Information recommendation method and device, storage medium and computer device

Abstract

The application discloses an information recommendation method and device, a storage medium and computer equipment. The method comprises the following steps: in response to an information recommendation request, a feature vector of recommendation information is obtained, wherein the recommendation information comprises user information, recommendation content and context information; based on the feature vector of the recommendation information, a plurality of expert vectors are obtained through an expert network set corresponding to each attribute, wherein the expert vector is obtained by fusing a sub-expert vector output by each sub-network in the expert network set, and each sub-network corresponds to an attribute value of an attribute; the plurality of expert vectors are connected through a gating network to obtain a fusion vector, the fusion vector is predicted through a tower network to obtain a prediction result of the recommendation information; based on the prediction result of the recommendation information, it is judged whether the recommendation content meets a preset recommendation condition, and if so, the recommendation content is pushed to a client where the user information is located. The above method can effectively improve the accuracy of content recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and in particular to an information recommendation method and device, a storage medium and a computer device. BACKGROUND

[0002] With the continuous development of Internet technology, the needs of users are becoming more and more diverse and personalized. At present, it has become a relatively common means to recommend commodity information to users through a personalized recommendation system and thereby improve the click rate and conversion rate of commodity information. However, in the personalized recommendation system, the distribution of user behavior history data is very different for different user types, different time periods, different product forms and different channels. This difference in data has a great influence on the training of the model.

[0003] At present, traditional single deep learning models cannot accurately depict the differences in user behavior data, and mutual interference occurs between different situations, resulting in inaccurate recommendation results. If a model is established and optimized for different situations, the amount of training data available for each single model will be insufficient, which will prevent the model from being fully learned. In addition, in a complex recommendation system, the amount of engineering development required to establish a model for each situation is huge, and the time and labor costs are high. More importantly, business data has a certain degree of similarity between different situations, but this similarity cannot be fully exploited under the existing method, resulting in low accuracy of information recommendation. SUMMARY

[0004] Therefore, the present application provides an information recommendation method and device, a storage medium and a computer device, which mainly aims to solve the technical problem of low accuracy of information recommendation in a recommendation system.

[0005] According to a first aspect of the present application, an information recommendation method is provided, which comprises:

[0006] In response to an information recommendation request, a feature vector of recommended information is obtained, wherein the recommended information comprises user information, recommended content and context information, and the recommended information further comprises a plurality of attributes and attribute values of each attribute;

[0007] Based on the feature vector of the recommended information, a plurality of expert vectors are obtained through a set of expert networks corresponding to each attribute, wherein each expert vector is obtained by fusing a sub-expert vector output by each sub-network in the set of expert networks, and each sub-network corresponds to an attribute value of the attribute;

[0008] The plurality of expert vectors are connected by a gating network to obtain a fusion vector, and the fusion vector is predicted by a tower network to obtain a prediction result of the recommendation information.

[0009] Based on the prediction result of the recommendation information, it is judged whether the recommended content meets the preset recommendation condition. If it meets, the recommended content is pushed to the client where the user information is located.

[0010] Optionally, in response to the information recommendation request, the feature vector of the recommendation information is obtained, including: in response to the information recommendation request, obtaining user information, recommended content and context information; determining the attributes of the recommendation information and the attribute values of each attribute according to the user information, the recommended content and the context information, wherein the attributes include at least one of user gender, user age, user state, user behavior, product category, time, place and channel; performing feature processing on the user information, the recommended content and the context information, and connecting the attribute values of each attribute in the user information, the recommended content, the context information and the recommendation information after the feature processing to obtain the feature vector of the recommendation information.

[0011] Optionally, based on the feature vector of the recommendation information, a plurality of expert vectors are obtained through a set of expert networks corresponding to each attribute, including: determining a main sub-network and an auxiliary sub-network corresponding to the recommendation information in each expert network set according to the attribute values of each attribute in the recommendation information, wherein the attribute value corresponding to the main sub-network is the same as the attribute value of the recommendation information, and the attribute value corresponding to the auxiliary sub-network is different from the attribute value of the recommendation information; mapping the feature vector of the recommendation information into a main expert vector through the main sub-network, and mapping the feature vector of the recommendation information into an auxiliary expert vector through the auxiliary sub-network; for each expert network set, aggregating the auxiliary expert vectors, and performing weighted fusion on the aggregated auxiliary expert vectors and the main expert vector to obtain an expert vector.

