Information recommendation and model training method and device, equipment and storage medium

Abstract

The invention provides an information recommendation method and device, an information recommendation model training method and device, equipment and a storage medium, and the recommendation model training method comprises the steps: carrying out the sampling of a global sample based on a preset exposure growth rate, and obtaining a training data set; the global sample comprises samples under each content recommendation scene; training a recommendation model by adopting the training data set to obtain a basic recommendation model suitable for each content recommendation scene; each of a recall layer model and a sorting layer model in the recommendation model comprises a main network and an offset network; and based on the personalized data set under each content recommendation scene, respectively adjusting the basic recommendation model to obtain a personalized recommendation model of each content recommendation scene. According to the method, the recommendation system model comprising the universal unbiased basic model and the personalized recommendation models can be obtained.

Description

Technical Field

[0001] The present application belongs to the field of model training technology, and specifically relates to an information recommendation and its model training method, device, equipment and storage medium. Background Art

[0002] Information flow applications usually contain different types of content, including news, pictures, books, audio, short videos or long videos, etc., and different types of content are usually displayed in different channels and different locations.

[0003] When cold-starting comprehensive content recommendations in new scenarios, it is impossible to train personalized recommendation models due to the lack of targeted training data. When using global samples as training data, the differences in content and consumption of each channel lead to deviations in recommendation model training. Therefore, when cold-starting in new scenarios, the personalized recommendation effect of the model is not very good. Summary of the invention

[0004] The present application proposes an information recommendation and model training method, device, equipment and storage medium, which can obtain a recommendation system model including a general unbiased basic model and multiple personalized recommendation models.

[0005] The first embodiment of the present application proposes a recommendation model training method, including:

[0006] Sampling the global samples based on a preset exposure growth rate to obtain a training data set; the global samples include samples in various content recommendation scenarios;

[0007] The training data set is used to train a recommendation model to obtain a basic recommendation model suitable for each content recommendation scenario; the recall layer model and the ranking layer model in the recommendation model both include a main network and a bias network;

[0008] Based on the personalized data sets in each content recommendation scenario, the basic recommendation models are adjusted respectively to obtain personalized recommendation models for each content recommendation scenario.

[0009] In some embodiments of the present application, sampling the global samples based on a preset exposure growth rate includes:

[0010] Determine the exposure number of each sample in the global sample respectively;

[0011] Based on a preset exposure growth rate and the exposure number of each sample, multiple samples are selected from the global samples; the exposure numbers of the multiple samples are increased according to the preset exposure growth rate.

[0012] In some embodiments of the present application, the selecting a plurality of samples from the global samples based on a preset exposure growth rate and the number of exposures of each sample includes:

[0013] Based on the exposure number of each sample, samples with the same exposure number are grouped into sets to obtain multiple sample sets;

[0014] Sorting the multiple sample sets based on the exposure number corresponding to each sample set, and selecting a preset number of sample sets with the highest exposure numbers from the sorted sample sets;

[0015] Selecting, from the preset number of sample sets, a plurality of target sample sets whose exposure numbers increase according to the preset exposure growth rate;

[0016] Based on the exposure number corresponding to each target sample set, samples are selected from each target sample set.

[0017] In some embodiments of the present application, the recall layer model includes a first main network and a first bias network; the step of training the recommendation model using the training data set includes:

[0018] Inputting the user feature data and the recommended content included in each sample in the training data set into the first main network respectively to obtain the relationship between the recommended content and the user feature;

[0019] Inputting the bias features in each sample into the first bias network respectively to obtain bias parameters of each sample data;

[0020] Based on the relationship between the recommended content and the user characteristics, and the bias parameters of each sample data, a predicted output value is determined; the predicted output value is used to indicate the degree of influence of the bias feature on whether the sample is clicked.

[0021] In some embodiments of the present application, the ranking layer model includes a second main network and a second bias network; the training of the recommendation model using the training data set also includes:

[0022] Based on the predicted output value, determining a new training data set;

[0023] Inputting new user feature data and new recommended content included in each sample of the new training data set into the second main network respectively, to obtain the relationship between the new recommended content and the new user feature;

[0024] Inputting the bias features in each sample in the new training data set into the second bias network respectively to obtain the bias parameters of each new sample data;

[0025] Based on the relationship between the new recommended content and the new user feature, and the bias parameter of each new sample data, a new predicted output value is determined; the new predicted output value is used to indicate the degree of influence of the bias feature on whether the sample is clicked.

[0026] In some embodiments of the present application, adjusting the basic recommendation model based on the personalized data set in each content recommendation scenario to obtain the personalized recommendation model for each content recommendation scenario includes:

[0027] Deleting the bias network in the basic recommendation model to obtain a new basic recommendation model;

[0028] Based on the personalized data sets in each scenario, the new basic recommendation model is fine-tuned to obtain the personalized recommendation models corresponding to each scenario.

[0029] The second aspect of the present application provides an information recommendation method, the method comprising:

[0030] Obtain user feature data of target users in the target content recommendation scenario;

[0031] Inputting the user feature data into a personalized recommendation model corresponding to the target content recommendation scenario, and outputting the content to be recommended; the personalized recommendation model is trained by the recommendation model training method described in the first aspect;

[0032] The content to be recommended is sent to the target user.

[0033] In some embodiments of the present application, inputting the user feature data into a personalized recommendation model corresponding to the target content recommendation scenario and outputting the content to be recommended includes:

[0034] If a target personalized recommendation model for the target content recommendation scenario has been trained, the user feature data is input into the target personalized recommendation model, and the content to be recommended is output.

[0035] In some embodiments of the present application, inputting the user feature data into a personalized recommendation model corresponding to the target content recommendation scenario and outputting the content to be recommended includes:

[0036] If the target personalized recommendation model for the target content recommendation scenario has not been trained yet, the user feature data is respectively input into the personalized recommendation models for each content recommendation scenario that has been trained, to obtain multiple groups of candidate content output by the personalized recommendation models for each content recommendation scenario;

[0037] The content to be recommended is screened out from the multiple groups of candidate content.

[0038] In some embodiments of the present application, the method further includes:

[0039] When a target personalized data set of the target content recommendation scenario is obtained, the basic recommendation model is adjusted based on the target personalized data set to obtain a target personalized recommendation model of the target content recommendation scenario.

