the first similarity, determine the first attention weight of the account attribute feature vector for each sample content data,

[0231] Projecting each sample content data in the sample content sequence data using the fourth neural network model to obtain a third historical content feature vector of each sample content data, the third historical content feature vector having a second dimension,

[0232] Determine the fourth dot product result of the first attention weight of each sample content data and the third historical content feature vector,

[0233] The fourth dot product results of all content data are accumulated to obtain sequence feature information.

[0234] The second implementation method:

[0235] Projecting each sample content data in the sample content sequence data using the fifth neural network model to obtain a fourth historical content feature vector for each content data,

[0236] Projecting each sample content data in the sample content sequence data using the sixth neural network model to obtain a fifth historical content feature vector for each content data,

[0237] Projecting each sample content data in the sample content sequence data using the seventh neural network model to obtain the sixth historical content feature vector of each content data,

[0238] Determine a second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each sample content data,

[0239] Determine, based on the second similarity, a second attention weight of the fifth historical content feature vector of each sample content data to the fourth historical content feature vector,

[0240] Determine the fifth dot product result of the second attention weight of each sample content data and the sixth historical content feature vector, and the fifth dot product result is used to represent the sequence feature information.

[0241] The fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

[0242] See also Figure 4 As shown, Figure 4 The figure is a schematic diagram of a content recommendation device according to an exemplary embodiment.

[0243] The device includes:

[0244] An acquisition unit 41 is configured to acquire historical content sequence data of a target account and content attribute data of content data to be recommended;

[0245] A determination unit 42 is configured to determine the total similarity between the historical content sequence data and the content attribute data;

[0246] The prediction unit 43 is configured to obtain a prediction result of an operation event of the target account on each content data in the historical content sequence data based on the historical content sequence data, the content attribute data, and the total similarity;

[0247] The output unit 44 is configured to recommend content to the target account based on the content data to be recommended.

[0248] In one example, the determining unit 42 may include:

[0249] a vector mean acquisition module configured to acquire a feature mean vector of the historical content sequence data using the first tower model;

[0250] an attribute feature acquisition module configured to acquire a content attribute feature vector of the content attribute data using the second tower model;

[0251] The total similarity determination module is configured to determine the total similarity between the historical content sequence data and the content attribute data using the feature average vector and the content attribute feature vector.

[0252] In one example, the vector mean acquisition module and the attribute feature acquisition module can respectively implement the calculation of the feature mean vector and the acquisition of the content attribute feature vector through a dual-tower model structure. In this case:

[0253] The vector mean acquisition module may be specifically configured to: project each content data in the historical content sequence data using a first neural network model to obtain a first historical content feature vector of each content data, wherein the first historical content feature vector has a first dimension; and obtain a feature average vector of the first historical content feature vectors of all content data in the historical content sequence data using a first tower model;

[0254] The attribute feature acquisition module may be specifically configured to: project the content attribute data using the second tower model to obtain a content attribute feature vector having a first dimension.

[0255] In one example, as a first method for determining the total similarity, the total similarity determination module may be specifically configured to: obtain a first dot product result of the feature average vector and the content feature vector to obtain the total similarity.

[0256] In another example, as a second method for determining the total similarity, the total similarity determination module may be specifically configured as follows:

[0257] Obtain the target account's account attribute data or account attribute data and its embedded feature data, and extract the sequence feature information of the historical content sequence data.

[0258] The sequence feature information is input into the first tower model to obtain a first output result with a second dimension. The first output result is used to characterize the event operation characteristics of the target account.

[0259] Input the content attribute data into the second tower model to obtain a second output result with a second dimension, and the second output result is used to represent the content attribute feature vector.

[0260] Merge the first output result and the feature average vector in vector dimension to obtain the first merged result.

[0261] Merge the second output result and the content attribute feature vector in vector dimension to obtain a second merged result.

[0262] A second dot product result of the first merging result and the second merging result is determined, where the second dot product result includes a dot product component representing the total similarity.