[0012] Optionally, for each expert network set, the auxiliary expert vectors are aggregated, and the aggregated auxiliary expert vectors and the main expert vector are weighted and fused to obtain an expert vector, including: for each expert network set, obtaining the correlation between each auxiliary sub-network and the main sub-network; determining the weight of each auxiliary expert vector according to the correlation between each auxiliary sub-network and the main sub-network; aggregating the auxiliary expert vectors according to the weight of each auxiliary expert vector to obtain an aggregated auxiliary expert vector; and performing weighted fusion on the main expert vector and the aggregated auxiliary expert vector through an update gate network to obtain an expert vector.

[0013] Optionally, before the step of obtaining the fusion vector by weighting connecting the plurality of expert vectors through the gating network, the method further comprises: mapping the feature vector of the recommendation information into a general expert vector through a general expert network, wherein the general expert network corresponds to the overall representation of the recommendation information; and the step of obtaining the fusion vector by weighting connecting the plurality of expert vectors through the gating network comprises: weighting connecting the expert vector output by each of the expert network sets and the general expert vector through the gating network to obtain the fusion vector.

[0014] Optionally, before the step of obtaining the fusion vector by weighting connecting the expert vector output by each of the expert network sets and the general expert vector through the gating network, the method further comprises: determining a target vector dimension of the general expert vector according to a vector dimension of each of the expert vectors and a preset feature proportion of the general expert vector; and expanding the vector dimension of the general expert vector to the target vector dimension, wherein the target vector dimension is greater than the vector dimension of each of the expert vectors.

[0015] Optionally, the number of the gating networks and the tower networks is one or more, and each of the gating networks and each of the tower networks corresponds to a prediction task; and the step of obtaining the fusion vector by weighting connecting the plurality of expert vectors through the gating network, and obtaining the prediction result of the recommendation information by predicting the fusion vector through the tower network comprises: for each of the prediction tasks, obtaining the weight of each of the expert network sets through the gating network corresponding to the prediction task, and weighting connecting the expert vector output by each of the expert network sets according to the weight of each of the expert network sets to obtain the fusion vector of the prediction task; for each of the prediction tasks, predicting the fusion vector of the prediction task through the tower network corresponding to the prediction task to obtain the prediction result of the prediction task; and obtaining the prediction result of the recommendation information according to the prediction result of each of the prediction tasks.

[0016] Optionally, the step of determining whether the recommendation content meets the preset recommendation condition based on the prediction result of the recommendation information, and if so, pushing the recommendation content to the client where the user information is located comprises: determining a recommendation threshold of each of the prediction tasks according to the recommendation condition; obtaining the prediction result of each of the prediction tasks, wherein the prediction task comprises at least one of a click rate prediction task, a conversion rate prediction task and a preference degree prediction task; for each of the prediction tasks, comparing the prediction result of the prediction task with the recommendation threshold of the prediction task; and if the prediction result of each of the prediction tasks is greater than the recommendation threshold of the prediction task, pushing the recommendation content to the client where the user information is located.

[0017] Optionally, the information recommendation method is applied to a pre-trained general recommendation model, the general recommendation model comprising a plurality of expert network sets, at least one gating network, and at least one tower network, wherein each of the expert network sets corresponds to an attribute of the recommendation information, the expert network set comprising a plurality of sub-networks and a mask attention mechanism, the sub-networks being configured to output a sub-expert vector corresponding to an attribute value of the attribute, and the mask attention mechanism being configured to fuse each of the sub-expert vectors into an expert vector; each of the gating networks corresponds to a prediction task and is configured to perform weighted connection on the expert vectors output by each of the expert network sets to obtain a fusion vector of the prediction task; and each of the tower systems corresponds to a prediction task and is configured to perform prediction on the fusion vector of the prediction task to obtain a prediction result of the prediction task.

[0018] Optionally, the mask attention mechanism comprises a Boolean mask, an attention operator, and an update gate network, wherein the Boolean mask is configured to generate an indication vector comprising elements corresponding to the plurality of sub-networks, wherein an element corresponding to a main sub-network is a first preset value, and an element corresponding to an auxiliary sub-network is a second preset value, the main sub-network corresponding to an attribute value identical to that of the recommendation information, and the auxiliary sub-network corresponding to an attribute value different from that of the recommendation information; the attention operator is configured to aggregate auxiliary expert vectors output by the auxiliary sub-networks; and the update gate network is configured to perform weighted fusion on the aggregated auxiliary expert vectors and a main expert vector output by the main sub-network to obtain an expert vector.

[0019] Optionally, the general recommendation model further comprises a general expert network, wherein the general expert network corresponds to an overall representation of the recommendation information and is configured to output a general expert vector corresponding to the overall representation of the recommendation information; and the gating network is further configured to perform weighted connection on the expert vectors output by each of the expert network sets and the general expert vector to obtain the fusion vector of the prediction task.

