Resource pushing method and device, electronic equipment, storage medium and product

Abstract

The invention provides a resource pushing method and device, electronic equipment, a storage medium and a product, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring a first historical resource, a second historical resource, text content of a candidate resource and generated duration of the candidate resource; extracting the text content features of the candidate resources to obtain first text features; obtaining a to-be-utilized access rate based on the first text feature and the generated duration; extracting a text content feature of the first historical resource to obtain a second text feature; extracting the text content feature of the second historical resource to obtain a third text feature; extracting splicing result features of the text content of the candidate resources and the to-be-utilized access rate to obtain splicing features; based on the second text feature, the third text feature and the splicing feature, calculating the degree of interest; and if the interest degree exceeds a threshold value, sending the resource content of the candidate resource to the target account. According to the invention, the experience feeling of viewing the pushed resources by the user can be improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a resource pushing method, device, electronic device, storage medium and product. Background Art

[0002] The user can view the resources pushed by the server on the client. The resources may be news, emails, or public account articles, etc. In the prior art, the server pushes the resources that the user is interested in to the account logged into the client.

[0003] However, the existing technology only considers whether the user is interested in the resource, but does not consider the timeliness of the resource. If the pushed resource meets the user's interest, but is relatively old from the current moment (for example, the server pushes last year's news to the account logged in to the client on the same day), the browsing value of the pushed resource is not high for the user, resulting in a low user experience in viewing the pushed resource. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a resource push method, device, electronic device, storage medium and product to improve the user's experience of viewing pushed resources. The specific technical solution is as follows:

[0005] In a first aspect, an embodiment of the present invention provides a resource push method, the method comprising:

[0006] Obtain the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generation time of each candidate resource at the current moment;

[0007] Extracting features from the text content of each candidate resource to obtain a first text feature;

[0008] Based on the first text feature and the generated time of each candidate resource, predict the access rate of pushing the candidate resource at the current moment to obtain the access rate to be utilized;

[0009] Extracting features from the text content of the first historical resource to obtain a second text feature; extracting features from the text content of the second historical resource to obtain a third text feature; and extracting features from a concatenation result of the text content of each candidate resource and the access rate to be utilized to obtain a concatenation feature;

[0010] Calculating the interest level of the target account in the candidate resource based on the second text feature, the third text feature, and the concatenated feature of each candidate resource;

[0011] If the interest level of the candidate resource exceeds a first preset threshold, the resource content of the candidate resource is sent to the target account.

[0012] Optionally, the extracting features from the text content of each candidate resource to obtain a first text feature includes:

[0013] Input the text content of each candidate resource into the first transformer layer of the pre-trained access prediction model to obtain the first text feature; wherein the access prediction model is trained based on the text content of the pushed sample resources, the generated time of the sample resources at the push time, and the real access rate of the sample resources; the generated time of a sample resource at the push time represents the difference between the push time of the sample resource and the generation time of the sample resource; the real access rate of a sample resource is the access rate of the sample resource when it is pushed at the push time of the sample resource; the push time of every two sample resources adjacent to each other is separated by the same preset time;

[0014] The method predicts the access rate of pushing the candidate resource at the current moment based on the first text feature and the generated time of each candidate resource to obtain the access rate to be utilized, including:

[0015] Inputting the concatenated first text feature and the generated duration into the first fully connected layer in the access prediction model to obtain a first dimensionality reduction feature;

[0016] The first dimension reduction feature is input into the first output layer of the access prediction model to obtain the access rate of pushing the candidate resource at the current moment as the access rate to be utilized.

[0017] Optionally, the extracting features from the text content of the first historical resource to obtain the second text feature includes:

[0018] Inputting the text content of the first historical resource into the second transformer layer of the pre-trained interest evaluation model to obtain a second text feature; wherein the interest evaluation model is trained based on the text content of the target sample resource, the actual access rate of the target sample resource, the text content of the resource pushed to the sample account, the text content of the resource browsed by the sample account, and the actual interest level of the sample account in the target sample resource in a preset time period; the target sample resource is the resource pushed to the sample account in the preset time period;

[0019] The extracting features of the text content of the second historical resource to obtain a third text feature includes:

[0020] Inputting the text content of the second historical resource into the second fully connected layer in the interest assessment model to obtain a third text feature;

[0021] The feature extraction of the splicing result of the text content and the access rate to be utilized of each candidate resource to obtain the splicing feature includes:

[0022] Inputting the concatenation result of the text content and the access rate to be utilized of each candidate resource into the third fully connected layer in the interest evaluation model to obtain a concatenated feature;

[0023] The calculating the interest degree of the target account in the candidate resource based on the second text feature, the third text feature, and the concatenated feature of each candidate resource includes:

[0024] Inputting the second text feature, the third text feature, and the concatenated feature of each candidate resource into the fourth fully connected layer in the interest evaluation model to obtain a second dimensionality reduction feature;

[0025] The second dimensionality reduction feature is input into the second output layer of the interest evaluation model to obtain the degree of interest of the target account in the candidate resource.

[0026] Optionally, the actual interest level of a sample account in the target sample resource in a preset time period is obtained according to the following steps;

[0027] For a sample account, obtain the total number of target sample resources pushed to the sample account within a preset time period, the exposure duration of each target sample resource, and the generated duration of each target sample resource at the time of pushing;

[0028] Using a preset formula, the total number of acquired resources, the generated time of each target sample resource at the push time, and the exposure time are calculated and processed to obtain the actual interest level of the sample account in the target sample resource in the preset time period;

[0029] Wherein, the preset formula is: ; Indicates the actual interest of a sample account in the target sample resource during a preset time period. Indicates the total number of target sample resources pushed to the sample account within the preset time period. represents the preset parameters, and Greater than 0 and less than 1, Indicates Whether the target sample resources are accessed, Indicates The exposure duration of target sample resources, Indicates The generated duration of each target sample resource at the time of push.

[0030] Optionally, before acquiring the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generation time of each candidate resource at the current moment, the method further includes:

[0031] Obtain the text content of each resource to be processed in the preset resource library and the account information of the target account;

[0032] Extracting features from the text content of each resource to be processed to obtain a fourth text feature; and extracting features from the account information of the target account to obtain an account feature;

[0033] Calculating the matching degree between the resource to be processed and the target account based on the concatenation result of the fourth text feature and the account feature;

[0034] If the matching degree of the resource to be processed is greater than a second preset threshold, the resource to be processed is determined as a candidate resource of the target account.

[0035] Optionally, feature extraction is performed on the text content of each resource to be processed to obtain a fourth text feature, including:

[0036] Inputting the text content of each resource to be processed into the third transformer layer of the pre-trained resource matching model to obtain a fourth text feature; wherein the resource matching model is trained based on the text content of the sample resource, the account information of the sample account, and a label indicating whether the sample account has accessed the sample resource;

[0037] The feature extraction of the target account information to obtain the account features includes:

[0038] Inputting the account information of the target account into the fifth fully connected layer in the resource matching model to obtain account features;

[0039] The calculating the matching degree between the resource to be processed and the target account based on the concatenation result of the fourth text feature and the account feature includes:

[0040] Inputting the concatenation result of the fourth text feature and the account feature into the sixth fully connected layer in the resource matching model to obtain a third dimensionality reduction feature;

[0041] The third dimensionality reduction feature is input into the third output layer in the resource matching model to obtain the matching degree between the resource to be processed and the target account.