[0263] In the case of adopting the second determination method of the total similarity, the prediction unit may be specifically configured to: normalize the second dot product result to obtain a prediction result.

[0264] In the case of adopting the first determination method of the total similarity, the prediction unit may be specifically configured as follows:

[0265] Obtain the target account's account attribute data or account attribute data and its embedded feature data, and extract the sequence feature information of the historical content sequence data.

[0266] The sequence feature information is input into the first tower model to obtain a first output result with a second dimension. The first output result is used to characterize the event operation characteristics of the target account.

[0267] Input the content attribute data into the second tower model to obtain a second output result with a second dimension, and the second output result is used to represent the content attribute feature vector.

[0268] Determine a third dot product result of the first output result and the second output result,

[0269] Normalize the sum of the third dot product result and the first dot product result to get the prediction result.

[0270] In one example, to extract sequence feature information, the prediction unit may be specifically configured as follows:

[0271] Projecting the account attribute data or the account attribute data and its embedded feature data using the second neural network model to obtain an account attribute feature vector having a second dimension,

[0272] Projecting each content data in the historical content sequence data using the third neural network model to obtain a second historical content feature vector of each content data, the second historical content feature vector having a second dimension,

[0273] determining a first similarity between the account attribute feature vector and the second historical content feature vector of each content data;

[0274] According to the first similarity, a first attention weight of the account attribute feature vector for each content data is determined,

[0275] Projecting each content data in the historical content sequence data using the fourth neural network model to obtain a third historical content feature vector for each content data, the third historical content feature vector having a second dimension,

[0276] Determine the fourth dot product result of the first attention weight of each content data and the third historical content feature vector,

[0277] The fourth dot product results of all content data are accumulated to obtain sequence feature information.

[0278] In another example, as an alternative solution for extracting sequence feature information, the prediction unit may also be specifically configured as follows:

[0279] Projecting each content data in the historical content sequence data using the fifth neural network model to obtain a fourth historical content feature vector for each content data,

[0280] Projecting each content data in the historical content sequence data using the sixth neural network model to obtain a fifth historical content feature vector for each content data.

[0281] The seventh neural network model is used to project each content data in the historical content sequence data to obtain the sixth historical content feature vector of each content data.

[0282] determining a second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each content data;

[0283] Determine, based on the second similarity, a second attention weight of the fifth historical content feature vector of each content data to the fourth historical content feature vector,

[0284] Determine the fifth dot product result of the second attention weight of each content data and the sixth historical content feature vector, and the fifth dot product result is used to represent the sequence feature information.

[0285] The fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

[0286] See also Figure 5 As shown, Figure 5 FIG. 1 is a schematic diagram of a content recommendation model training device according to an exemplary embodiment. The content recommendation model training device in this embodiment may include:

[0287] An acquiring unit 51 is configured to acquire sample content sequence data and content attribute sample data of the sample content data of a sample account;

[0288] The training unit 52 is configured to:

[0289] The sample content sequence data and the content attribute sample data are input into the content recommendation model. The content recommendation model is used to determine the total similarity between the sample content sequence data and the content attribute sample data, and based on the sample content sequence data, the content attribute sample data and the total similarity, obtain the prediction result of the sample account's operation event for each sample content data in the sample content sequence data.

[0290] According to the prediction results output by the content recommendation model, the target loss function value of the content recommendation model is determined, and according to the target loss function value, the model parameters of the content recommendation model are updated until the training end conditions are met.

[0291] The content recommendation model in this embodiment is configured as follows Figure 4 The determination unit 42 and the prediction unit 43 in the content recommendation apparatus shown have the same operating principle but different input data.

[0292] The content recommendation model used as a training object can be run based on a dual-tower model architecture. If so, the content recommendation model can be configured as follows: using the first tower model to obtain the feature average vector of the historical content sequence sample data, and using the second tower model to obtain the content attribute feature vector of the content attribute sample data; using the feature average vector and the content attribute feature vector, determine the total similarity between the sample content sequence data and the content attribute sample data.