[0020] According to a second aspect of the present application, there is provided an information recommendation device, comprising:

[0021] an information acquisition module configured to acquire a feature vector of recommendation information in response to an information recommendation request, wherein the recommendation information comprises user information, recommendation content, and context information, and further comprises a plurality of attributes and attribute values of each of the attributes;

[0022] The information processing module is configured to obtain a plurality of expert vectors based on a feature vector of the recommendation information through an expert network set corresponding to each attribute, wherein each expert vector is obtained by fusing a sub-expert vector output by each sub-network in the expert network set, and each sub-network corresponds to an attribute value of the attribute.

[0023] The information prediction module is configured to obtain a fusion vector by performing weighted connection on the plurality of expert vectors through a gating network, and obtain a prediction result of the recommendation information by performing prediction on the fusion vector through a tower network.

[0024] The information pushing module is configured to determine whether the recommendation content meets a preset recommendation condition based on the prediction result of the recommendation information, and push the recommendation content to a client where the user information is located if the recommendation content meets the preset recommendation condition.

[0025] Optionally, the information acquisition module is specifically configured to acquire user information, recommendation content and context information in response to an information recommendation request, determine an attribute of recommendation information and an attribute value of each attribute according to the user information, the recommendation content and the context information, wherein the attribute includes at least one of user gender, user age, user state, user behavior, commodity category, time, place and channel, perform feature processing on the user information, the recommendation content and the context information, and connect the user information, the recommendation content, the context information and the attribute value of each attribute in the recommendation information after the feature processing to obtain a feature vector of the recommendation information.

[0026] Optionally, the information processing module is specifically configured to determine a main sub-network and an auxiliary sub-network corresponding to the recommendation information in each expert network set according to the attribute value of each attribute in the recommendation information, wherein the attribute value corresponding to the main sub-network is the same as the attribute value of the recommendation information, and the attribute value corresponding to the auxiliary sub-network is different from the attribute value of the recommendation information, map the feature vector of the recommendation information into a main expert vector through the main sub-network, map the feature vector of the recommendation information into an auxiliary expert vector through the auxiliary sub-network, aggregate the auxiliary expert vectors for each expert network set, and perform weighted fusion on the aggregated auxiliary expert vectors and the main expert vector to obtain an expert vector.

[0027] Optionally, the information processing module is further configured to: for each of the expert network set, obtain a correlation of each of the auxiliary sub-networks with the main sub-network; determine a weight of each of the auxiliary expert vectors according to the correlation of each of the auxiliary sub-networks with the main sub-network; aggregate the auxiliary expert vectors according to the weight of each of the auxiliary expert vectors to obtain an aggregated auxiliary expert vector; and obtain an expert vector by performing weighted fusion on the main expert vector and the aggregated auxiliary expert vector through the update gate network.

[0028] Optionally, the information processing module is further configured to map the feature vector of the recommendation information into a general expert vector through a general expert network, where the general expert network corresponds to the overall representation of the recommendation information; and the information prediction module is further configured to perform weighted connection on the expert vector output by each of the expert network set and the general expert vector through the gate network to obtain the fusion vector.

[0029] Optionally, the information processing module is further configured to determine a target vector dimension of the general expert vector according to a vector dimension of each of the expert vectors and a preset feature proportion of the general expert vector; and expand the vector dimension of the general expert vector to the target vector dimension, where the target vector dimension is greater than the vector dimension of each of the expert vectors.

[0030] Optionally, the number of the gate networks and the tower networks is one or more, and each of the gate networks and the tower networks corresponds to a prediction task; and the information prediction module is specifically configured to: for each of the prediction tasks, obtain a weight of each of the expert network set through the gate network corresponding to the prediction task, and perform weighted connection on the expert vector output by each of the expert network set according to the weight of each of the expert network set to obtain a fusion vector of the prediction task; for each of the prediction tasks, perform prediction on the fusion vector of the prediction task through the tower network corresponding to the prediction task to obtain a prediction result of the prediction task; and obtain a prediction result of the recommendation information according to the prediction result of each of the prediction tasks.

[0031] Optionally, the information pushing module is specifically configured to: determine a recommendation threshold of each of the prediction tasks according to the recommendation condition; obtain a prediction result of each of the prediction tasks, where the prediction task includes at least one of a click rate prediction task, a conversion rate prediction task and a preference degree prediction task; for each of the prediction tasks, compare the prediction result of the prediction task with the recommendation threshold of the prediction task; and if the prediction result of each of the prediction tasks is greater than the recommendation threshold of the prediction task, push the recommendation content to a client where the user information is located.