[0040] The embodiment of the third aspect of the present application provides a recommendation model training device, including:

[0041] A training data acquisition module, used to sample global samples based on a preset exposure growth rate to obtain a training data set; the global samples include samples in various content recommendation scenarios;

[0042] A basic model training module, used to train a recommendation model using the training data set to obtain a basic recommendation model suitable for each content recommendation scenario; the recall layer model and the ranking layer model in the recommendation model both include a main network and a bias network;

[0043] The personalized model training module is used to adjust the basic recommendation model based on the personalized data sets in each content recommendation scenario to obtain a personalized recommendation model for each content recommendation scenario.

[0044] The fourth aspect of the present application provides an information recommendation device, the device comprising:

[0045] A data acquisition module is used to acquire user characteristic data of target users in a target content recommendation scenario;

[0046] A content determination module, configured to input the user feature data into a personalized recommendation model corresponding to the target content recommendation scenario, and output the content to be recommended; the personalized recommendation model is trained by the recommendation model training method described in the first aspect;

[0047] The content sending module is used to send the content to be recommended to the target user.

[0048] An embodiment of the fifth aspect of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect or the second aspect above.

[0049] The embodiment of the sixth aspect of the present application provides a computer-readable storage medium on which a computer program is stored, and the program is executed by a processor to implement the method described in the first aspect or the second aspect above.

[0050] The technical solution provided in the embodiments of the present application has at least the following technical effects or advantages:

[0051] In the embodiment of the present application, the global samples are first sampled based on the preset exposure growth rate to obtain a training data set; then the recommendation model is trained using the training data set to obtain a basic recommendation model, and then the personalized data set under each content recommendation scenario is used to fine-tune the basic recommendation model to obtain the personalized recommendation model of each content recommendation scenario. In this way, by selecting the training data set based on the preset exposure growth rate, more samples with smaller exposure numbers can be collected, which improves the diversity of samples in the training data set, reduces the degree to which the trained basic recommendation model is affected by samples with larger exposure numbers, and improves the unbiasedness of the basic recommendation model. In addition, the recall layer model and the ranking layer model in the basic recommendation model both include a main network and a bias network. Based on the debiasing effect of the bias network, the bias problem caused by different exposure channels and different exposure positions of the global samples can be solved. Through the above two methods, an unbiased and reliable basic recommendation model can be trained when using global sample training. Then, the personalized data set under each content recommendation scenario is used to fine-tune the basic recommendation model, and the personalized recommendation model of each content recommendation scenario can be quickly obtained, thereby improving the scalability of the model.

[0052] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] By reading the detailed description of the preferred embodiment below, various other advantages and benefits will become clear to those of ordinary skill in the art. The accompanying drawings are only used for the purpose of illustrating the preferred embodiment and are not considered to be limitations of the present application. In addition, the same reference symbols are used to represent the same components throughout the accompanying drawings.

[0054] In the attached picture:

[0055] Figure 1 A schematic diagram showing a flow chart of a recommendation model training method provided in some embodiments of the present application;

[0056] Figure 2 A specific flow chart of step S2 in some embodiments of the present application is shown;

[0057] Figure 3 A schematic diagram of the structure of the recall layer model in some embodiments of the present application is shown;

[0058] Figure 4 A schematic diagram of the structure of a sorting layer model in some embodiments of the present application is shown;

[0059] Figure 5 A specific flow chart of step S3 in some embodiments of the present application is shown;

[0060] Figure 6 A schematic diagram showing a flow chart of an information recommendation method provided in some embodiments of the present application;

[0061] Figure 7 A schematic diagram showing the structure of a recommendation model training device provided in some embodiments of the present application is shown;

[0062] Figure 8 A schematic diagram showing the structure of an information recommendation device provided in some embodiments of the present application is shown;

[0063] Fig. 9 A schematic diagram of an electronic device provided by an embodiment of the present application is shown;

[0064] Fig.10 A schematic diagram of a storage medium provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0065] The exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0066] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by technicians in the field to which this application belongs.

[0067] For information flow applications, such as some video websites, e-commerce websites, etc., they usually contain different types of content, such as but not limited to news, pictures, books, audio, short videos or long videos, etc., and different types of content are usually displayed in different channels and different locations. Here, the channel can be understood as the website location where the user is located when the content is recommended to the user, which is related to the path the user takes to browse the website. For example, an e-commerce website can include various types of products, and the specific classification of products can be regarded as a channel. In order to increase the diversity of products and activities, different access points may be set up on the website, such as discounted products, live broadcast rooms, etc., which can also be regarded as content channels. The location of the content can be understood as the display page and page location where the recommended content is located when it is displayed to the user, such as the third page on the second line of the first page, etc.

[0068] In order to recommend content that better meets user needs to improve user experience satisfaction, the basic recommendation model can be trained based on the data of existing scenarios to obtain the corresponding recommendation models for each scenario. The existing scenarios here include but are not limited to existing users performing category searches, existing users performing keyword searches, and users browsing from the homepage. When recommending content to users in existing scenarios, the basic recommendation model can be trained based on the existing user feature data in the existing scenarios and the corresponding recommended content to obtain a personalized recommendation model. However, when cold-starting content recommendations (recommending new users or new items) in a new scenario, a personalized recommendation model cannot be trained due to the lack of targeted training data.

[0069] In related technologies, global samples can be used as training data. However, due to differences in content and consumption among channels, there are deviations in the training of the recommendation model. Therefore, when cold starting in a new scenario, the personalized recommendation effect of the model is not very good.