[0042] In a second aspect, an embodiment of the present invention provides a resource push device, the device comprising:

[0043] A first acquisition module is used to acquire the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generation time of each candidate resource at the current moment;

[0044] A first extraction module, used to extract features from the text content of each candidate resource to obtain a first text feature;

[0045] A prediction module, configured to predict the access rate of pushing the candidate resource at the current moment based on the first text feature and the generated time of each candidate resource, and obtain the access rate to be utilized;

[0046] A second extraction module is used to extract features from the text content of the first historical resource to obtain a second text feature; extract features from the text content of the second historical resource to obtain a third text feature; and extract features from a splicing result of the text content of each candidate resource and the access rate to be utilized to obtain a splicing feature;

[0047] A calculation module, configured to calculate the interest level of the target account in the candidate resource based on the second text feature, the third text feature, and a concatenated feature of each candidate resource;

[0048] The sending module is used to send the resource content of the candidate resource to the target account if the interest level of the candidate resource exceeds a first preset threshold.

[0049] An embodiment of the present invention further provides an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0050] Memory, used to store computer programs;

[0051] The processor is used to implement the above resource pushing method when executing the program stored in the memory.

[0052] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the resource pushing method described above is implemented.

[0053] An embodiment of the present invention further provides a computer program product, including a computer program, and the computer program implements the above-mentioned resource pushing method when executed by a processor.

[0054] Beneficial effects of the embodiments of the present invention:

[0055] A resource push method provided by an embodiment of the present invention can obtain the text content of a first historical resource that has been pushed to a target account, the text content of a second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generated time of each candidate resource at the current moment; perform feature extraction on the text content of each candidate resource to obtain a first text feature; based on the first text feature and the generated time of each candidate resource, predict the access rate of pushing the candidate resource at the current moment to obtain a to-be-used access rate; perform feature extraction on the text content of the first historical resource to obtain a second text feature; perform feature extraction on the text content of the second historical resource to obtain a third text feature; and perform feature extraction on the splicing result of the text content of each candidate resource and the to-be-used access rate to obtain a splicing feature; based on the second text feature, the third text feature, and the splicing feature of each candidate resource, calculate the target account's interest in the candidate resource; if the interest in the candidate resource exceeds a first preset threshold, send the resource content of the candidate resource to the target account.

[0056] In the embodiment of the present invention, in the process of predicting the access rate of pushing the candidate resource at the current moment, the generated time length and the first text feature of the candidate resource at the current moment are used. The generated time length can reflect the timeliness. The longer the generated time length, the lower the timeliness. The first text feature can include the text feature of the candidate resource itself. Thus, the determined access rate to be utilized takes into account the generated time length of the candidate resource at the current moment and the influence of the text content of the candidate resource. In the process of calculating the interest of the target account in the candidate resource, the text features of the first historical resource and the second historical resource and the access rate to be utilized are used. The first historical resource is the pushed resource, and the second historical resource is the browsed resource. The first historical resource and the second historical resource can include the text features of the historical resource that the target account is interested in. Thus, the interest of the target account in the candidate resource is calculated by combining the text features of the historical resource that the target account is interested in, the timeliness of the candidate resource and the text features of the candidate resource itself. Therefore, the candidate resource whose interest level exceeds the first preset threshold meets the interest of the target account and has timeliness, which can improve the browsing value of the pushed resource and improve the user's experience of viewing the pushed resource.

[0057] Of course, it is not necessary to achieve all of the advantages described above at the same time to implement any product or method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can also be obtained based on these drawings.

[0059] Figure 1 A schematic diagram of a flow chart of a first resource push method provided in an embodiment of the present invention;

[0060] Figure 2 A schematic diagram of a flow chart of a second resource push method provided in an embodiment of the present invention;

[0061] Figure 3 A schematic diagram of a flow chart of a third resource push method provided in an embodiment of the present invention;

[0062] Figure 4 A schematic diagram of a fourth resource push method provided in an embodiment of the present invention;

[0063] Figure 5 A schematic diagram of a fifth resource push method provided in an embodiment of the present invention;

[0064] Figure 6 A schematic diagram of the principle of a resource push method provided by an embodiment of the present invention;

[0065] Figure 7 A schematic diagram of the structure of a resource push device provided by an embodiment of the present invention;

[0066] Figure 8 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field based on the present invention belong to the scope of protection of the present invention.

[0068] In the prior art, the server can push resources to the account of the logged-in client. The pushed resources can be determined according to a fixed algorithm or manually selected. Specifically, professionals can determine the resources to be pushed, the push title, and the account to be pushed according to the content of the resources, or the server can determine the resources to be pushed, the push title, and the account to be pushed according to the user's interests according to a pre-set algorithm. The server sends the manually determined resources and push title to the account to be pushed.

[0069] However, the existing technology only considers whether the user is interested in the resource, but does not consider the timeliness of the resource. If the pushed resource meets the user's interest, but is far away from the current moment (for example, the server pushes last year's news to the client on the same day), the browsing value of the pushed resource is not high for the user, resulting in a low user experience in viewing the pushed resource.

[0070] In order to improve the user's experience of viewing pushed resources, embodiments of the present invention provide a resource pushing method, device, electronic device, storage medium and product.

[0071] The resource push method provided in the embodiment of the present invention can be applied to the server. The server can obtain candidate resources and determine the resources to be pushed to the account from the candidate resources; the server can also interact with the client to push resources to the account logged in to the client.

[0072] A resource push method provided by an embodiment of the present invention may include the following steps:

[0073] Obtain the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generation time of each candidate resource at the current moment;

[0074] Extracting features from the text content of each candidate resource to obtain a first text feature;

[0075] Based on the first text feature and the generated time of each candidate resource, predict the access rate of pushing the candidate resource at the current moment to obtain the access rate to be utilized;

[0076] Performing feature extraction on the text content of the first historical resource to obtain a second text feature; performing feature extraction on the text content of the second historical resource to obtain a third text feature; and performing feature extraction on the splicing result of the text content of each candidate resource and the access rate to be utilized to obtain a splicing feature;

[0077] Calculate the interest degree of the target account in the candidate resource based on the second text feature, the third text feature, and the concatenated feature of each candidate resource;

[0078] If the interest level of the candidate resource exceeds a first preset threshold, the resource content of the candidate resource is sent to the target account.

[0079] In the embodiment of the present invention, in the process of predicting the access rate of pushing the candidate resource at the current moment, the generated time length and the first text feature of the candidate resource at the current moment are used. The generated time length can reflect the timeliness. The longer the generated time length, the lower the timeliness. The first text feature can include the text feature of the candidate resource itself. Thus, the determined access rate to be utilized takes into account the generated time length of the candidate resource at the current moment and the influence of the text content of the candidate resource. In the process of calculating the interest of the target account in the candidate resource, the text features of the first historical resource and the second historical resource and the access rate to be utilized are used. The first historical resource is the pushed resource, and the second historical resource is the browsed resource. The first historical resource and the second historical resource can include the text features of the historical resource that the target account is interested in. Thus, the interest of the target account in the candidate resource is calculated by combining the text features of the historical resource that the target account is interested in, the timeliness of the candidate resource and the text features of the candidate resource itself. Therefore, the candidate resource whose interest level exceeds the first preset threshold meets the interest of the target account and has timeliness, which can improve the browsing value of the pushed resource and improve the user's experience of viewing the pushed resource.

[0080] The following describes a resource push method provided by an embodiment of the present invention in conjunction with the accompanying drawings. Figure 1 , Figure 1 A schematic flow chart of a resource push method provided in an embodiment of the present invention, the method may include steps S101-S106.

[0081] S101, obtaining the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generation time of each candidate resource at the current moment.