[0032] Optionally, the information recommendation method is applied to a pre-trained general recommendation model, the general recommendation model comprising a plurality of expert network sets, at least one gating network, and at least one tower network, wherein each of the expert network sets corresponds to an attribute of the recommendation information, the expert network set comprising a plurality of sub-networks and a mask attention mechanism, the sub-networks being configured to output a sub-expert vector corresponding to an attribute value of the attribute, and the mask attention mechanism being configured to fuse each of the sub-expert vectors into an expert vector; each of the gating networks corresponds to a prediction task, and is configured to perform weighted connection on the expert vectors output by each of the expert network sets to obtain a fusion vector of the prediction task; and each of the tower systems corresponds to a prediction task, and is configured to perform prediction on the fusion vector of the prediction task to obtain a prediction result of the prediction task.

[0033] Optionally, the mask attention mechanism comprises a Boolean mask, an attention operator, and an update gate network, wherein the Boolean mask is configured to generate an indication vector comprising elements corresponding to the plurality of sub-networks respectively, wherein an element corresponding to a main sub-network is a first preset value, and an element corresponding to an auxiliary sub-network is a second preset value, the attribute value corresponding to the main sub-network being the same as the attribute value of the recommendation information, and the attribute value corresponding to the auxiliary sub-network being different from the attribute value of the recommendation information; the attention operator is configured to aggregate auxiliary expert vectors output by the auxiliary sub-networks; and the update gate network is configured to perform weighted fusion on the aggregated auxiliary expert vectors and a main expert vector output by the main sub-network to obtain an expert vector.

[0034] Optionally, the general recommendation model further comprises a general expert network, wherein the general expert network corresponds to an overall representation of the recommendation information, and is configured to output a general expert vector corresponding to the overall representation of the recommendation information; and the gating network is further configured to perform weighted connection on the expert vectors output by each of the expert network sets and the general expert vector to obtain the fusion vector of the prediction task.

[0035] According to a third aspect of the present application, there is provided a storage medium having a computer program stored thereon, the program being executed by a processor to implement the information recommendation method.

[0036] According to a fourth aspect of the present application, there is provided a computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the information recommendation method when executing the program.

[0037] The information recommendation method, device, storage medium and computer equipment provided by the application can convert the feature vector of the recommendation information into the sub-specialist vector corresponding to different attribute values through multiple sub-networks, obtain the specialist vector corresponding to each attribute through multiple specialist network sets, and enable the recommendation information to be differentially learned and represented under different attributes and different attribute values. Meanwhile, the sub-specialist vector is fused to obtain the specialist vector in each specialist network set, the multiple specialist vectors are weighted and connected to obtain the fusion vector, and the recommendation information is predicted based on the fusion vector, so that it is determined whether to push the recommendation content to the user according to the prediction result, the commonality of each differential representation is mined, and the accuracy of the recommendation information prediction is improved. Based on this, the above method can effectively mine the difference and similarity between the recommendation information under different attributes and different attribute values, and compared with the traditional recommendation method, the above method can effectively improve the accuracy of content recommendation.

[0038] The above description is only a summary of the technical solutions of the application, in order to enable the technical means of the application to be implemented according to the content of the specification, and in order to enable the above and other purposes, characteristics and advantages of the application to be more obvious and easy to understand, the following specific embodiments of the application are described. BRIEF DESCRIPTION OF DRAWINGS

[0039] The accompanying drawings, which are included to provide a further understanding of the application, constitute a part of this application and illustrate embodiments of the application and explanations thereof, and do not constitute an improper limitation on the application. In the drawings:

[0040] Figure 1 A structure schematic diagram of a general recommendation model provided by an embodiment of the application is shown;

[0041] Figure 2 A flow schematic diagram of an information recommendation method provided by an embodiment of the application is shown;

[0042] Figure 3 A structure schematic diagram of another general recommendation model provided by an embodiment of the application is shown;

[0043] Figure 4 A flow schematic diagram of another information recommendation method provided by an embodiment of the application is shown;

[0044] Figure 5 A structure schematic diagram of an information recommendation device provided by an embodiment of the application is shown. DETAILED DESCRIPTION

[0045] The application will be described in detail below with reference to the drawings and in combination with the embodiments. It should be noted that the embodiments in the application and the characteristics in the embodiments can be combined with each other without conflict.

[0046] In the era of big data, personalized recommendation has become a ubiquitous service in various application scenarios. The purpose of personalized recommendation is to identify items (such as advertisements, videos, products, dishes, etc.) and select the most preferred recommendation content for users from a large pool of content. This problem is usually modeled as a classification problem and used to predict whether a user will click, purchase, or watch a particular recommended content.