[0293] For example, the content recommendation model is specifically configured as follows: using a first neural network model to project each sample content data in the sample content sequence data to obtain a first historical content feature vector for each sample content data, the first historical content feature vector having a first dimension, and using a first tower model to obtain a feature average vector of the first historical content feature vectors of all sample content data in the sample content sequence data.

[0294] For example, the content recommendation model is specifically configured to project the content attribute sample data using the second tower model to obtain a content attribute feature vector having a first dimension.

[0295] The content recommendation model used as a training object can also support two methods of determining the total similarity.

[0296] When the first method is used to determine the total similarity, the content recommendation model may be specifically configured to obtain a first dot product result of the feature average vector and the content feature vector, and the first dot product result is used to represent the total similarity.

[0297] When the second method is used to determine the total similarity, the content recommendation model can be specifically configured as follows: obtaining account attribute sample data of the sample account or the account attribute sample data and its embedded feature data, and extracting sequence feature information of the sample content sequence data, inputting the sequence feature information into the first tower model, and obtaining a first output result with a second dimension, the first output result is used to characterize the event operation characteristics of the sample account, inputting the content attribute sample data into the second tower model, and obtaining a second output result with a second dimension, the second output result is used to characterize the content attribute feature vector, merging the first output result and the feature average vector in the vector dimension to obtain a first merged result, merging the second output result and the content attribute feature vector in the vector dimension to obtain a second merged result, determining a second dot product result of the first merged result and the second merged result, and the second dot product result includes a dot integral component representing the total similarity.

[0298] If the second method is used to determine the total similarity, the content recommendation model may be specifically configured to obtain a prediction result in the following manner: normalizing the second dot product result to obtain a prediction result.

[0299] If the first method is used to determine the total similarity, the content recommendation model can be specifically configured to obtain a prediction result in the following manner: obtain account attribute sample data or account attribute sample data and its embedded feature data, and extract sequence feature information of historical content sequence sample data, input the sequence feature information into the first tower model, obtain a first output result with a second dimension, and the first output result is used to characterize the event operation characteristics of the sample account, input the content attribute sample data into the second tower model, obtain a second output result with a second dimension, and the second output result is used to characterize the content attribute feature vector, determine the third dot product result of the first output result and the second output result, normalize the sum of the third dot product result and the first dot product result, and obtain a prediction result.

[0300] In one example, in order to achieve the extraction of sequence feature information, the content recommendation model can be specifically configured as follows: using a second neural network model to project account attribute sample data or account attribute sample data and its embedded feature sample data to obtain an account attribute feature vector with a second dimension, using a third neural network model to project each sample content data in the sample content sequence data to obtain a second historical content feature vector for each sample content data, the second historical content feature vector has a second dimension, determining a first similarity between the account attribute feature vector and the second historical content feature vector of each sample content data, and determining a first attention weight of the account attribute feature vector for each sample content data based on the first similarity, using a fourth neural network model to project each sample content data in the sample content sequence data to obtain a third historical content feature vector for each sample content data, the third historical content feature vector has a second dimension, determining a fourth dot product result of the first attention weight of each sample content data and the third historical content feature vector, accumulating the fourth dot product results of all content data to obtain sequence feature information.

[0301] In another example, in order to achieve the extraction of sequence feature information, the content recommendation model can also be specifically configured as follows: using the fifth neural network model to project each sample content data in the sample content sequence data to obtain the fourth historical content feature vector of each content data, using the sixth neural network model to project each sample content data in the sample content sequence data to obtain the fifth historical content feature vector of each content data, using the seventh neural network model to project each sample content data in the sample content sequence data to obtain the sixth historical content feature vector of each content data, determining the second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each sample content data, determining the second attention weight of the fifth historical content feature vector of each sample content data to the fourth historical content feature vector based on the second similarity, determining the fifth dot product result of the second attention weight of each sample content data and the sixth historical content feature vector, and the fifth dot product result is used to characterize the sequence feature information, wherein the fourth historical content feature vector, the fifth historical content feature vector and the sixth historical content feature vector all have a third dimension.