[0082] It is understandable that the server can obtain the text content of each candidate resource from the preset resource library. The preset resource library can be a local database of the server or a database in other devices. The preset resource library stores the text content of the candidate resources to be pushed to the target account. Among them, the preset resource library can also store the generation time of each candidate resource. The server can obtain the generation time of each candidate resource, and then use the difference between the current time and the generation time of each candidate resource as the generated time of the candidate resource at the current time. The generated time of the candidate resource can reflect the timeliness of the candidate resource. The generated time of the candidate resource at the current time is inversely correlated with the timeliness. The longer the generated time, the lower the timeliness.

[0083] The server may obtain the text content of the first historical resource and the text content in the second historical resource from a database storing the text content of the resources pushed by the target account and the text content of the resources browsed by the target account.

[0084] The text content of each resource may include at least one of the following: resource category, resource author, resource keyword, resource tag, and resource title. Each resource may be news, email, or public account article.

[0085] S102: extract features from the text content of each candidate resource to obtain a first text feature.

[0086] It is understandable that the server can input the text content of each candidate resource into the feature extraction network in the trained deep learning model to extract features and obtain the first text feature. Exemplarily, the feature extraction network can be a transformer network layer or a fully connected network layer.

[0087] S103: Based on the first text feature and the generated time of each candidate resource, predict the access rate of pushing the candidate resource at the current moment to obtain the access rate to be utilized.

[0088] It is understandable that the first text feature of the candidate resource may include the feature of the candidate resource itself, and the time the candidate resource has been generated may reflect the timeliness of the candidate resource. The server may predict the access rate of the candidate resource being pushed at the current moment based on the timeliness feature and text feature of the candidate resource.

[0089] Among them, the access rate of a resource is the ratio of the number of times the resource is browsed after being pushed to the number of times it is exposed; resource exposure means that the resource is displayed to the user on the client for user selection, and resource browsing means that the resource is clicked to access. For example, the client homepage displays multiple news headlines, and the news to which each displayed news headline belongs is the exposed resource. The user clicks on a news headline on the client homepage, and the client accesses the specific content of the news to which the news headline belongs. The accessed news is the browsed resource.

[0090] Optionally, in one implementation, the server can determine resources similar to the text features of the candidate resources. The determined resources can be resources that have been pushed to each account. The resources whose generated time at the push time is the same as the generated time of the candidate resources are determined as the target resources; the access rate of the target resources is used as a reference to predict the access rate of the candidate resources. Exemplarily, if there are multiple target resources to be found, the average of the access rates of multiple target resources is used as the predicted access rate; if there is one target resource to be found, the access rate in the target resource is used as the predicted access rate. For example, the candidate resource is news, and the text features of the candidate resource indicate that the news belongs to the financial type. The generated time of the candidate resource at the current moment is one hour, the target resources a and b are both financial news, and the difference between the push time and the generation time of the target resources a and b is one hour. The access rate of the target resource a is 20%, and the access rate of the target resource b is 40%. Then 30% can be used as the access rate of the candidate resource pushed at the current moment. In the embodiment of the present invention, the access rate of the target resource similar to the candidate resource can be used as a reference, and the access rate of the candidate resource can be accurately predicted in the case of a reference.

[0091] Optionally, in another implementation, the server can use a pre-trained click prediction model to process the first text feature and generated duration of each candidate resource to obtain a predicted access rate of the candidate resource at the current moment. This implementation will be described in detail in subsequent embodiments.

[0092] S104, extracting features from the text content of the first historical resource to obtain a second text feature; extracting features from the text content of the second historical resource to obtain a third text feature; and extracting features from the concatenation result of the text content of each candidate resource and the access rate to be utilized to obtain a concatenation feature.

[0093] It is understandable that the server can input the text content of the first historical resource into the feature extraction network in the trained deep learning model for feature extraction to obtain the second text feature; input the text content of the second historical resource into the feature extraction network in the trained deep learning model for feature extraction to obtain the third text feature; input the concatenation result of the text content of each candidate resource and the access rate to be used into the feature extraction network in the trained deep learning model for feature extraction to obtain the concatenation feature. Exemplarily, the feature extraction network can be a transformer network layer or a fully connected network layer.

[0094] S105, calculating the target account's interest in the candidate resource based on the second text feature, the third text feature, and the concatenated feature of each candidate resource.

[0095] It is understandable that the second text feature and the third text feature may include text features of resources that the target account is interested in, and the spliced ​​features may include text features of the candidate resource itself and the access rate to be utilized, and the access rate to be utilized is predicted based on the text features of the candidate resource and the generated duration reflecting timeliness. The server can calculate the target account's interest in the candidate resource based on the text features of the resource that the target account is interested in, the text features of the candidate resource itself contained in the spliced ​​resource, and the timeliness of the candidate resource.

[0096] Among them, the degree of interest of an account in a resource can be a revenue score value. If the revenue score is higher, the resource may be exposed for a longer time in the client logged in by the account, the probability of the account accessing the resource is higher, the timeliness of the resource is higher, and the generated time of the resource at the time of push is shorter.

[0097] Optionally, in one implementation, the server may determine accounts with interests similar to those of the target account, and determine, among the historical resources that have been pushed to the determined account, resources that have text features similar to those of the candidate resources and have been generated for the same length of time as the candidate resources; and use the average of the revenue scores of the determined resources as the revenue score of the candidate resource.

[0098] Optionally, in another implementation, the server may use a pre-trained interest assessment model to process the second text feature, the third text feature, and the concatenated features of each candidate resource to calculate the target account's interest in the candidate resource. This implementation will be described in detail in subsequent embodiments.

[0099] It is understandable that steps S102-S105 can be implemented using various network structures in a pre-trained large model, which can predict the access rate of candidate resources pushed at the current moment, and calculate the target account's interest in the candidate resources. Optionally, in another implementation, steps S102-S103 can be implemented using an access prediction model that can predict the access rate of candidate resources pushed at the current moment; steps S104-S105 can be implemented using an interest assessment model that calculates the target account's interest in candidate resources. Compared to a large model, the access prediction model and the interest assessment model are used to implement different functions respectively, and the results output by the access prediction model and the interest assessment model are more accurate.

[0100] S106: If the interest level of the candidate resource exceeds a first preset threshold, the resource content of the candidate resource is sent to the target account.

[0101] It is understandable that if the interest level of a candidate resource is higher, the candidate resource may be exposed for a longer time on the client logged in by the target account, the candidate resource is more likely to be accessed by the target account, and the candidate resource has been generated for a shorter time at the current moment. The server can determine whether the interest level of the candidate resource exceeds a first preset threshold, and send the resource content of the candidate resource whose interest level exceeds the first preset threshold to the target account. The resource content of the candidate resource may include at least one of the following: a resource title or a resource introduction, etc.

[0102] In the embodiment of the present invention, in the process of predicting the access rate of pushing the candidate resource at the current moment, the generated time length and the first text feature of the candidate resource at the current moment are used. The generated time length can reflect the timeliness. The longer the generated time length, the lower the timeliness. The first text feature can include the text feature of the candidate resource itself. Thus, the determined access rate to be utilized takes into account the generated time length of the candidate resource at the current moment and the influence of the text content of the candidate resource. In the process of calculating the interest of the target account in the candidate resource, the text features of the first historical resource and the second historical resource and the access rate to be utilized are used. The first historical resource is the pushed resource, and the second historical resource is the browsed resource. The first historical resource and the second historical resource can include the text features of the historical resource that the target account is interested in. Thus, the interest of the target account in the candidate resource is calculated by combining the text features of the historical resource that the target account is interested in, the timeliness of the candidate resource, and the text features of the candidate resource itself. Therefore, the candidate resource whose interest level exceeds the first preset threshold meets the interest of the target account and has timeliness, which can improve the browsing value of the pushed resource and improve the user's experience of viewing the pushed resource.