[0047] Currently, the complexity of user behavior data is increasing, and many recommendation systems involve many factors of users and environments that significantly affect the prediction of user preferences by the recommendation system. On the one hand, users can be divided into different groups according to different criteria (such as gender, age, user status, etc.), and different groups have different preferences for recommended content. For example, men and women usually have different purchase habits when shopping online. On the other hand, the recommendation environment can also be divided into different scenarios according to context information (such as time, location, presentation style, access channel, etc.). For example, in dish recommendation, customers' preferences change significantly at different times of the day (such as breakfast, lunch, afternoon tea, etc.) and different locations. In addition, users also access the recommendation system through different channels (such as mobile applications, external links of third-party applications, message sharing links, etc.), which also leads to different preferences of users. The above factors can be referred to as attributes of recommendation information. When training samples with multiple different attributes through a single multi-layer neural network, the diversity of data produces a strong interference effect, resulting in inaccurate recommendation results.

[0048] To solve the problem of inaccurate recommendation by a single multi-layer neural network, the existing technology proposes a mixture-of-experts (MOE) learning framework. This learning model aims to use different expert networks to learn different parts of the input data differently. The outputs of all experts are aggregated through a gate network, and the weights of each expert network are determined. Finally, the prediction result of the prediction task is output through a tower network. However, most multi-expert learning networks use implicit expert networks and do not use explicit attributes such as population, time period, and scenario in prior knowledge, as well as explicit division of input space under each attribute (such as early, mid, and late under the time period angle). Therefore, the problem of data differentiation between different situations has not been fundamentally solved, and there is no explicit connection between the existing multi-expert and multi-angle optimization, lacking interpretability.

[0049] To solve the above problems, in one embodiment, as Figure 1As shown, a general recommendation model MAEE (Mixture of Aspect-Explicit-Expert) is proposed, which models the user's preference for the recommended content in the context information as different prediction tasks (such as predicting the probability of the user clicking / purchasing / preferencing a recommended content), and the model mainly consists of three parts:

[0050] 1. Expert network set AEE (Aspect-Explicit-Expert)

[0051] The MAEE model contains multiple AEE networks, and each AEE aims to learn the differentiated representation of user preferences from different attributes (such as population, time period, etc.). In each aspect, the input space is divided according to the attribute, for example, under the time attribute, the AEE network can be divided into three time periods: morning, noon, and evening. Further, each AEE contains several class-specific sub-networks (class-specific sub-experts) for learning the differentiated representation between different attribute values under each attribute. Finally, the sub-expert vectors output for different attribute values under the same attribute can be fused by a masked attention mechanism (Masked Attention Mechanism) to obtain an expert vector.

[0052] 2. Gating network (Expert Gates)

[0053] The gating network can determine the importance of different AEEs for the prediction task and fuse the representation results (different expert vectors) of different AEEs by weighting. Different prediction tasks have their own independent gating networks, and each AEE has different weights for different prediction tasks.

[0054] 3. Tower network (Task-specific Tower)

[0055] The tower network, also known as the target task-specific network, can convert the result of the fusion of the representation of each expert network (fusion vector) into the prediction result of the prediction task, where each prediction task has an independent prediction network and independently outputs the prediction result of different prediction tasks. For tasks with binary prediction results, a sigmoid activation function (a function that maps variables between 0 and 1) can be used in the last layer of the tower network. For example, the click / purchase / like prediction tasks in most personalized recommendation instances are binary tasks.

[0056] It should be noted that during the training of the general recommendation model described above, cross-entropy loss can be applied to each prediction task, and the weighted sum of all losses in backpropagation can be used to jointly optimize multiple tasks. In fact, the weights of each prediction task can be tuned as static hyperparameters or dynamically during the training process to improve performance.

[0057] The general recommendation model provided in this embodiment transforms the feature vectors of recommendation information into sub-expert vectors corresponding to different attribute values ​​through multiple sub-networks. By obtaining expert vectors for each attribute through multiple expert network sets, the recommendation information can be differentiated for learning and representation under different attributes and values. Simultaneously, by fusing the sub-expert vectors within each expert network set to obtain expert vectors, and by weightedly connecting multiple expert vectors to obtain a fusion vector, and by predicting recommendation information based on the fusion vector, commonalities can be mined from each pair of differential representations, thereby improving the accuracy of recommendation prediction. Based on this, the above general recommendation model can effectively mine the differences and similarities between recommendation information under different attributes and with different attribute values. Compared to traditional multi-expert networks, this recommendation model can effectively improve the accuracy of content recommendation.