[0302] See also Figure 6 As shown, Figure 6 This is a schematic diagram of an electronic device according to an exemplary embodiment. The electronic device includes a processor; a memory for storing executable instructions of the processor;

[0303] The processor is configured to execute executable instructions to implement the steps of any content recommendation method and / or the steps of the content recommendation model training method.

[0304] In an exemplary embodiment, a computer-readable storage medium including at least one instruction is also provided, such as a memory including at least one instruction. The at least one instruction can be executed by a processor in a computer device to complete the steps of the content recommendation method and / or the steps of the content recommendation model training method in the above embodiment.

[0305] Optionally, the above-mentioned computer-readable storage medium may be a non-temporary computer-readable storage medium. For example, the non-temporary computer-readable storage medium may include ROM (Read-Only Memory), RAM (Random-Access Memory), CD-ROM (Compact Disc Read-Only Memory), magnetic tape, floppy disk and optical data storage device, etc.

[0306] In an exemplary embodiment, a computer program product is also provided, comprising one or more instructions, which can be executed by a processor of a computer device to complete the steps of the content recommendation method and / or the steps of the content recommendation model training method of each of the above-mentioned embodiments.

[0307] Those skilled in the art will readily appreciate other embodiments disclosed herein after considering the specification and practicing the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations disclosed herein that adhere to the general principles disclosed herein and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered merely as exemplary, and the true scope and spirit of this disclosure are indicated by the following claims.

[0308] It should be understood that the present disclosure is not limited to the precise structures 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 content recommendation method, characterized in that: include: Obtain the target account's historical content sequence data and content attribute data of the content data to be recommended, Obtaining the feature average vector of the historical content sequence data, Obtaining a content attribute feature vector having a first dimension of the content attribute data, Acquiring account attribute data of the target account or the account attribute data and its embedded feature data, and extracting sequence feature information of the historical content sequence data; Inputting the sequence feature information into a first tower model for processing feature data on the target account side to obtain a first output result, wherein the first output result is used to characterize the event operation feature of the target account, The content attribute data is input into a second tower model for content-side feature data processing to obtain a second output result, wherein the second output result is used to represent a content attribute feature vector having a second dimension. Merging the first output result and the feature average vector in a vector dimension to obtain a first merged result, merging the second output result and the content attribute feature vector having the first dimension in a vector dimension to obtain a second merged result, and determining a second dot product result of the first merged result and the second merged result, wherein the second dot product result includes a first dot product result component for representing the total similarity between the historical content sequence data and the content attribute data, and a third dot product result component between the first output result and the second output result; or determining a third dot product result of the first output result and the second output result, obtaining a first dot product result of the feature average vector and the content feature vector having the first dimension, and calculating the sum of the third dot product result and the first dot product result to obtain a second dot product result; Based on the second dot product result, obtaining a prediction result of the target account's operation event on each content data in the historical content sequence data, wherein the prediction result represents an estimated value of the operation event; According to the prediction result, content recommendation based on the to-be-recommended content data is performed for the target account.

2. The content recommendation method according to claim 1, characterized in that The obtaining of the feature average vector of the historical content sequence data includes: Using the first tower model, obtaining a feature average vector of the historical content sequence data; The step of obtaining a content attribute feature vector having a first dimension of the content attribute data includes: The second tower model is used to obtain a content attribute feature vector having a first dimension of the content attribute data.

3. The content recommendation method according to claim 2, characterized in that: The obtaining of the feature average vector of the historical content sequence data by using the first tower model includes: Projecting each of the content data in the historical content sequence data using a first neural network model to obtain a first historical content feature vector of each of the content data, wherein the first historical content feature vector has a first dimension, The feature average vector is obtained by using the first tower model, wherein the feature average vector represents: an average value of the first historical content feature vectors of all the content data in the historical content sequence data.

4. The content recommendation method according to claim 2, characterized in that: The method of using the second tower model to obtain a content attribute feature vector having a first dimension of the content attribute data includes: The content attribute data is projected using a second tower model to obtain a content attribute feature vector having a first dimension.