[0103] Optionally, a pre-trained access prediction model may be used to implement steps S102-S103. Figure 2 As shown, in Figure 1 On the basis of the resource push method shown, step S102 includes step S1021, and step S103 includes steps S1031-S1032.

[0104] S1021, input the text content of each candidate resource into the first transformer layer in the pre-trained access prediction model to obtain the first text feature.

[0105] It is understandable that the text content of each candidate resource can be input into the first transformer layer of the pre-trained access prediction model, and the first transformer layer can be used to extract features from the text content of the candidate resource to obtain the first text feature. Exemplarily, the first transformer layer can be a network structure of a bidirectional encoder from a transformer (Bidirectional Encoder Representations from Transformers, BERT). The network structure of BERT includes the encoding part of the transformer network structure. The encoding part is composed of multiple layers of stacked self-attention and feedforward application networks, which can focus on more semantic relationships in the text content, improve the effectiveness of the extracted features, and improve the accuracy of the subsequent predicted access rate.

[0106] S1031, input the concatenated first text feature and the generated duration into the first fully connected layer in the access prediction model to obtain a first dimensionality reduction feature.

[0107] It is understandable that the server can input the concatenated first text feature and the generated duration into the first fully connected layer in the access prediction model for feature extraction to obtain the first reduced dimension feature. Exemplarily, the first fully connected layer can reduce the concatenated features of the concatenated first text feature and the generated duration into one-dimensional data.

[0108] S1032, input the first dimension reduction feature into the first output layer of the access prediction model to obtain the access rate of pushing the candidate resource at the current moment as the access rate to be utilized.

[0109] It is understandable that the server can input the first dimension reduction feature into the first output layer in the access prediction model, and map the first dimension reduction feature through the preset activation function in the first output layer to obtain the access rate of the candidate resource pushed at the current moment as the access rate to be utilized. Exemplarily, the preset activation function can be a sigmoid function, the first dimension reduction feature is one-dimensional data, and the first dimension reduction feature can be mapped to between 0 and 1 through the Sigmoid function to obtain the predicted access rate.

[0110] Among them, the access prediction model is: trained based on the text content of the pushed sample resources, the generated time of the sample resources at the push time and the actual access rate of the sample resources; the generated time of a sample resource at the push time represents the difference between the push time of the sample resource and the generation time of the sample resource; the actual access rate of a sample resource is the access rate of the sample resource when it is pushed at the push time of the sample resource; the push times of every two adjacent sample resources are separated by the same preset time.

[0111] It is understandable that the access prediction model can be a deep learning regression model, that is, the access prediction model is trained using a sample data set with known labels so that the access prediction model can find the mapping relationship between input data and output data, so that the output of the access prediction model is close to the known labels.

[0112] The device used to train the access prediction model can be a server or other device. The sample resources used to train the access prediction model are pushed resources. The generated time of the pushed sample resources at the time of pushing is the difference between the push time of the sample resources and the generation time of the sample resources. For example, if the sample resources are generated at 2 pm and pushed to the account at 3 pm, the generated time of the sample resources at the time of pushing is one hour. The actual access rate of the sample resources used for training is the ratio of the number of times the sample resources are accessed to the number of times they are exposed after being pushed.

[0113] Optionally, multiple sample resources with consecutive push times can be used for training, and the push times of every two sample resources with adjacent push times can be separated by a preset time length, wherein the preset time length can be 5 minutes. For example, the push times of the three sample resources used for training can be 2 o'clock, 2:05 and 2:10. In the embodiment of the present invention, sample resources with adjacent push times of a preset time length are used for training, and resources pushed at different times can be used as samples to improve the training effectiveness of the access prediction model.

[0114] In the process of training the access prediction model, the text content of a pushed sample resource, the generated time of the sample resource at the time of pushing, and the real access rate of the sample resource can be obtained; then the text content of a pushed sample resource and the generated time of the sample resource at the time of pushing are input into the access prediction model to be trained, and the first transformer layer in the access prediction model can be used to extract features of the text content of the sample resource to obtain the text features of the sample resource; the first fully connected layer of the access prediction model can be used to reduce the dimension of the spliced ​​text content and the generated time to obtain the first reduced dimension feature; the first output layer in the access prediction model is used to map the first dimension feature to obtain the predicted access rate; and the loss value of the predicted access rate and the real access rate is calculated. When it is judged that the access prediction model has not converged based on the loss value, the network parameters of the access prediction model are adjusted, and the steps of obtaining the text content of a pushed sample resource, the generated time of the sample resource at the time of pushing, and the real access rate of the sample resource are returned to continue training the access prediction model.

[0115] During the training of the access prediction model, the model parameters are adjusted to make the access rate output by the access prediction model as close as possible to the actual access rate. The access rate output by the trained access prediction model can represent the predicted access rate of the candidate resources pushed at the current moment.

[0116] Optionally, the server may predict the access rate of the candidate resource being pushed at the current moment by using the text content of each candidate resource, the access data volume of all resources of the category to which the candidate resource belongs, and the generation time of the candidate resource at the current moment.

[0117] It is understandable that the category to which the candidate resource belongs may be a specific resource type, a resource creator, or a resource tag. The specific resource type may be a financial type or an entertainment type, etc. The resource creator may be the generator of the resource. The resources generated by the same resource creator usually have the same characteristics. For example, the access rate of the resource generated by resource creator A is usually higher. The resource tag may be the topic classification identifier of the resource, which is usually determined by the resource creator or identified by the server. For example, the resource tag may be a brand name, an activity name, etc. The amount of access data may be the number of times the resource is accessed and the number of times the resource is pushed to the account for display. The amount of access data of all resources in the category to which the candidate resource belongs may be the number of times all resources in the same category as the candidate resource are accessed and the number of times they are pushed to the account for display. For example, if the candidate resource is news, and the news belongs to the financial type of news, the amount of access data of all resources in the category to which the candidate resource belongs is: the number of times all news belonging to the financial type is accessed and the number of times it is pushed to the account for display.

[0118] Among them, the access data volume of all resources in the category to which the candidate resource belongs are real numbers and can be stored using a dense matrix. The access data volume of all resources in the category to which the candidate resource belongs can also be called dense data. Correspondingly, the access prediction model also includes a fully connected layer for extracting features from the access data volume of all resources in the category to which the candidate resource belongs. The extracted features can be called dense features. The text content of each candidate resource is discrete data. The process of extracting features from the text content of each candidate resource to obtain the first text feature is to map the discrete data to a vector space. This process can also be called embedding, and the first text feature can also be called an embedding feature.

[0119] Accordingly, in the process of training the access resource model, the text content of a pushed sample resource, the access data volume of all resources of the category to which the sample resource belongs, the generated time of the sample resource at the time of push, and the real access rate of the sample resource can be obtained; the text content of a pushed sample resource, the access data volume of all resources of the category to which the sample resource belongs, and the generated time of the sample resource at the time of push are input into the access prediction model to be trained for training. The specific training process is similar to the training process of the access resource model mentioned above, and will not be elaborated here.