[0070] Based on the above reasons, an embodiment of the present application provides a recommendation model training method, which first samples the global samples based on a preset exposure growth rate to obtain a training data set; then uses the training data set to train the recommendation model (for ease of understanding, the recommendation model before training can be referred to as the original recommendation model) to obtain a basic recommendation model, and then uses the personalized data set under each content recommendation scenario to fine-tune the basic recommendation model to obtain the personalized recommendation model of each content recommendation scenario. In this way, by selecting a training data set based on a preset exposure growth rate, more samples with smaller exposure numbers can be collected, which improves the diversity of samples in the training data set, reduces the degree to which the trained basic recommendation model is affected by samples with larger exposure numbers, and improves the unbiasedness of the basic recommendation model. In addition, the recall layer model and the ranking layer model in the basic recommendation model both include a main network and a bias network. Based on the debiasing effect of the bias network, the bias problem caused by different exposure channels and different exposure positions of the global samples can be solved. Through the above two methods, an unbiased and reliable basic recommendation model can be trained when using global sample training. Then, personalized data sets in various content recommendation scenarios are used to fine-tune the basic recommendation model, so that personalized recommendation models for various content recommendation scenarios can be quickly obtained, thereby improving the scalability of the model.

[0071] Among them, the global samples include samples in various content recommendation scenarios, and the samples in various content recommendation scenarios may include the characteristic information of the requesting user and the recommended content in each content recommendation scenario, as well as the characteristic information of the recommended content. The characteristic information of the requesting user includes but is not limited to the user name, whether it is a new user, the browsing path of the recommended content, whether the recommended content is clicked, etc. The characteristic information of the content includes but is not limited to the channel and location where the content is sent, the user who receives the recommended content, and whether the content is clicked, etc.

[0072] The execution subject of the recommendation model training method can be a computer, server, or server cluster used for model training, or a recommendation model training device that specifically executes the recommendation model training method on these devices, and this embodiment does not make any specific limitations on this.

[0073] The specific structure of the above original recommendation model can be any information recommendation model as long as its recall layer model and ranking layer model both include a main network and a bias network.

[0074] It is understandable that the above only mentions the recall layer model and the ranking layer model of the original recommendation model, but it does not mean that the original recommendation model only includes the recall layer model and the ranking layer model. In addition to the recall layer model and the ranking layer model, it can also include other necessary models, such as the data layer model, the personalization layer model and the feedback layer model. The data layer model mainly involves the collection, storage and preprocessing of data. The data here may include user behavior data, item information, user characteristics, etc. Based on the data layer model, the original data can be cleaned, deduplicated, normalized, etc. for subsequent model use. The personalization layer model is mainly responsible for customizing the recommendation results according to the user's personalized characteristics and preferences. The personalization layer model can adjust the recommendation results according to the user's historical behavior, interest tags, social relationships and other information to provide a recommendation experience that is more in line with the user's interests and preferences. The feedback layer model is responsible for collecting user feedback information on the recommendation results, such as user clicks, purchases, ratings and other behaviors. This feedback information can be used to train models, optimize recommendation strategies and improve the performance of recommendation systems.

[0075] The recommendation model training method provided in the embodiment of the present application is described in detail below with reference to the accompanying drawings.

[0076] See also Figure 1 , is a flow chart of the recommendation model training method provided in the embodiment of the present application, such as Figure 1 As shown, the recommendation model training method may include the following steps:

[0077] Step S1, sampling the global samples based on a preset exposure growth rate to obtain a training data set; the global samples include samples in various content recommendation scenarios.

[0078] Among them, the exposure growth rate is the growth rate of the number of exposures, and the number of exposures can be understood as the number of times the content in the sample is recommended to the user (regardless of whether the user clicks on it, as long as the recommended content is displayed to the user, it is considered effective exposure). When sampling based on the global sample, the content in the sample can be sorted in order from high to low based on the number of exposures of the content in the sample, and then the top N content (N can be any natural number) with the highest number of exposures, that is, the TopN exposed content, and its related feature information and user behavior information are selected to form a training data set. In order to reduce the impact of content with a high number of exposures on model training, the growth rate of the number of exposures between the TopK exposures (K can be any natural number, K is less than or equal to N) and the TopK-1 exposures can be limited. For example, assuming that the growth rate of the limited exposure number is r, if the exposure number of the content ranked K+1 in the exposure number is M, then the exposure number of the Kth content is limited to M*(1+r), and the Kth content is randomly sampled based on this exposure number.

[0079] Based on the introduction of the above exposure growth rate, such as Figure 2 As shown, the above step S1 may include the following specific steps: step S11, respectively determining the exposure number of each sample in the global sample; step S12, selecting multiple samples from the global sample based on a preset exposure growth rate and the exposure number of each sample; the exposure numbers of the multiple samples increase according to the preset exposure growth rate.

[0080] The preset exposure growth rate may be any integer value, and those skilled in the art may set it according to actual conditions, and this embodiment does not specifically limit this.

[0081] In practical applications, after determining the exposure number of each sample in the global sample, a reasonable exposure growth rate can be set according to the exposure number of each sample. For example, if the exposure numbers of each sample are relatively concentrated, a smaller exposure growth rate can be set to prevent the exposure number of the subsequently selected sample from being too large or too small. If the exposure numbers of each sample are relatively dispersed, a larger exposure growth rate can be set. Then, based on the exposure number of each sample and the set exposure growth rate, sampling can be performed from the global sample in an increasing manner according to the preset exposure growth rate to select samples with relatively dispersed exposure numbers as training data sets, thereby increasing the diversity of samples and reducing the bias effect of samples with larger exposure numbers on model training.

[0082] Furthermore, the above-mentioned step of selecting multiple samples from the global samples based on the preset exposure growth rate and the exposure number of each sample may include the following processing: based on the exposure number of each sample, grouping the samples with the same exposure number into sets to obtain multiple sample sets; sorting the multiple sample sets based on the exposure number corresponding to each sample set, and selecting a preset number of sample sets with the highest exposure numbers from each sorted sample set; selecting multiple target sample sets whose exposure numbers increase according to the preset exposure growth rate from the preset number of sample sets; and selecting samples from each target sample set based on the exposure number corresponding to each target sample set.

[0083] In actual applications, there are a lot of contents in the global samples, and there will be a lot of samples with the same exposure number. When selecting samples with a certain exposure number, since there are a large number of samples with this exposure number, the number of samples corresponding to this exposure number can be selected from the samples with this exposure number. For example, if the highest exposure number is a, then select a sample from the set of samples with an exposure number of a. If the set of samples with an exposure number of a contains less than a samples, then all samples with an exposure number of a can be selected, or other selection rules can be preset, such as selecting 80% or one-half of the exposure number.