5. The content recommendation method according to claim 1, characterized in that: The first output result has a second dimension.

6. The content recommendation method according to claim 5, characterized in that: The obtaining, based on the second dot product result, a prediction result of the target account's operation event on each content data in the historical content sequence data includes: Normalize the second dot product result to obtain the prediction result.

7. The method according to claim 1, characterized in that The extracting sequence feature information of the historical content sequence data includes: Projecting the account attribute data or the account attribute data and its embedded feature data using a second neural network model to obtain an account attribute feature vector having a second dimension, Projecting each content data in the historical content sequence data using a third neural network model to obtain a second historical content feature vector of each content data, wherein the second historical content feature vector has the second dimension, determining a first similarity between the account attribute feature vector and the second historical content feature vector of each content data; Determine a first attention weight of the account attribute feature vector for each of the content data based on the first similarity, Projecting each of the content data in the historical content sequence data using a fourth neural network model to obtain a third historical content feature vector for each of the content data, wherein the third historical content feature vector has the second dimension, Determine a fourth dot product result of the first attention weight of each content data and the third historical content feature vector, The fourth dot product results of all the content data are accumulated to obtain the sequence feature information.

8. The method according to claim 1, characterized in that The extracting sequence feature information of the historical content sequence data includes: Projecting each of the content data in the historical content sequence data using a fifth neural network model to obtain a fourth historical content feature vector for each of the content data, Projecting each of the content data in the historical content sequence data using a sixth neural network model to obtain a fifth historical content feature vector for each of the content data, Projecting each of the content data in the historical content sequence data using the seventh neural network model to obtain a sixth historical content feature vector for each of the content data, determining a second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each content data; determining a second attention weight of the fifth historical content feature vector of each content data to the fourth historical content feature vector according to the second similarity, determining a fifth dot product result of the second attention weight of each content data and the sixth historical content feature vector, wherein the fifth dot product result is used to represent the sequence feature information; The fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

9. A method for training a content recommendation model, characterized in that: include: Get the sample content sequence data and content attribute sample data of the sample content data of the sample account, Get the feature average vector of historical content sequence sample data, Obtaining a content attribute feature vector with a first dimension of the content attribute sample data, Acquire account attribute sample data of the sample account or the account attribute sample data and its embedded feature data, and extract sequence feature information of the sample content sequence data, The sequence feature information is input into a first tower model for processing feature data on the target account side to obtain a first output result, which is used to characterize the event operation characteristics of the sample account. Inputting the content attribute sample data into a second tower model for content-side feature data processing to obtain a second output result, wherein the second output result is used to represent a content attribute feature vector having a second dimension; Merging the first output result and the feature average vector in a vector dimension to obtain a first merged result, merging the second output result and the content attribute feature vector in a vector dimension to obtain a second merged result, and determining a second dot product result of the first merged result and the second merged result, wherein the second dot product result includes a first dot product component representing the total similarity between the sample content sequence data and the content attribute sample data, and a third dot product result component of the first output result and the second output result; Alternatively, determining a third dot product result of the first output result and the second output result, obtaining a first dot product result of the feature average vector and the content feature vector having the first dimension, and calculating a sum of the third dot product result and the first dot product result to obtain a second dot product result; Based on the second dot product result, obtaining a prediction result of the operation event of the sample account on each sample content data in the sample content sequence data; Determine the target loss function value of the model according to the prediction result output by the model, According to the target loss function value, the model parameters of the model are updated until the training end condition is met.

10. The training method according to claim 9, characterized in that: The step of obtaining the feature average vector of the historical content sequence sample data includes: Using the first tower model, the feature average vector of the historical content sequence sample data is obtained; The step of obtaining a content attribute feature vector having a first dimension of the content attribute sample data includes: The second tower model is used to obtain the content attribute feature vector of the content attribute sample data.