[0120] In an embodiment of the present invention, the access rate is predicted by utilizing the text content of each candidate resource, the amount of access data of all resources in the category to which the candidate resource belongs, and the generated duration of the candidate resource at the current moment. This can take into account more characteristics of the candidate resources and improve the accuracy of the predicted access rate.

[0121] Optionally, a pre-trained interest evaluation model may be used to implement steps S104-S105. Figure 3 As shown, in Figure 1 On the basis of the resource push method shown, step S104 includes step S1041, and step S105 includes steps S1051-S1052.

[0122] S1041, input the text content of the first historical resource into the second transformer layer of the pre-trained interest assessment model to obtain the second text feature; input the text content of the second historical resource into the second fully connected layer of the interest assessment model to obtain the third text feature; input the concatenation result of the text content of each candidate resource and the access rate to be used into the third fully connected layer of the interest assessment model to obtain the concatenation feature.

[0123] It is understandable that the server can input the text content of the first historical resource into the second transformer layer in the pre-trained interest assessment model, and use the second transformer layer to extract features from the text content of the first historical resource to obtain second text features; wherein the second transformer layer is a transformer network structure.

[0124] The server can input the text content of the second historical resource into the second fully connected layer in the interest evaluation model, and use the second fully connected layer to extract features from the text content of the second historical resource to obtain a third text feature. The text content of each candidate resource and the access rate to be used of the candidate resource can be spliced, and the splicing result can be feature extracted using the third fully connected layer in the interest evaluation model to obtain a splicing feature.

[0125] S1051, input the second text feature, the third text feature, and the concatenated feature of each candidate resource into the fourth fully connected layer in the interest evaluation model to obtain a second dimensionality reduction feature.

[0126] It is understandable that the server can input the second text feature, the third text feature, and the concatenated feature of each candidate resource into the fourth fully connected layer in the interest evaluation model for feature extraction to obtain a reduced second dimension reduction feature. The second dimension reduction feature can be one-dimensional data.

[0127] Among them, the second fully connected layer, the third fully connected layer and the fourth fully connected layer can be a fully connected neural network structure.

[0128] S1052, input the second dimensionality reduction feature into the second output layer of the interest evaluation model to obtain the target account's interest level in the candidate resource.

[0129] It is understandable that the server can input the second dimensionality reduction feature into the second output layer in the interest assessment model, and use the preset activation function in the second output layer to map the second dimensionality reduction feature to obtain the target account's interest level in the candidate resource.

[0130] Among them, the interest evaluation model is trained based on the text content of the target sample resources, the actual access rate of the target sample resources, the text content of the resources pushed to the sample account, the text content of the resources browsed by the sample account, and the actual interest level of the sample account in the target sample resources in a preset time period; the target sample resources are the resources pushed to the sample account in a preset time period.

[0131] It can be understood that the interest assessment model can be a deep learning regression model, which uses the actual interest level of the sample account as a known label to train the interest assessment model, so that the interest assessment model can find the mapping relationship between the input data and the output data, so that the output of the interest assessment model is close to the known label.

[0132] The device used to train the interest evaluation model can be a server or other device. The text content of the resources that have been pushed to the sample account is the relevant content of the sample account, which can also be called the resource sequence that has been pushed to the user. The resources that have been pushed to the sample account can be resources pushed in the recent period of time, and in order to improve the diversity of resources, the resources in the recent period of time can include target sample resources and more resources. For example, the target sample resources are resources pushed within 24 hours of the day, and the resources pushed in the recent period of time can be resources pushed within the past week, including the target sample resources pushed on the day; the text content of the resources that the sample account has browsed can also be called a sequence of resources browsed by the user; the features extracted from the text content of the resources that have been pushed to the sample account and the text content of the resources that the sample account has browsed can also be collectively referred to as user features. The target sample resources are resources pushed to the sample account in a preset time period, and the preset time period is a pre-set period of time. For example, the target sample resources can be resources pushed to the sample account within 24 hours of the day when the interest evaluation model is trained. In addition, whether the sample account is interested in the target sample resource is known, for example, the following information is pre-stored: whether the target sample resource is browsed by the sample account, the generation time of the target sample resource at the push time, the exposure time of the target sample resource, and the total number of target sample resources. The total number of target sample resources may be greater than or equal to 0. The exposure time of the target sample resource may be the display time on the recommendation page of the sample account when the sample account has not visited the target sample resource.

[0133] In the process of training the interest evaluation model, the text content of the target sample resources, the actual access rate of the target sample resources, the text content of the resources pushed to the sample account, and the text content of the resources browsed by the sample account can be obtained and input into the interest evaluation model to be trained, and the second transformer layer in the interest evaluation model is used to extract features of the text content of the resources pushed to the sample account to obtain the text features of the pushed resources; the second fully connected layer in the interest evaluation model is used to extract features of the text content of the resources browsed by the sample account to obtain the text features of the browsed resources; the third fully connected layer in the interest evaluation model is used to extract features of the text content of the target sample resources and the actual access rate of the target sample resources to obtain the splicing features; the fourth fully connected layer in the interest evaluation model is used to perform dimensionality reduction processing on the text features of the pushed resources, the text features of the browsed resources, and the splicing features to obtain the reduced dimensionality features; the second output layer in the interest evaluation model is used to map the reduced dimensionality features to obtain the predicted interest level of the sample account in the target sample resources; and the preset loss function is used to calculate the loss value between the predicted interest level and the actual interest level. When it is determined based on the loss value that the interest evaluation model has not converged, the network parameters of the interest evaluation model are adjusted, and the step of obtaining the text content of the pushed target sample resources, the actual access rate of the target sample resources, the text content of the resources pushed to the sample account, and the text content of the resources browsed by the sample account is returned to continue training the interest evaluation model.

[0134] During the process of training the interest evaluation model, the model parameters are adjusted to make the interest level output by the interest evaluation model as close as possible to the actual interest level. The interest level output by the trained interest evaluation model can represent the interest level of the target account in the candidate resources.

[0135] Among them, the preset loss function used when training the interest evaluation model can be a loss function for regression tasks, for example:

[0136]

[0137] Shows genuine interest. Indicates the level of interest in the forecast; is a preset parameter, which can be 0.1. Represents the loss value. When the absolute value of the difference between the actual interest level and the predicted interest level is not greater than the preset parameter, use Calculate the loss value; when the absolute value of the difference between the actual interest level and the predicted interest level is greater than the preset parameter, use , calculate the loss value.

[0138] Optionally, in another implementation, the server can use the interest evaluation model to obtain the target account's interest level in each candidate resource to be sent. The server can push candidate resources with lower interest levels to the target account, add random variables to the pushed resources used in the subsequent determination of candidate resources, collect more possibilities for iterative training of the model, and improve the accuracy of the resource matching model.

[0139] Optionally, in one implementation, the actual interest level of a sample account in the target sample resource in a preset time period is obtained according to the following steps;

[0140] For a sample account, obtain the total number of target sample resources pushed to the sample account within a preset time period, the exposure duration of each target sample resource, and the generated duration of each target sample resource at the time of pushing;

[0141] Using a preset formula, the total number of acquired resources, the generated time of each target sample resource at the push time, and the exposure time are calculated and processed to obtain the actual interest level of the sample account in the target sample resource in the preset time period;

[0142] The preset formula is: ; Indicates the actual interest of a sample account in the target sample resource during a preset time period. Indicates the total number of target sample resources pushed to the sample account within the preset time period. represents the preset parameters, and Greater than 0 and less than 1, Indicates Whether the target sample resources are accessed, Indicates The exposure duration of target sample resources, Indicates The generated duration of each target sample resource at the time of push.