[0058] Based on the above general recommendation model, in one embodiment, such as Figure 2 As shown, an information recommendation method is provided. Taking the application of this method to a computer device as an example, it includes the following steps:

[0059] 101. In response to the information recommendation request, obtain the feature vector of the recommendation information, wherein the recommendation information includes user information, recommendation content and context information, and the recommendation information also includes multiple attributes and the attribute value of each attribute.

[0060] Specifically, when a computer device receives an information recommendation request, it can obtain the feature vector of the recommendation information to be predicted. This feature vector can be obtained by feature processing a single recommendation or multiple recommendations. In this embodiment, each recommendation consists of user information, recommendation content, and contextual information. The user information, recommendation content, and contextual information can include various attributes, such as user gender, user age, user status, user behavior, product category, time, location, and channel. For each attribute, the recommendation information also includes a corresponding attribute value, such as a female user, a user age between 20 and 25 years old, a new user status, a food delivery category, a midday time, a location in an office building area, and a channel on a mobile application, etc.

[0061] In the embodiment, when the recommendation information is one piece, the prediction result output by the general recommendation model is the probability of the user's click / purchase / preference for the recommended content in the piece of recommendation information; when the recommendation information is multiple pieces, the prediction result output by the general recommendation model is the probability of the user's click / purchase / preference for the recommended content in each piece of recommendation information. Generally, in order to improve the prediction efficiency, when the recommendation information is multiple pieces, the user information in each piece of recommendation information can be set to be the same, the recommended content can be set to be different, and the context information can be set to be the same as the current scene in which the user is located (such as the same time and place), so that the general recommendation model outputs the probability of the user's click / purchase / preference for different recommended contents in the current scene, thereby facilitating the recommendation system to simultaneously recommend multiple recommended contents to the user.

[0062] 102. Based on the feature vector of the recommendation information, a plurality of expert vectors are obtained through the expert network set corresponding to each attribute, wherein each expert vector is obtained by fusing the sub-expert vectors output by each sub-network in the expert network set, and each sub-network corresponds to an attribute value of an attribute.

[0063] Specifically, after obtaining the feature vector of the recommendation information, the feature vector of the recommendation information can be input into the general recommendation model described in the above embodiment, and a sub-expert vector corresponding to each attribute value can be output through each sub-network in each expert network set in the general recommendation model, and then a plurality of sub-expert vectors under the same attribute can be fused into an expert vector corresponding to each attribute.

[0064] In the embodiment, each expert network set aims to learn the differentiated representation of user preferences from different attributes, and each sub-network in the expert network set is divided according to the attribute values of the attributes. Each sub-network is used to learn the differentiated representation between different attribute values under each attribute. Further, the sub-expert vectors output for different attribute values under the same attribute can be fused by a mask attention mechanism to obtain an expert vector.

[0065] 103. The plurality of expert vectors are weighted and connected through a gate network to obtain a fusion vector, and the fusion vector is predicted through a tower network to obtain the prediction result of the recommendation information.

[0066] Specifically, after obtaining the plurality of expert vectors, each expert vector can be forwarded to the gate network, and the gate network can determine the importance of different expert network sets for the prediction task and weight and fuse different expert vectors to obtain a fusion vector. Further, the fusion vector can be forwarded to the tower network, and the tower network can convert the fusion vector into a prediction result of the prediction task. Finally, the prediction results of all prediction tasks are aggregated to obtain the prediction result of the recommendation information.

[0067] In the embodiment, different prediction tasks correspond to respective independent gating networks and tower networks. The gating networks determine different weights of each expert network set for different prediction tasks, and the tower networks can independently output prediction results of different prediction tasks. It is assumed that the number of prediction tasks is k. When k = 1, it is a single-task scenario, and when k > 1, it is a multi-task scenario. The prediction tasks can be at least one of a click rate prediction task, a conversion rate prediction task, and a preference prediction task. Correspondingly, the prediction results of the recommendation information can include at least one of a click rate prediction result, a conversion rate prediction result, and a preference prediction result of the user for the recommended content.

[0068] 104. Based on the prediction results of the recommendation information, it is determined whether the recommended content meets a preset recommendation condition. If yes, the recommended content is pushed to a client where the user information is located.

[0069] Specifically, after obtaining the prediction results of each prediction task, it is determined whether each recommended content meets a preset recommendation condition according to a recommendation condition preset in the recommendation system. If yes, the recommended content in the recommendation information is pushed to the user. If no, the recommended content in the recommendation information is not pushed to the user.