11. The training method according to claim 10, characterized in that: The obtaining of the feature average vector of the sample content sequence data by using the first tower model includes: Projecting each sample content data in the sample content sequence data using a first neural network model to obtain a first historical content feature vector of each sample content data, wherein the first historical content feature vector has a first dimension. The feature average vector of the first historical content feature vectors of all the sample content data in the sample content sequence data is obtained by using the first tower model.

12. The training method according to claim 10, characterized in that: The method of obtaining a content attribute feature vector having a first dimension of the content attribute sample data by using the second tower model includes: The content attribute sample data is projected using a second tower model to obtain a content attribute feature vector having a first dimension.

13. The training method according to claim 9, characterized in that: The first output result has a second dimension.

14. The training method according to claim 13, characterized in that: The obtaining, based on the second dot product result, a prediction result of the operation event of the sample account on each sample content data in the sample content sequence data includes: Normalize the second dot product result to obtain the prediction result.

15. The training method according to claim 9, characterized in that: The extracting sequence feature information of the sample content sequence data includes: Projecting the account attribute sample data or the account attribute sample data and its embedded feature sample data using a second neural network model to obtain an account attribute feature vector having a second dimension, Projecting each sample content data in the sample content sequence data using a third neural network model to obtain a second historical content feature vector of each sample content data, wherein the second historical content feature vector has the second dimension. determining a first similarity between the account attribute feature vector and the second historical content feature vector of each of the sample content data; Determine, based on the first similarity, a first attention weight of the account attribute feature vector for each of the sample content data; Projecting each of the sample content data in the sample content sequence data using a fourth neural network model to obtain a third historical content feature vector for each of the sample content data, wherein the third historical content feature vector has the second dimension, Determine a fourth dot product result of the first attention weight of each sample content data and the third historical content feature vector, The fourth dot product results of all the content data are accumulated to obtain the sequence feature information.

16. The training method according to claim 9, characterized in that: The extracting sequence feature information of the sample content sequence data includes: Projecting each of the sample content data in the sample content sequence data using a fifth neural network model to obtain a fourth historical content feature vector for each of the content data, Projecting each of the sample content data in the sample content sequence data using a sixth neural network model to obtain a fifth historical content feature vector for each of the content data, Projecting each of the sample content data in the sample content sequence data using the seventh neural network model to obtain a sixth historical content feature vector for each of the content data, determining a second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each sample content data; determining a second attention weight of the fifth historical content feature vector of each sample content data to the fourth historical content feature vector according to the second similarity, determining a fifth dot product result of the second attention weight of each sample content data and the sixth historical content feature vector, wherein the fifth dot product result is used to represent the sequence feature information; The fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

17. A content recommendation device, characterized in that: The device includes: An acquisition unit is configured to acquire the historical content sequence data of the target account and the content attribute data of the content data to be recommended, The determining unit is configured to obtain a feature average vector of the historical content sequence data, obtain a content attribute feature vector having a first dimension of the content attribute data, The prediction unit is configured to obtain the account attribute data of the target account or the account attribute data and its embedded feature data, extract sequence feature information of the historical content sequence data, input the sequence feature information into a first tower model for target account side feature data processing, and obtain a first output result, wherein the first output result is used to characterize the event operation feature of the target account. Inputting the content attribute data into a second tower model for content-side feature data processing to obtain a second output result, wherein the second output result is used to represent a content attribute feature vector having a second dimension; Merging the first output result and the feature average vector in a vector dimension to obtain a first merged result, merging the second output result and the content attribute feature vector having the first dimension in a vector dimension to obtain a second merged result, and determining a second dot product result of the first merged result and the second merged result, wherein the second dot product result includes a first dot product result component for representing the total similarity between the historical content sequence data and the content attribute data, and a third dot product result component of the first output result and the second output result; Alternatively, determining a third dot product result of the first output result and the second output result, obtaining a first dot product result of the feature average vector and the content feature vector having the first dimension, and calculating a sum of the third dot product result and the first dot product result to obtain a second dot product result; Based on the second dot product result, a prediction result of the target account's operation event on each content data in the historical content sequence data is obtained, wherein the prediction result represents an estimated value of the operation event. The output unit is configured to perform content recommendation based on the content data to be recommended to the target account.