[0143] It is understandable that only by checking whether the sample account browses the target sample resource to determine whether the sample account is interested in the target sample resource, less information is needed for reference. If the target sample resource is browsed and displayed on the recommendation page for a long time, it can be said that the sample account has a great interest in the target sample resource. In the preset formula, 1 indicates that the target sample resource has been browsed. 0 means that the target sample resource has not been browsed, and the target sample resource that has not been browsed is not considered. In other words, through the preset formula , you can filter out target sample resources that have not been viewed. If the exposure time is longer, the interest level is higher; if the generated time is longer, the interest level is lower. The denominator in the preset formula is The error caused by the number of target sample resources being 0 can be avoided, and the sensitivity of the degree of interest can be expanded by taking the square root of n a times. In this solution, the accurate degree of interest can be determined by a preset formula, thereby improving the accuracy of model training.

[0144] Optional, such as Figure 4 As shown, in Figure 1 Based on the resource push method shown, the method further includes steps S401-S404.

[0145] S401, obtaining the text content of each resource to be processed in a preset resource library and the account information of the target account.

[0146] It is understandable that the server can receive the real-time resource stream of Kafka sent by the resource provider, and the resource stream contains the text content of multiple resources. The server filters the received resource stream based on the flink distributed stream processing framework and stores the filtered resource stream in the preset resource library. The server can obtain the text content of each resource to be processed in the preset resource library and the account information of the target account. Among them, the account information of the target account may include at least one of the following: the text content of the resources browsed by the target account, the number of clicks of the target account on each category of resources, and the age of the user to whom the target account belongs and the registration time of the target account.

[0147] S402, extracting features from the text content of each resource to be processed to obtain a fourth text feature; and extracting features from the account information of the target account to obtain an account feature.

[0148] It can be understood that the feature extraction network in the pre-trained deep learning model can be used to perform feature extraction on the text content of each resource to be processed to obtain the fourth text feature; the feature extraction network in the pre-trained deep learning model can be used to perform feature extraction on the account information of the target account to obtain the account feature.

[0149] S403: Calculate the matching degree between the resource to be processed and the target account based on the concatenation result of the fourth text feature and the account feature.

[0150] It can be understood that the fourth text feature is the text feature of the resource to be processed, and the account feature is the feature of the account information of the target account. By comparing the similarity between the fourth text feature and the account feature, the matching degree between the resource to be processed and the target account can be determined.

[0151] S404: If the matching degree of the resource to be processed is greater than a second preset threshold, the resource to be processed is determined as a candidate resource of the target account.

[0152] It is understandable that it can be determined whether the matching degree of the resource to be processed is greater than the second preset threshold. If so, the resource to be processed is determined as a candidate resource to be pushed to the target account.

[0153] In an embodiment of the present invention, candidate resources to be pushed to a target account can be used to establish a personalized candidate resource pool for the target account, and push resources can be subsequently determined from the candidate resources, which can increase the target account's interest in the pushed resources and improve the user's experience in browsing resources.

[0154] Optional, such as Figure 5 As shown, in Figure 4 On the basis of the resource push method shown, step S402 includes step S4021, step S403 includes step S4031, and step S404 includes steps S4041-S4042.

[0155] S4021, input the text content of each resource to be processed into the third transformer layer in the pre-trained resource matching model to obtain a fourth text feature.

[0156] It is understandable that the server can input the text content of each resource to be processed into the third transformer layer in the pre-trained resource matching model, and use the third transformer layer to extract features of the text content of the resource to be processed to obtain a fourth text feature.

[0157] S4031, input the account information of the target account into the fifth fully connected layer in the resource matching model to obtain account features.

[0158] It is understandable that the fifth fully connected layer in the resource matching model can be used to extract features from the account information of the target account to obtain account features.

[0159] S4041, input the concatenation result of the fourth text feature and the account feature into the sixth fully connected layer in the resource matching model to obtain the third dimensionality reduction feature.

[0160] It is understandable that the sixth fully connected layer in the resource matching model can be used to perform dimensionality reduction processing on the concatenation result of the fourth text feature and the account feature to obtain a third dimensionality reduction feature. Exemplarily, the third dimensionality reduction feature can be one-dimensional data.

[0161] S4042, input the third dimensionality reduction feature into the third output layer in the resource matching model to obtain the matching degree between the resource to be processed and the target account.

[0162] It is understandable that the third dimensionality reduction feature can be input into the third output layer in the resource matching model, and the third dimensionality reduction feature can be mapped using the activation function preset in the third output layer to obtain the matching degree between the resource to be processed and the target account.

[0163] The resource matching model is trained based on the text content of the sample resource, the account information of the sample account, and a label indicating whether the sample account accesses the sample resource.

[0164] It is understandable that the resource matching model can be a deep learning regression model, that is, the resource matching model is trained using a sample data set with known labels, so that the resource matching model can find the mapping relationship between input data and output data, so that the output of the resource matching model is close to the known label. The device used to train the resource matching model can be a server or other device.

[0165] In the process of training the resource matching model, the text content of the sample resource, the account information of the sample account, and the label indicating whether the sample account accesses the sample resource can be obtained; then, the text content of the sample resource and the account information of the sample account are input into the resource matching model to be trained to obtain the predicted matching degree output by the resource matching model; based on the label indicating whether the sample account accesses the sample resource and the predicted matching degree, the loss value is calculated; when it is judged based on the loss value that the resource matching model has not converged, the network parameters of the resource matching model are adjusted, and the step of obtaining the text content of the sample resource, the account information of the sample account, and the label indicating whether the sample account accesses the sample resource is returned to continue training the resource matching model.

[0166] During the training process of the resource matching model, the model parameters are adjusted to make the matching degree output by the resource matching model as close as possible to the real label of whether the sample account accesses the sample resource. The matching degree output by the trained resource matching model can represent the predicted probability of the target account accessing the resource to be processed.

[0167] In the embodiment of the present invention, the resource matching model can be used to screen out candidate resources with a high probability of being accessed by the target account from the resources to be processed, and then push resources are determined from the candidate resources, which can increase the target account's interest in the pushed resources and improve the user's experience of browsing resources.

[0168] Optionally, the server may determine the matching degree between the resource to be processed and the target account by using the text content of each resource to be processed, the amount of access data of the category to which the resource to be processed belongs, and the account information of the target account.

[0169] Specifically, the resource matching model also includes a fully connected layer for feature extraction of the access data volume of the category to which the resource to be processed belongs. Through the fully connected layer, the features of the access data volume of the category to which the resource to be processed belongs can be extracted; then the sixth fully connected layer is used to extract the features of the splicing result of the fourth text feature, the account feature and the access data volume feature to obtain the feature extraction result, and the third output layer is used to map the feature extraction result to obtain the matching degree between the resource to be processed and the target resource. Correspondingly, when training the resource matching model, the text content of the sample resource, the access data volume of the category to which the sample resource belongs, and the account information of the sample account can be input into the resource matching model to be trained for training. In the embodiment of the present invention, the text content of the resource to be processed, the access data volume of the category to which the resource to be processed belongs, and the account information of the target account can be used to determine the matching degree between the resource to be processed and the target account, using more features and improving the accuracy of the matching degree.

[0170] The resource push method will be introduced below in conjunction with a schematic diagram of the principle of the resource push method.