[0084] Step S2, using the training data set to train the recommendation model to obtain a basic recommendation model suitable for each content recommendation scenario; the recall layer model and the ranking layer model in the recommendation model both include a main network and a bias network.

[0085] Among them, the recall layer model and the ranking layer model are both important components of the recommendation model. The main task of the recall layer model is to quickly screen out some candidate samples from a large number of samples. The design goal of the recall layer model is a high recall rate, that is, to cover the user's interest areas as much as possible to provide diverse recommendation results. The output of the recall layer model is used as the input of the ranking layer model. Through the ranking layer model, each sample can be further finely sorted and personalized. Through the pre-screening of the recall layer model, the computational complexity of the ranking layer model can be reduced, and the performance and efficiency of the recommendation model can be improved.

[0086] In some embodiments, the recall layer model may include a first main network and a first bias network. The above step of using the training data set to train the recommendation model may include the following processing: inputting the user feature data and recommended content included in each sample in the training data set into the first main network respectively to obtain the relationship between the recommended content and the user features; inputting the bias features in each sample into the first bias network respectively to obtain the bias parameters of each sample data; based on the relationship between the recommended content and the user features, and the bias parameters of each sample data, determining the predicted output value; the predicted output value is used to indicate the degree of influence of the bias feature on whether the sample is clicked.

[0087] Among them, bias features include but are not limited to basic features such as the channel and location of the content in the sample, as well as statistical features such as the number of exposures, and any feature that can affect whether the content in the sample is clicked by the user after exposure can become a bias feature. The bias parameter can be used to indicate the degree of influence of the bias feature on whether the sample is clicked, and is a quantified value. In this embodiment, its specific value is not specified. For example, if 1 is considered to be an inevitable influence, the bias parameter can be a value greater than 0 and less than 1.

[0088] In this embodiment, the recall layer model includes not only the main network (Main Net), but also a bias network (BiasNet, also called a debiasing network). For the convenience of distinguishing from the main network and bias network of the sorting layer, they can be respectively referred to as the first main network and the first bias network, wherein the first main network is the main structure of the recall layer model, and can be any neural network model with a recall function (quickly screening out a part of candidate items from a large-scale item library for subsequent sorting operations). This embodiment does not specifically limit its specific structure, such as a twin tower model, a multi-vector recall model, etc. The first bias network is used to process the features that need to be debiased (hereinafter referred to as bias features), and extract more comprehensive information as much as possible to improve the fit of the model, that is, to connect all the debiasing features, and then pass through several layers of full connection to finally obtain a bias parameter (bias logits). This embodiment does not specifically limit the specific structure of the bias network, and it can be any neural network model with a debiasing function.

[0089] Specifically, the structure of the recall layer model can be as follows Figure 3 As shown, the first main network includes a feature embedding layer, a multi-head self-attention layer, a connection layer, a multi-layer perceptron layer and an L2 normalization layer, wherein the feature embedding layer is used to perform preliminary feature embedding on the user configuration information, the user behavior information and the target behavior information to be recalled, and convert the information into corresponding vectors, which can be recorded as user configuration vector, user behavior vector and target behavior vector respectively.

[0090] The multi-head self-attention layer can map the input sequence into multiple attention heads, each head learns different attention weights, and then merges the outputs of these heads into the final attention representation, so that different semantic information can be processed in parallel, thereby more effectively capturing the association relationship in the sequence. In this embodiment, the recall layer model can learn the association information between multiple user behavior features through the multi-head self-attention layer and form a final attention representation vector to better select user behavior information related to the query, which can be used to capture the semantic relationship between the query and multiple user behavior features, helping to improve the recall effect.

[0091] The connection layer (Concat layer) is used to perform the operation of connecting (splicing) multiple input tensors. It splices and fuses multiple input tensors along the specified axis (usually the feature axis) to generate a new output tensor. In the recall model, the Concat layer connects the user configuration vector and the above-mentioned attention representation vector, thereby fusing the user configuration vector and the attention representation vector together and splicing them into a richer feature vector to better represent and capture different feature patterns, thereby enhancing the expressiveness of the recall model.

[0092] Multilayer Perceptron (MLP) is a commonly used neural network structure consisting of multiple fully connected layers. Each neuron is connected to all neurons in the previous layer, and information transmission and nonlinear transformation are achieved through weights and activation functions. In the recall model, the MLP layer can be used to process input features to learn potential feature vectors representing users and samples. These potential feature vectors can be used to calculate the user's preference or similarity to the sample. In the recommendation system, the goal of the recall model is to quickly select a small number of candidate items from a large number of samples so that the subsequent sorting model can further optimize the recommendation results. In this embodiment, different features can be combined through the MLP layer to produce a higher-order feature representation, thereby improving the recall model's ability to model complex interactions between users and samples.

[0093] The L2 normalization layer is a commonly used neural network layer, which is used to normalize the input data so that the L2 norm (i.e., Euclidean distance) of the feature vector of each sample is 1. It can help speed up the training process, improve the generalization ability of the model, and prevent the problem of gradient disappearance. In the recall model, the L2 normalization layer can be used to normalize the feature vector of the sample. By L2 normalizing the input feature vector, the scale difference between different features will not have a significant impact on the similarity measurement, thereby better measuring the similarity between different features. In this embodiment, the degree of association between different feature vectors can be measured by calculating the cosine similarity between them. The feature fusion vector (output of the MLP layer) and the target feature vector of the user configuration and user behavior are L2 normalized respectively, and the modulus length of the vector can be scaled to the same size, that is, converted into a vector on the unit sphere. In this way, the role of each vector in the distance measurement is equal, so that different vectors have higher comparability, and vectors are easier to compare and match, thereby more accurately measuring the similarity between different features, helping the system to better perform recall recommendations.