18. The content recommendation device according to claim 17, wherein: The determining unit includes: The vector mean acquisition module is configured to obtain the feature average vector of the historical content sequence data using the first tower model, The attribute feature acquisition module is configured to acquire a content attribute feature vector of the content attribute data using the second tower model.

19. The content recommendation device according to claim 18, characterized in that The vector mean acquisition module is specifically configured as follows: Projecting each of the content data in the historical content sequence data using a first neural network model to obtain a first historical content feature vector of each of the content data, wherein the first historical content feature vector has a first dimension, The feature average vector is obtained by using the first tower model, and the feature average vector represents: an average value of the first historical content feature vectors of all the content data in the historical content sequence data.

20. The content recommendation device according to claim 18, wherein: The attribute feature acquisition module is specifically configured as follows: The content attribute data is projected using a second tower model to obtain a content attribute feature vector having a first dimension.

21. The content recommendation device according to claim 18, wherein: The first output result has a second dimension.

22. The content recommendation device according to claim 17, wherein: The prediction unit is specifically configured as follows: Normalize the second dot product result to obtain the prediction result.

23. The content recommendation device according to claim 17, wherein: The sequence feature information is obtained in the following manner: Projecting the account attribute data or the account attribute data and its embedded feature data using a second neural network model to obtain an account attribute feature vector having a second dimension, Projecting each content data in the historical content sequence data using a third neural network model to obtain a second historical content feature vector of each content data, wherein the second historical content feature vector has the second dimension, determining a first similarity between the account attribute feature vector and the second historical content feature vector of each content data; Determine a first attention weight of the account attribute feature vector for each of the content data based on the first similarity, Projecting each of the content data in the historical content sequence data using a fourth neural network model to obtain a third historical content feature vector for each of the content data, wherein the third historical content feature vector has the second dimension, Determine a fourth dot product result of the first attention weight of each content data and the third historical content feature vector, The fourth dot product results of all the content data are accumulated to obtain the sequence feature information.

24. The content recommendation device according to claim 17, wherein: The sequence feature information is obtained in the following manner: Projecting each of the content data in the historical content sequence data using a fifth neural network model to obtain a fourth historical content feature vector for each of the content data, Projecting each of the content data in the historical content sequence data using a sixth neural network model to obtain a fifth historical content feature vector for each of the content data, Projecting each of the content data in the historical content sequence data using the seventh neural network model to obtain a sixth historical content feature vector for each of the content data, determining a second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each content data; determining a second attention weight of the fifth historical content feature vector of each content data to the fourth historical content feature vector according to the second similarity, determining a fifth dot product result of the second attention weight of each content data and the sixth historical content feature vector, wherein the fifth dot product result is used to represent the sequence feature information; The fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

25. A training device for a content recommendation model, characterized in that: include: an acquisition unit configured to acquire sample content sequence data and content attribute sample data of the sample content data of the sample account, The training unit is configured as follows: Get the feature average vector of historical content sequence sample data, Obtaining a content attribute feature vector with a first dimension of the content attribute sample data, Acquire account attribute sample data of the sample account or the account attribute sample data and its embedded feature data, and extract sequence feature information of the sample content sequence data, The sequence feature information is input into a first tower model for processing feature data on the target account side to obtain a first output result, which is used to characterize the event operation characteristics of the sample account. Inputting the content attribute sample data into a second tower model for content-side feature data processing to obtain a second output result, wherein the second output result is used to represent a content attribute feature vector having a second dimension; Merging the first output result and the feature average vector in a vector dimension to obtain a first merged result, merging the second output result and the content attribute feature vector in a vector dimension to obtain a second merged result, and determining a second dot product result of the first merged result and the second merged result, wherein the second dot product result includes a first dot product component representing the total similarity between the sample content sequence data and the content attribute sample data, and a third dot product result component of the first output result and the second output result; Alternatively, determining a third dot product result of the first output result and the second output result, obtaining a first dot product result of the feature average vector and the content feature vector having the first dimension, and calculating a sum of the third dot product result and the first dot product result to obtain a second dot product result; Based on the second dot product result, a prediction result of the sample account on the operation event of each sample content data in the sample content sequence data is obtained, wherein the prediction result represents an estimated value of the operation event. According to the prediction results, the target loss function value is determined. According to the target loss function value, the model parameters of the model are updated until the training end condition is met.