[0070] In the embodiment, the recommendation condition can be preset according to actual conditions. For example, the recommendation condition can be that the recommended content meets the preset recommendation condition when the prediction results of each prediction task reach a preset probability threshold. The recommendation condition can also be that the recommended content meets the preset recommendation condition when the prediction results of one or more prediction tasks reach a preset probability threshold. The recommendation condition can also be that the recommended content meets the preset recommendation condition when the rankings of the prediction results of one or more prediction tasks are within a preset ranking. When the prediction results of the recommended content meet the preset recommendation condition, the recommended content can be pushed to the user.

[0071] The information recommendation method provided in the embodiment can convert the feature vector of the recommendation information into the sub-expert vectors corresponding to different attribute values through multiple sub-networks, obtain the expert vectors corresponding to each attribute through multiple expert network sets, and can enable the recommendation information to be differentially learned and represented under different attributes and different attribute values. Meanwhile, the sub-expert vectors are fused to obtain the expert vectors in each expert network set, the multiple expert vectors are weighted and connected to obtain the fused vector, and the recommendation information is predicted based on the fused vector, so as to determine whether to push the recommendation content to the user according to the prediction result, which can mine the commonness of each differential representation, thereby improving the accuracy of the recommendation information prediction. Based on this, the above method can effectively mine the differences and similarities between the recommendation information under different attributes and different attribute values, and compared with the traditional recommendation method, the above method can effectively improve the accuracy of content recommendation.

[0072] Further, as a refinement and expansion of the above embodiment, in order to completely describe the implementation process of the embodiment, in one embodiment, a design method of a mask attention mechanism in a general recommendation model is provided. In the embodiment, the mask attention mechanism includes a Boolean mask, an attention operator, and an update gate network. In the expert network set, each sub-network is a multi-layer neural network, which can map the feature vector of the recommendation information to a multi-dimensional space. Further, for an attribute of the recommendation information, when the recommendation information includes a certain attribute value of the attribute, the mask attention mechanism can determine the sub-expert network corresponding to the corresponding attribute value as the main sub-network, and determine the other sub-expert networks as auxiliary sub-networks. For example, when the user gender attribute in the recommendation information is female, for the expert network set corresponding to the gender attribute, the sub-network corresponding to the female can be determined as the main sub-network, and the sub-network corresponding to the male can be determined as the auxiliary sub-network. Further, the mask attention mechanism can aggregate the auxiliary sub-expert vectors output by the auxiliary sub-networks to update the main expert vector output by the main sub-network.

[0073] Specifically, the mask attention mechanism can be composed of the following three operators:

[0074] (1) Boolean mask: used to generate an indication vector, which includes a plurality of elements corresponding to each sub-network in the expert network set, wherein the element corresponding to the main sub-network can be a first preset value (such as 1), and the element corresponding to the auxiliary sub-network can be a second preset value (such as 0). In the embodiment, the attribute value corresponding to the main sub-network is the same as the attribute value of the recommendation information, and the attribute value corresponding to the auxiliary sub-network is different from the attribute value of the recommendation information.

[0075] (2) Attention operator: for each expert network set, after the main sub-network and the auxiliary sub-network in the expert network set are determined, the auxiliary sub-expert vectors output by all auxiliary sub-networks can be further aggregated into an auxiliary representation to modify the main sub-expert vector output by the main sub-network. The aggregation process is realized by an attention operator, and the weight of each auxiliary sub-network can be determined by its relevance to the main sub-network.

[0076] (3) Update gate network: the final output of the expert network set is an expert vector obtained by combining the main sub-expert vector and the aggregated auxiliary sub-expert vector through the update gate network, to balance the specific information related to the attribute value of the recommended information and the common information irrelevant to the attribute value of the recommended information.

[0077] In this embodiment, the main sub-network and the auxiliary sub-network in the expert network set are determined by a Boolean mask, the auxiliary sub-expert vectors output by the auxiliary sub-networks are aggregated by an attention operator, and finally the aggregated auxiliary sub-expert vectors and the main sub-expert vector are fused by an update gate network to obtain an expert vector, which can enable the recommended information to be differentiated under different attribute values, and can also mine commonality for each differentiated representation, thereby effectively improving the accuracy of the recommended information representation.

[0078] In one embodiment, as shown in Figure 3 A design method of a general recommendation model is also provided. The general recommendation model includes a plurality of expert network sets, a general expert network, at least one gate network, and at least one tower network. The general expert network corresponds to the overall representation of the recommended information and is used to output a general expert vector corresponding to the overall representation of the recommended information. Further, the gate network can perform weighted connection on the expert vector output by each expert network set and the general expert vector to obtain a fusion vector of a prediction task, and the tower network can perform prediction on the fusion vector to obtain a prediction result of the prediction task.