[0094] In this embodiment, the first main network first embeds the user configuration features, user behavior features and target behavior features through the embedding layer to obtain the corresponding feature vectors. Then, on the one hand, the feature vectors corresponding to the multiple user behavior features are processed through the multi-head self-attention layer to obtain a more integrated attention feature vector, and the feature vector corresponding to the user configuration feature and the attention feature vector are connected, and the connected feature vector is output, and then the connected feature vector is processed through the MLP layer to obtain a higher-order feature vector, and the higher-order feature vector is L2 normalized to obtain the L2 normalized vector corresponding to the user feature. On the other hand, the target behavior feature is processed by the MLP layer and L2 normalized in turn to obtain the L2 normalized vector corresponding to the target feature. Then the two L2 normalized vectors are convolved to obtain the output result of the first main network. The first bias network embeds the debiased feature through the embedding layer to obtain the feature vector corresponding to the bias feature, and then the feature vector corresponding to the bias feature is processed by one or more MLP layers to obtain the bias parameter corresponding to the bias feature. Finally, the softmax function can be used in batch processing to convert the output result of the first main network and the bias parameter into a probability distribution, so that the recall model can efficiently screen large-scale candidate sets to obtain items related to user queries or needs.

[0095] It can be understood that the structure of the above-mentioned first main network and the structure of the first bias network are only one implementation of this embodiment, and this embodiment is not limited to this. As long as the first main network can realize the function of sample recall, the first bias network can eliminate or minimize the influence of bias features on sample selection.

[0096] Furthermore, the sorting layer model includes a second main network and a second bias network. Then the above step of using the training data set to train the recommendation model also includes the following processing: based on the predicted output value, determine a new training data set; input the new user feature data and new recommended content included in each sample of the new training data set into the second main network respectively, and obtain the relationship between the new recommended content and the new user feature; input the bias features in each sample in the new training data set into the second bias network respectively, and obtain the bias parameters of each new sample data; based on the relationship between the new recommended content and the new user feature, and the bias parameters of each new sample data, determine a new predicted output value; the new predicted output value is used to indicate the degree of influence of the bias feature on whether the sample is clicked.

[0097] The main network (Main Net) of the sorting layer model, that is, the second main network, is the main structure of the sorting layer model, which can be any neural network model with a sorting function. The present embodiment does not specifically limit its specific structure, such as a point-to-point sorting model, a queue-by-queue sorting model, an overall queuing model, and a learning queuing model. The bias network of the sorting layer model, that is, the second bias network, is also used to process the bias features and extract more comprehensive information as much as possible to improve the fit of the model, that is, to connect all the debiasing features, and then pass through several layers of full connection to finally obtain a bias parameter. The bias parameter can also be used to indicate the degree of influence of the bias feature on whether the sample is clicked. The present embodiment does not specifically limit the specific structure of the second bias network, which can be any neural network model with a debiasing function.

[0098] In the process of training the recall layer model, the following loss function (1) can be used to optimize the recall model. After each training cycle, the loss value of the training cycle is calculated, and when the loss value reaches convergence, the training of the recall layer model is terminated.

[0099]

[0100] Among them, B represents the number of training samples; EXP(sim(x,y)) represents the user's interest in the content corresponding to the feature item; logits bias Represents the bias parameter of the first bias network output; T represents the temperature coefficient, which is a hyperparameter and can be customized; x represents the training feature, y represents the training sample label (for example, the label may include "0" and "1", 1 represents that the content in the sample is clicked after exposure, and 0 represents that the content in the sample is not clicked after exposure); M represents the set of positive samples and negative samples.

[0101] In some other embodiments, the structure of the sorting layer model can be as follows: Figure 4As shown, the second main network includes a feature embedding layer, a multi-head self-attention layer, an attention mechanism of a deep interest network, a connection layer, and a multi-layer perceptron layer, wherein the specific operations and functions of the feature embedding layer, the multi-head self-attention layer, the connection layer, and the multi-layer perceptron layer are similar to the principles and functions of the recall layer model. The specific operations can refer to the execution process of each layer in the first main network, as long as the initial input features are replaced with the sample content to be sorted. In this embodiment, the attention mechanism of the Deep Interest Network (DIN) is introduced. In the sorting stage, user interest is used as an important feature dimension, and the feature items that are more similar to the user's interest will be ranked in the front position. Specifically, the deep interest network uses the Attention mechanism to capture the correlation between the target feature item and the feature item in the user behavior sequence, thereby realizing interest modeling of specific target feature items and applying it to the sorting stage. The second bias network refers to the structure of the first bias network and may also include an embedding layer and two multi-layer perceptron layers. The debiased feature is first embedded through the embedding layer to obtain a feature vector corresponding to the bias feature, and then the feature vector corresponding to the bias feature is processed by one or more MLP layers to obtain a bias parameter corresponding to the bias feature.

[0102] In this embodiment, the second main network of the sorting layer model can also first embed the user configuration features, user behavior features and target behavior features through the embedding layer to obtain the corresponding feature vectors. Then, on the one hand, the feature vectors corresponding to multiple user behavior features are processed through the multi-head self-attention layer to obtain a more integrated attention feature vector. And the correlation between the feature vector corresponding to the target behavior feature and the attention feature vector is captured through the attention mechanism of the deep interest network, and the user behavior feature vectors are sorted according to the degree of correlation, and the sorted output results are connected with the feature vectors corresponding to the user configuration features and the feature vectors corresponding to the target behavior features, and the connected feature vectors are output, and then the connected feature vectors are processed through the MLP layer to obtain a higher-order feature vector, that is, the output result of the second main network, which can also be called the main parameter. The second bias network can also embed the debiasing feature through the embedding layer to obtain the feature vector corresponding to the bias feature, and then the feature vector corresponding to the bias feature is processed by one or more MLP layers to obtain the bias parameter corresponding to the bias feature. Finally, the main parameters and bias parameters can be connected and fused to obtain the final parameter value, which can be called final logits. This parameter value can be a probability value, which is used to indicate the degree of influence of objective factors on whether users click on recommended content. Its value can be 0-1, where 0 indicates that the objective factor represented by the corresponding feature has no influence on whether users click on recommended content, and 1 indicates that the objective factor represented by the corresponding feature has a direct influence on whether users click on recommended content. When this objective factor appears, users may or may not click on the recommended content. The sorting layer model calculates the parameter value of each sample and then sorts them according to the parameter value.

[0103] During the training process of the above-mentioned sorting layer model, the following loss function (2) can be used to optimize the sorting layer model. After each training cycle, the loss value of the training cycle is calculated, and the training of the sorting layer model is terminated when the loss value reaches convergence.