[0171] like Figure 6 As shown, the server can receive the Kafka resource stream sent by the resource provider, filter the received resource stream based on the Flink distributed stream processing framework, and store the filtered resource stream in the preset resource library, that is, the resource screening process. Then, the text content of each resource to be processed in the preset resource library and the account information of the target account are obtained, and the resource features and user features are obtained. The resource features and user features can be stored in the redis database, that is, the resource features are obtained from the resource feature database, and the user features are obtained from the user feature database; the resource matching model is used to process the obtained text content and account information, and the matching degree between each resource to be processed and the target account is obtained, so as to screen each resource to be processed in the preset resource library and determine the candidate resources to be sent to the target account, that is, to establish a personalized resource candidate pool.

[0172] The server can obtain the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generated time of each candidate resource at the current moment. Then, the access prediction model is used to process the text content and generated time of each candidate resource to obtain the access rate to be used of the candidate resource; that is, the implementation process of the timeliness analysis task.

[0173] The server can use the interest evaluation model to process the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generated duration of each candidate resource at the current moment, to obtain the target account's interest level in the candidate resource; that is, the realization of the benefit analysis task.

[0174] If the interest level of the candidate resource exceeds a first preset threshold, the server sends the resource content of the candidate resource to the target account, that is, determines the final push resource.

[0175] Optionally, the server also has a frequency control module, through which the determined final push resource is pushed to the target account when the current number of pushes to the target account does not exceed a push number threshold.

[0176] The embodiment of the present invention also provides a resource push device, such as Figure 7 As shown, the device comprises:

[0177] A first acquisition module 710 is used to acquire the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generation time of each candidate resource at the current moment;

[0178] A first extraction module 720, configured to extract features from the text content of each candidate resource to obtain a first text feature;

[0179] The prediction module 730 is used to predict the access rate of pushing the candidate resource at the current moment based on the first text feature and the generated time of each candidate resource, so as to obtain the access rate to be utilized;

[0180] The second extraction module 740 is used to extract features from the text content of the first historical resource to obtain a second text feature; extract features from the text content of the second historical resource to obtain a third text feature; and extract features from the splicing result of the text content of each candidate resource and the access rate to be used to obtain a splicing feature;

[0181] A calculation module 750, configured to calculate the interest level of the target account in the candidate resource based on the second text feature, the third text feature, and a concatenated feature of each candidate resource;

[0182] The sending module 760 is configured to send the resource content of the candidate resource to the target account if the interest level of the candidate resource exceeds a first preset threshold.

[0183] Optionally, the first extraction module 720 is specifically used to input the text content of each candidate resource into the first transformer layer of a pre-trained access prediction model to obtain a first text feature; wherein the access prediction model is: trained based on the text content of the pushed sample resources, the generated time of the sample resources at the push time, and the real access rate of the sample resources; the generated time of a sample resource at the push time represents the difference between the push time of the sample resource and the generation time of the sample resource; the real access rate of a sample resource is the access rate of the sample resource when the sample resource is pushed at the push time of the sample resource; the push times of every two sample resources adjacent to each other are separated by the same preset time;

[0184] The prediction module 730 is specifically used to input the concatenated first text feature and the generated duration into the first fully connected layer in the access prediction model to obtain a first dimensionality reduction feature; input the first dimensionality reduction feature into the first output layer in the access prediction model to obtain the access rate of the candidate resource pushed at the current moment as the access rate to be utilized.

[0185] Optionally, the second extraction module 740 is specifically used to input the text content of the first historical resource into the second transformer layer of the pre-trained interest evaluation model to obtain a second text feature; wherein the interest evaluation model is trained based on the text content of the target sample resource, the actual access rate of the target sample resource, the text content of the resource pushed to the sample account, the text content of the resource browsed by the sample account, and the actual interest level of the sample account in the target sample resource in a preset time period; the target sample resource is the resource pushed to the sample account in the preset time period;

[0186] Inputting the text content of the second historical resource into the second fully connected layer in the interest assessment model to obtain a third text feature;

[0187] Inputting the concatenation result of the text content and the access rate to be utilized of each candidate resource into the third fully connected layer in the interest evaluation model to obtain a concatenated feature;

[0188] Inputting the second text feature, the third text feature, and the concatenated feature of each candidate resource into the fourth fully connected layer in the interest evaluation model to obtain a second dimensionality reduction feature;

[0189] The second dimensionality reduction feature is input into the second output layer of the interest evaluation model to obtain the degree of interest of the target account in the candidate resource.

[0190] Optionally, the actual interest level of a sample account in the target sample resource in a preset time period is obtained according to the following steps;

[0191] For a sample account, obtain the total number of target sample resources pushed to the sample account within a preset time period, the exposure duration of each target sample resource, and the generated duration of each target sample resource at the time of pushing;

[0192] Using a preset formula, the total number of acquired resources, the generated time of each target sample resource at the push time, and the exposure time are calculated and processed to obtain the real interest level of the sample account in the target sample resource in the preset time period;

[0193] Wherein, the preset formula is: ; Indicates the actual interest of a sample account in the target sample resource during a preset time period. Indicates the total number of target sample resources pushed to the sample account within the preset time period. represents the preset parameters, and Greater than 0 and less than 1, Indicates Whether the target sample resources are accessed, Indicates The exposure duration of target sample resources, Indicates The generated duration of each target sample resource at the time of push.

[0194] Optionally, the device further comprises:

[0195] The second acquisition module is used to acquire the text content of each resource to be processed in the preset resource library and the account information of the target account;

[0196] A third extraction module is used to extract features from the text content of each resource to be processed to obtain a fourth text feature; and to extract features from the account information of the target account to obtain an account feature;

[0197] A first determination module, configured to calculate a matching degree between the resource to be processed and the target account based on a concatenation result of the fourth text feature and the account feature;

[0198] The second determination module is configured to determine the resource to be processed as a candidate resource of the target account if the matching degree of the resource to be processed is greater than a second preset threshold.

[0199] an optional third extraction module, specifically used to input the text content of each resource to be processed into the third transformer layer of a pre-trained resource matching model to obtain a fourth text feature; wherein the resource matching model is trained based on the text content of the sample resource, the account information of the sample account, and a label indicating whether the sample account has accessed the sample resource;

[0200] Inputting the account information of the target account into the fifth fully connected layer in the resource matching model to obtain account features;

[0201] A first determination module is specifically used to input the concatenation result of the fourth text feature and the account feature into the sixth fully connected layer in the resource matching model to obtain a third dimensionality reduction feature;

[0202] The third dimensionality reduction feature is input into the third output layer in the resource matching model to obtain the matching degree between the resource to be processed and the target account.

[0203] In the embodiment of the present invention, in the process of predicting the access rate of pushing the candidate resource at the current moment, the generated time length and the first text feature of the candidate resource at the current moment are used. The generated time length can reflect the timeliness. The longer the generated time length, the lower the timeliness. The first text feature can include the text feature of the candidate resource itself. Thus, the determined access rate to be utilized takes into account the generated time length of the candidate resource at the current moment and the influence of the text content of the candidate resource. In the process of calculating the interest of the target account in the candidate resource, the text features of the first historical resource and the second historical resource and the access rate to be utilized are used. The first historical resource is the pushed resource, and the second historical resource is the browsed resource. The first historical resource and the second historical resource can include the text features of the historical resource that the target account is interested in. Thus, the interest of the target account in the candidate resource is calculated by combining the text features of the historical resource that the target account is interested in, the timeliness of the candidate resource and the text features of the candidate resource itself. Therefore, the candidate resource whose interest level exceeds the first preset threshold meets the interest of the target account and has timeliness, which can improve the browsing value of the pushed resource and improve the user's experience of viewing the pushed resource.