26. The training device according to claim 25, characterized in that The training unit is configured to: Using the first tower model, the feature average vector of the historical content sequence sample data is obtained, The second tower model is used to obtain the content attribute feature vector of the content attribute sample data.

27. The training device according to claim 26, characterized in that The training unit is specifically configured as follows: Projecting each sample content data in the sample content sequence data using a first neural network model to obtain a first historical content feature vector of each sample content data, wherein the first historical content feature vector has a first dimension. The feature average vector is obtained by using the first tower model, wherein the feature average vector represents: an average value of the first historical content feature vectors of all the sample content data in the sample content sequence data.

28. The training device according to claim 26, characterized in that The training unit is specifically configured as follows: The content attribute sample data is projected using a second tower model to obtain a content attribute feature vector having a first dimension.

29. The training device according to claim 25, characterized in that The first output result has a second dimension.

30. The training device according to claim 25, characterized in that The training unit is specifically configured as follows: Normalize the second dot product result to obtain the prediction result.

31. The training device according to claim 25, characterized in that The training unit is specifically configured as follows: Projecting the account attribute sample data or the account attribute sample data and its embedded feature sample data using a second neural network model to obtain an account attribute feature vector having a second dimension, Projecting each sample content data in the sample content sequence data using a third neural network model to obtain a second historical content feature vector of each sample content data, wherein the second historical content feature vector has the second dimension. determining a first similarity between the account attribute feature vector and the second historical content feature vector of each of the sample content data; Determine, based on the first similarity, a first attention weight of the account attribute feature vector for each of the sample content data; Projecting each of the sample content data in the sample content sequence data using a fourth neural network model to obtain a third historical content feature vector for each of the sample content data, wherein the third historical content feature vector has the second dimension, Determine a fourth dot product result of the first attention weight of each sample content data and the third historical content feature vector, The fourth dot product results of all the content data are accumulated to obtain the sequence feature information.

32. The training device according to claim 25, characterized in that The training unit is specifically configured as follows: Projecting each of the sample content data in the sample content sequence data using a fifth neural network model to obtain a fourth historical content feature vector for each of the content data, Projecting each of the sample content data in the sample content sequence data using a sixth neural network model to obtain a fifth historical content feature vector for each of the content data, Projecting each of the sample content data in the sample content sequence data using the seventh neural network model to obtain a sixth historical content feature vector for each of the content data, determining a second similarity between the fourth historical content feature vector and the fifth historical content feature vector of each sample content data; determining a second attention weight of the fifth historical content feature vector of each sample content data to the fourth historical content feature vector according to the second similarity, determining a fifth dot product result of the second attention weight of each sample content data and the sixth historical content feature vector, wherein the fifth dot product result is used to represent the sequence feature information; The fourth historical content feature vector, the fifth historical content feature vector, and the sixth historical content feature vector all have a third dimension.

33. An electronic device, characterized in that: include: processor; a memory for storing executable instructions for the processor; The processor is configured to execute the executable instructions to implement the content recommendation method according to any one of claims 1 to 8 and / or the content recommendation model training method according to any one of claims 9 to 17.

34. A computer-readable storage medium, characterized in that When at least one instruction in the computer-readable storage medium is executed by a processor of an electronic device, the electronic device is enabled to implement the content recommendation method as described in any one of claims 1 to 8 and / or the training method of the content recommendation model as described in any one of claims 9 to 17.

35. A computer program product, comprising a computer program, which, when executed by a processor, implements the content recommendation method according to any one of claims 1 to 8 and / or the content recommendation model training method according to any one of claims 9 to 17.

Images