[0104]

[0105] Among them, B represents the number of training samples; logits bias represents the bias parameter output by the second bias network; x represents the training feature, y represents the training sample label (for example, the label may include "0" and "1", 1 represents that the content in the sample is clicked after being exposed, and 0 represents that the content in the sample is not clicked after being exposed); H(x i ) represents the result predicted by the main network; log(H(x i )+logits bias Represents the entropy of the predicted value when the sample is positive; log(1-H(x i)-logits bias Represents the entropy of the predicted value when a negative sample is encountered.

[0106] It can be understood that the structure of the above-mentioned second main network and the structure of the second bias network are only one implementation of this embodiment, and this embodiment is not limited to this. As long as the second main network can realize the function of sample recall, the second bias network can eliminate or minimize the influence of bias features on sample selection.

[0107] Step S3: based on the personalized data sets in each content recommendation scenario, respectively adjust the basic recommendation model to obtain a personalized recommendation model for each content recommendation scenario.

[0108] Among them, personalized data sets can be understood as data sets composed of sample content in the same recommendation scenario. For example, for a recommendation system applied to an e-commerce platform, in the scenario of inputting keywords to search for products, the data of different users in this scenario (including user configuration features, user behavior features, and target behavior features) can be used as a personalized data set in this scenario.

[0109] In practical applications, after training the recall layer model including the first main network and the first bias network, and the sorting layer model including the second main network and the second bias network, an unbiased and reliable basic recommendation model including the recall layer model and the sorting layer model can be obtained. Then, in order to obtain a specific recommendation model applied to a specific application scenario, the personalized data set under each content recommendation scenario can be used to fine-tune the basic recommendation model to obtain multiple different personalized recommendation models, which are respectively suitable for their corresponding content recommendation scenarios. In this way, by using data from different scenarios to fine-tune the basic recommendation model, personalized models for different scenarios can be quickly obtained, thereby improving the scalability of the model.

[0110] In some embodiments, Figure 5 As shown, step S3 may include the following processing steps: step S31, deleting the bias network in the basic recommendation model to obtain a new basic recommendation model; step S32, based on the personalized data sets in each scenario, fine-tuning the new basic recommendation model to obtain the personalized recommendation models corresponding to each scenario.

[0111] In step S2 of the embodiment of the present application, an unbiased and reliable basic recommendation model is trained. In the implementation process of this step S3, the bias network in the basic recommendation model is first deleted to obtain a new basic recommendation model, and then the data of personalized scenarios are used to perform fine-tuning training in the model without the bias net. The user mind in the personalized scenario can be added to the new basic recommendation model, thereby improving the recommendation effect of the recommendation system in the personalized scenario.

[0112] Specifically, during the fine-tuning training of the new basic recommendation model, the loss function can also be used to optimize the new basic model. The loss function used here can refer to the above formulas (1) and (2), and the parts related to the bias network in formulas (1) and (2) can be removed.

[0113] In summary, the recommendation model training method provided in this embodiment first samples the global samples based on the preset exposure growth rate to obtain a training data set; then uses the training data set to train the recommendation model to obtain a basic recommendation model, and then uses the personalized data set in each content recommendation scenario to fine-tune the basic recommendation model to obtain the personalized recommendation model of each content recommendation scenario. In this way, by selecting the training data set based on the preset exposure growth rate, more samples with smaller exposure numbers can be collected, the diversity of samples in the training data set is improved, the dominance of the head sample data is avoided, the degree of influence of the trained basic recommendation model on the samples with larger exposure numbers is reduced, and the unbiasedness of the basic recommendation model is improved. In addition, the recall layer model and the ranking layer model in the basic recommendation model both include the main network and the bias network. Based on the debiasing effect of the bias network, the bias problem caused by different exposure channels and different exposure positions of the global samples can be solved. Through the above two methods, an unbiased and reliable basic recommendation model can be trained when using global sample training. Then, the personalized data set in each content recommendation scenario is used to fine-tune the basic recommendation model, and the personalized recommendation model of each content recommendation scenario can be quickly obtained, thereby improving the scalability of the model.

[0114] Some other embodiments of the present application also provide an information recommendation method, such as Figure 6 As shown, the method includes the following steps.

[0115] Step S10: obtaining user characteristic data of a target user in a target content recommendation scenario.

[0116] Among them, the target user can be understood as the user who currently applies the method to recommend information, the target content recommendation scenario can be understood as the content recommendation scenario applicable to the target user, and the user feature data includes but is not limited to the configuration features and behavior features of the target user.

[0117] Step S20: input the user feature data into a personalized recommendation model corresponding to the target content recommendation scenario, and output the content to be recommended.

[0118] The personalized recommendation model is obtained by training the recommendation model training method described above. It is understandable that each embodiment of the recommendation model training method described above can be applied to the embodiment of the information recommendation method to achieve the same technical effect, and will not be described in detail here.

[0119] In some embodiments, step S20 may include the following processing: if a target personalized recommendation model for a target content recommendation scenario has been trained, user feature data is input into the target personalized recommendation model, and the content to be recommended is output.

[0120] In this embodiment, after obtaining the user feature data of the target user in the target content recommendation scenario, it is possible to first determine whether the corresponding personalized recommendation model has been trained, that is, whether the current recommendation system includes the target personalized recommendation model for the target content recommendation scenario, and if the target personalized recommendation model for the target content recommendation scenario has been trained, the user feature data is input into the target personalized recommendation model, and the content to be recommended is output, thereby obtaining more accurate content to be recommended and achieving better model recommendation effects.

[0121] In other embodiments, step S20 may also include the following processing: if the target personalized recommendation model for the target content recommendation scenario has not been trained yet, the user feature data is respectively input into the personalized recommendation models of each content recommendation scenario that has been trained, to obtain multiple groups of candidate content output by the personalized recommendation models of each content recommendation scenario; and the content to be recommended is screened out from the multiple groups of candidate content.