[0204] The embodiment of the present invention further provides an electronic device, such as Figure 8 As shown, it includes a processor 801 , a communication interface 802 , a memory 803 and a communication bus 804 , wherein the processor 801 , the communication interface 802 , and the memory 803 communicate with each other via the communication bus 804 .

[0205] Memory 803, used for storing computer programs;

[0206] The processor 801 is used to implement the aforementioned resource push method when executing the program stored in the memory 803.

[0207] The communication bus mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0208] The communication interface is used for communication between the above electronic device and other devices.

[0209] The memory may include a random access memory (RAM) or a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0210] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0211] In another embodiment of the present invention, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned resource push methods are implemented.

[0212] In another embodiment of the present invention, a computer program product including instructions is provided. When the computer program product is run on a computer, the computer executes any resource pushing method in the above embodiments.

[0213] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.

[0214] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0215] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0216] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A resource push method, characterized in that: The method comprises: Obtain the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generation time of each candidate resource at the current moment; Extracting features from the text content of each candidate resource to obtain a first text feature; Based on the first text feature and the generated time of each candidate resource, predict the access rate of pushing the candidate resource at the current moment to obtain the access rate to be utilized; Extracting features from the text content of the first historical resource to obtain a second text feature; extracting features from the text content of the second historical resource to obtain a third text feature; and extracting features from a concatenation result of the text content of each candidate resource and the access rate to be utilized to obtain a concatenation feature; Calculating the interest level of the target account in the candidate resource based on the second text feature, the third text feature, and the concatenated feature of each candidate resource; If the interest level of the candidate resource exceeds a first preset threshold, the resource content of the candidate resource is sent to the target account.

2. The method according to claim 1, characterized in that The step of extracting features from the text content of each candidate resource to obtain a first text feature includes: Input the text content of each candidate resource into the first transformer layer of the pre-trained access prediction model to obtain the first text feature; wherein the access prediction model is trained based on the text content of the pushed sample resources, the generated time of the sample resources at the push time, and the real access rate of the sample resources; the generated time of a sample resource at the push time represents the difference between the push time of the sample resource and the generation time of the sample resource; the real access rate of a sample resource is the access rate of the sample resource when it is pushed at the push time of the sample resource; the push time of every two sample resources adjacent to each other is the same preset time distance; The method predicts the access rate of pushing the candidate resource at the current moment based on the first text feature and the generated time of each candidate resource to obtain the access rate to be utilized, including: Inputting the concatenated first text feature and the generated duration into the first fully connected layer in the access prediction model to obtain a first dimensionality reduction feature; The first dimension reduction feature is input into the first output layer of the access prediction model to obtain the access rate of pushing the candidate resource at the current moment as the access rate to be utilized.

3. The method according to claim 1, characterized in that The extracting features of the text content of the first historical resource to obtain second text features includes: Inputting the text content of the first historical resource into the second transformer layer of the pre-trained interest evaluation model to obtain a second text feature; wherein the interest evaluation model is trained based on the text content of the target sample resource, the actual access rate of the target sample resource, the text content of the resource pushed to the sample account, the text content of the resource browsed by the sample account, and the actual interest level of the sample account in the target sample resource in a preset time period; the target sample resource is the resource pushed to the sample account in the preset time period; The extracting features of the text content of the second historical resource to obtain a third text feature includes: Inputting the text content of the second historical resource into the second fully connected layer in the interest assessment model to obtain a third text feature; The feature extraction of the splicing result of the text content and the access rate to be utilized of each candidate resource to obtain the splicing feature includes: Inputting the concatenation result of the text content and the access rate to be utilized of each candidate resource into the third fully connected layer in the interest evaluation model to obtain a concatenated feature; The calculating the interest degree of the target account in the candidate resource based on the second text feature, the third text feature, and the concatenated feature of each candidate resource includes: Inputting the second text feature, the third text feature, and the concatenated feature of each candidate resource into the fourth fully connected layer in the interest evaluation model to obtain a second dimensionality reduction feature; The second dimensionality reduction feature is input into the second output layer of the interest evaluation model to obtain the degree of interest of the target account in the candidate resource.

4. The method according to claim 3, characterized in that The actual interest level of a sample account in the target sample resource in a preset time period is obtained by following the steps below; For a sample account, obtain the total number of target sample resources pushed to the sample account within a preset time period, the exposure duration of each target sample resource, and the generated duration of each target sample resource at the time of pushing; Using a preset formula, the total number of acquired resources, the generated time of each target sample resource at the push time, and the exposure time are calculated and processed to obtain the real interest level of the sample account in the target sample resource in the preset time period; Wherein, the preset formula is: ; Indicates the actual interest of a sample account in the target sample resource during a preset time period. Indicates the total number of target sample resources pushed to the sample account within the preset time period. represents the preset parameters, and Greater than 0 and less than 1, Indicates Whether the target sample resources are accessed, Indicates The exposure duration of target sample resources, Indicates The generated duration of each target sample resource at the time of push.

5. The method according to any one of claims 1 to 4, characterized in that: Before acquiring the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generation time of each candidate resource at the current moment, the method further includes: Obtain the text content of each resource to be processed in the preset resource library and the account information of the target account; Extracting features from the text content of each resource to be processed to obtain a fourth text feature; and extracting features from the account information of the target account to obtain an account feature; Calculating the matching degree between the resource to be processed and the target account based on the concatenation result of the fourth text feature and the account feature; If the matching degree of the resource to be processed is greater than a second preset threshold, the resource to be processed is determined as a candidate resource of the target account.

6. The method according to claim 5, characterized in that Perform feature extraction on the text content of each resource to be processed to obtain a fourth text feature, including: Inputting the text content of each resource to be processed into the third transformer layer of the pre-trained resource matching model to obtain a fourth text feature; wherein the resource matching model is trained based on the text content of the sample resource, the account information of the sample account, and a label indicating whether the sample account has accessed the sample resource; The feature extraction of the target account information to obtain the account features includes: Inputting the account information of the target account into the fifth fully connected layer in the resource matching model to obtain account features; The calculating the matching degree between the resource to be processed and the target account based on the concatenation result of the fourth text feature and the account feature includes: Inputting the concatenation result of the fourth text feature and the account feature into the sixth fully connected layer in the resource matching model to obtain a third dimensionality reduction feature; The third dimensionality reduction feature is input into the third output layer in the resource matching model to obtain the matching degree between the resource to be processed and the target account.

7. A resource push device, characterized in that: The device comprises: A first acquisition module is used to acquire the text content of the first historical resource that has been pushed to the target account, the text content of the second historical resource that has been browsed by the target account, the text content of each candidate resource to be pushed to the target account, and the generation time of each candidate resource at the current moment; A first extraction module, used to extract features from the text content of each candidate resource to obtain a first text feature; A prediction module, configured to predict the access rate of pushing the candidate resource at the current moment based on the first text feature and the generated time of each candidate resource, and obtain the access rate to be utilized; A second extraction module is used to extract features from the text content of the first historical resource to obtain a second text feature; extract features from the text content of the second historical resource to obtain a third text feature; and extract features from a splicing result of the text content of each candidate resource and the access rate to be utilized to obtain a splicing feature; A calculation module, configured to calculate the interest level of the target account in the candidate resource based on the second text feature, the third text feature, and a concatenated feature of each candidate resource; The sending module is used to send the resource content of the candidate resource to the target account if the interest level of the candidate resource exceeds a first preset threshold.

8. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, for implementing any of the methods described in claims 1-6 when executing a program stored in a memory.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

10. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 6 when executed by a processor.

Images