[0122] In actual applications, there may be a cold start of a new scenario. In this case, there may be no content recommendation scenario applicable to the target user. Based on all the trained personalized recommendation models, one or more different recommended contents can be obtained respectively. Then, the recommended contents with higher relevance to the target user can be selected from these recommended contents, that is, multiple recommended contents with higher parameter values ​​output by the sorting layer model, for content recommendation. This can achieve good recommendation effects when the personalized scenario is cold started.

[0123] Furthermore, the information recommendation method may also include the following processing: when a target personalized data set of a target content recommendation scenario is obtained, the basic recommendation model is adjusted based on the target personalized data set to obtain a target personalized recommendation model of the target content recommendation scenario.

[0124] In this embodiment, after the cold start of the scene, the target personalized data set of the new scene can be obtained, and then the basic recommendation model can be fine-tuned based on the obtained target personalized data set to obtain the target personalized recommendation model corresponding to the target content recommendation scene. Thereafter, when information recommendation is performed based on the target content recommendation scene, information recommendation can be performed based on the newly obtained target personalized recommendation model to further improve the recommendation accuracy and recommendation effect in the target content recommendation scene.

[0125] Step S30: sending the content to be recommended to the target user.

[0126] The information recommendation method provided in this embodiment adopts the recommendation model trained by the above-mentioned recommendation model training method to determine the recommended content. When making content recommendations, the target content recommendation scenario is determined based on the user feature data of the target user, and the recommended content is determined based on the personalized recommendation model under the target content recommendation scenario. This can obtain more accurate content to be recommended and achieve better model recommendation effects.

[0127] Some embodiments of the present application further provide a recommendation model training device, which is used to execute the recommendation model training method provided in any of the above embodiments. Figure 7 A schematic diagram of the recommendation model training device is shown in FIG. Figure 7 As shown, the recommendation model training device includes:

[0128] A training data acquisition module is used to sample global samples based on a preset exposure growth rate to obtain a training data set; the global samples include samples in various content recommendation scenarios;

[0129] The basic model training module is used to train the recommendation model using the training data set to obtain a basic recommendation model suitable for various content recommendation scenarios; the recall layer model and the ranking layer model in the recommendation model both include a main network and a bias network;

[0130] The personalized model training module is used to adjust the basic recommendation model based on the personalized data set in each content recommendation scenario to obtain the personalized recommendation model for each content recommendation scenario.

[0131] It can be understood that the recommendation model training device provided in this embodiment and the recommendation model training method provided in the embodiment of the present application are based on the same inventive concept, and can at least achieve the same beneficial effects of the recommendation model training method, and the various implementation methods of the recommendation model training method embodiment are also applicable to the embodiment of the recommendation model training device, which will not be repeated here.

[0132] Some embodiments of the present application further provide an information recommendation device, which is used to execute the information recommendation method provided in any of the above embodiments. Figure 8 A schematic diagram of the information recommendation device is shown in FIG. Figure 8 As shown, the device comprises:

[0133] A data acquisition module is used to acquire user characteristic data of target users in a target content recommendation scenario;

[0134] A content determination module, used to input user feature data into a recommendation model corresponding to a target content recommendation scenario, and output the content to be recommended; the recommendation model is trained by the recommendation model training method of any of the above implementation modes;

[0135] The content sending module is used to send the content to be recommended to the target user.

[0136] It can be understood that the information recommendation device provided in this embodiment and the information recommendation method provided in the embodiment of the present application are based on the same inventive concept, and can at least achieve the same beneficial effects of the information recommendation method, and the various implementation methods of the information recommendation method embodiment are also applicable to the embodiment of the information recommendation device, which will not be repeated here.

[0137] It should be noted that the data involved in this application (including but not limited to data used for model training, stored data, displayed data, etc.) are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0138] The present application also provides an electronic device to perform the above-mentioned recommendation model training method or information recommendation method. Fig. 9 , which shows a schematic diagram of an electronic device provided by some embodiments of the present application. Fig. 9 As shown, the electronic device 4 includes: a processor 400, a memory 401, a bus 402 and a communication interface 403, and the processor 400, the communication interface 403 and the memory 401 are connected via the bus 402; the memory 401 stores a computer program that can be run on the processor 400, and when the processor 400 runs the computer program, it executes the recommendation model training method or information recommendation method provided in any of the aforementioned embodiments of the present application.

[0139] The memory 401 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk storage. The communication connection between the device network element and at least one other network element is realized through at least one communication interface 403 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0140] The bus 402 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. Among them, the memory 401 is used to store programs, and the processor 400 executes the program after receiving the execution instruction. The recommendation model training method or information recommendation method disclosed in any implementation of the aforementioned embodiment of the present application can be applied to the processor 400, or implemented by the processor 400.

[0141] The processor 400 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 400. The above processor 400 can be a general-purpose processor, including a processor (Central Processing Unit, referred to as CPU), a network processor (Network Processor, referred to as NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a readily available programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application can be directly embodied as a hardware decoding processor to be executed, or the hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 401, and the processor 400 reads the information in the memory 401 and completes the steps of the above method in combination with its hardware.

[0142] The electronic device provided in the embodiment of the present application and the recommendation model training method or information recommendation method provided in the embodiment of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented therein.

[0143] The present application also provides a computer-readable storage medium corresponding to the recommendation model training method or information recommendation method provided in the above-mentioned embodiment. Fig.10 The computer-readable storage medium shown is a CD 50 on which a computer program (i.e., a program product) is stored. When the computer program is run by the processor, it will execute the recommendation model training method or information recommendation method provided by any of the aforementioned embodiments.

[0144] It should be noted that examples of computer-readable storage media may also include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical or magnetic storage media, which are not listed here one by one.

[0145] An embodiment of the present application also provides a computer program product, including a computer program, which is executed by a processor to implement the recommendation model training method or information recommendation method of any of the above embodiments.

[0146] The computer-readable storage medium and computer program product provided in the above-mentioned embodiments of the present application are based on the same inventive concept as the recommendation model training method or information recommendation method provided in the embodiments of the present application, and have the same beneficial effects as the methods adopted, run or implemented by the application programs stored therein.

[0147] It should be noted that:

[0148] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known structures and technologies are not shown in detail so as not to obscure the understanding of this description.

[0149] Similarly, it should be understood that in order to streamline the present application and aid in understanding one or more of the various inventive aspects, in the above description of the exemplary embodiments of the present application, the various features of the present application are sometimes grouped together into a single embodiment, figure, or description thereof. However, the disclosed method should not be interpreted as reflecting the following schematic diagram: the claimed application requires more features than the features explicitly recited in the individual claims. More specifically, as reflected in the claims below, the inventive aspects are less than all the features of the individual embodiments disclosed above. Therefore, the claims that follow the specific embodiment are hereby expressly incorporated into the specific embodiment, wherein each claim itself serves as a separate embodiment of the present application.

[0150] In addition, those skilled in the art will appreciate that, although some embodiments described herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present application and form different embodiments. For example, in the claims below, any one of the claimed embodiments may be used in any combination.

[0151] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A recommendation model training method, characterized in that: include: Sampling the entire domain samples based on a preset exposure growth rate to obtain a training data set; The global samples include samples in various content recommendation scenarios; The training data set is used to train a recommendation model to obtain a basic recommendation model suitable for each content recommendation scenario; the recall layer model and the ranking layer model in the recommendation model both include a main network and a bias network; Based on the personalized data sets in each content recommendation scenario, the basic recommendation models are adjusted respectively to obtain personalized recommendation models for each content recommendation scenario.

2. The method according to claim 1, characterized in that The sampling of the global samples based on the preset exposure growth rate includes: Determine the exposure number of each sample in the global sample respectively; Based on a preset exposure growth rate and the exposure number of each sample, multiple samples are selected from the global samples; the exposure numbers of the multiple samples are increased according to the preset exposure growth rate.

3. The method according to claim 2, characterized in that The selecting a plurality of samples from the global samples based on a preset exposure growth rate and the exposure number of each sample includes: Based on the exposure number of each sample, samples with the same exposure number are grouped into sets to obtain multiple sample sets; Sorting the multiple sample sets based on the exposure number corresponding to each sample set, and selecting a preset number of sample sets with the highest exposure numbers from the sorted sample sets; Selecting, from the preset number of sample sets, a plurality of target sample sets whose exposure numbers increase according to the preset exposure growth rate; Based on the exposure number corresponding to each target sample set, samples are selected from each target sample set.

4. The method according to claim 1, characterized in that: The recall layer model includes a first main network and a first bias network; the training of the recommendation model using the training data set includes: Inputting the user feature data and the recommended content included in each sample in the training data set into the first main network respectively to obtain the relationship between the recommended content and the user feature; Inputting the bias features in each sample into the first bias network respectively to obtain bias parameters of each sample data; Based on the relationship between the recommended content and the user characteristics, and the bias parameters of each sample data, a predicted output value is determined; the predicted output value is used to indicate the degree of influence of the bias feature on whether the sample is clicked.

5. The method according to claim 4, characterized in that The sorting layer model includes a second main network and a second bias network; the training of the recommendation model using the training data set also includes: Based on the predicted output value, determining a new training data set; Inputting new user feature data and new recommended content included in each sample of the new training data set into the second main network respectively, to obtain the relationship between the new recommended content and the new user feature; Inputting the bias features in each sample in the new training data set into the second bias network respectively to obtain the bias parameters of each new sample data; Based on the relationship between the new recommended content and the new user feature, and the bias parameter of each new sample data, a new predicted output value is determined; the new predicted output value is used to indicate the degree of influence of the bias feature on whether the sample is clicked.

6. The method according to claim 1, characterized in that The adjusting the basic recommendation models based on the personalized data sets in the various content recommendation scenarios to obtain the personalized recommendation models for the various content recommendation scenarios respectively includes: Deleting the bias network in the basic recommendation model to obtain a new basic recommendation model; Based on the personalized data sets in each scenario, the new basic recommendation model is fine-tuned to obtain the personalized recommendation models corresponding to each scenario.

7. An information recommendation method, characterized in that: The method comprises: Obtain user feature data of target users in the target content recommendation scenario; Inputting the user feature data into a personalized recommendation model corresponding to the target content recommendation scenario, and outputting the content to be recommended; the personalized recommendation model is trained by the recommendation model training method according to any one of claims 1 to 6; The content to be recommended is sent to the target user.

8. The method according to claim 7, characterized in that The step of inputting the user feature data into a personalized recommendation model corresponding to the target content recommendation scenario and outputting content to be recommended includes: If a target personalized recommendation model for the target content recommendation scenario has been trained, the user feature data is input into the target personalized recommendation model, and the content to be recommended is output.

9. The method according to claim 7 or 8, characterized in that: The step of inputting the user feature data into a personalized recommendation model corresponding to the target content recommendation scenario and outputting content to be recommended includes: If the target personalized recommendation model for the target content recommendation scenario has not been trained yet, the user feature data is respectively input into the personalized recommendation models for each content recommendation scenario that has been trained, to obtain multiple groups of candidate content output by the personalized recommendation models for each content recommendation scenario; The content to be recommended is screened out from the multiple groups of candidate content.

10. The method according to claim 9, characterized in that The method further comprises: When a target personalized data set of the target content recommendation scenario is obtained, a basic recommendation model is adjusted based on the target personalized data set to obtain a target personalized recommendation model of the target content recommendation scenario.

11. A recommendation model training device, characterized in that: include: A training data acquisition module, used to sample the global samples based on a preset exposure growth rate to obtain a training data set; The global samples include samples in various content recommendation scenarios; A basic model training module, used to train a recommendation model using the training data set to obtain a basic recommendation model suitable for each content recommendation scenario; the recall layer model and the ranking layer model in the recommendation model both include a main network and a bias network; The personalized model training module is used to adjust the basic recommendation model based on the personalized data sets in each content recommendation scenario to obtain a personalized recommendation model for each content recommendation scenario.

12. An information recommendation device, characterized in that: The device comprises: A data acquisition module is used to acquire user characteristic data of target users in a target content recommendation scenario; A content determination module, configured to input the user feature data into a personalized recommendation model corresponding to the target content recommendation scenario, and output the content to be recommended; the personalized recommendation model is trained by the recommendation model training method according to any one of claims 1 to 6; The content sending module is used to send the content to be recommended to the target user.

13. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method according to any one of claims 1 to 10.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the method according to any one of claims 1 to 10.

Images