[0079] In the above embodiments, the general recommendation model is also constructed on the basis of multiple sets of expert networks, wherein the general expert network is also a multi-layer neural network, which can output a general representation corresponding to the overall representation of the recommendation information independent of each attribute of the recommendation information. Further, for each prediction task, all expert representations, including individual expert vectors and general expert vectors, can be aggregated, and then all expert representations are connected by a gating network to obtain a fusion vector. Specifically, assuming that the number of prediction tasks is k, when k = 1, it is a single-task scenario, and when k > 1, it is a multi-task scenario. The general recommendation model with K tasks of L attributes includes 4 components: (1) a set of expert networks responsible for L attributes, (2) a general expert network focusing on the commonality of all attributes, (3) a gating network that aggregates all expert representations of K tasks, and (4) a tower network that finally outputs k tasks. In this embodiment, it is not necessary to limit the dimensions of the expert vectors and the general expert vectors to be the same. For example, for some prediction tasks, the dimension of the general expert vector can be expanded to increase the feature proportion of the general expert vector in the fusion vector, thereby avoiding the weakening of the influence of the general expert vector on the fusion vector, and further improving the accuracy of task prediction. By constructing a general expert network in the general expert network, the embodiment can further enhance the commonality mining of the difference representation, thereby improving the accuracy of the recommendation information prediction.

[0080] Further, based on the general recommendation model proposed in each of the above embodiments, as a refinement and extension of the above information recommendation method embodiment, in order to completely describe the implementation process of the present embodiment, an information recommendation method is provided, as shown in Figure 4 The method comprises the following steps:

[0081] 201, in response to an information recommendation request, obtaining a feature vector of recommendation information, wherein the recommendation information includes user information, recommendation content and context information, and the recommendation information further includes multiple attributes and attribute values of each attribute.

[0082] Specifically, when the computer device receives an information recommendation request, it can obtain the user information, recommendation content and context information to be predicted. Then, according to the user information, recommendation content and context information, the attributes of the recommendation information and the attribute values of each attribute can be determined, wherein the attributes can include at least one of user gender, user age, user state, user behavior, commodity category, time, place and channel, and the attribute values can be the potential values corresponding to each attribute. Further, the obtained user information, recommendation content and context information can be feature processed, and the feature processed user information, recommendation content and context information can be connected, and the attribute values of each attribute in the recommendation information can be added in the connected feature vector to obtain the feature vector of the recommendation information.

[0083] In the embodiment, the feature vector of the recommendation information can be obtained by one piece of recommendation information or by multiple pieces of recommendation information. When the recommendation information is multiple pieces, the user information in each piece of recommendation information can be set to be the same, the recommendation content can be set to be different, and the context information can be set to be the same as the current scene of the user (such as the same time and place). In this way, the output of the general recommendation model is the probability of clicking / purchasing / preference of the same user for different recommendation contents in the current scene, thereby facilitating the recommendation system to simultaneously recommend multiple recommendation contents to the user. In addition, adding the attribute value of each attribute in the feature vector of the recommendation information can also facilitate the general recommendation model to select the main sub-network and the auxiliary sub-network in each expert network set, thereby improving the prediction efficiency.

[0084] 202、based on the feature vector of the recommendation information, through the expert network set corresponding to each attribute, a plurality of expert vectors are obtained, wherein the expert vector is obtained by fusing the sub-expert vectors output by each sub-network in the expert network set, and each sub-network corresponds to an attribute value of an attribute.

[0085] Specifically, first, the main sub-network and the auxiliary sub-network corresponding to the recommendation information can be determined in each expert network set according to the attribute value of each attribute in the recommendation information, wherein the attribute value corresponding to the main sub-network is the same as the attribute value of the recommendation information, and the attribute value corresponding to the auxiliary sub-network is different from the attribute value of the recommendation information. Then, the feature vector of the recommendation information can be mapped to the main expert vector through the main sub-network, and the feature vector of the recommendation information can be mapped to the auxiliary expert vector through the auxiliary sub-network. Finally, for each expert network set, the auxiliary expert vectors can be aggregated, and the aggregated auxiliary expert vectors and the main expert vector can be weighted and fused to obtain the expert vector.

[0086] Further, the process of aggregating the expert vector can be realized by the following method: first, the correlation between each auxiliary sub-network and the main sub-network can be obtained for each expert network set. Then, the weight of each auxiliary expert vector can be determined according to the correlation between each auxiliary sub-network and the main sub-network. Further, the auxiliary expert vectors can be aggregated according to the weight of each auxiliary expert vector to obtain the aggregated auxiliary expert vector. Finally, the main expert vector and the aggregated auxiliary expert vector can be weighted and fused through the update gate network to obtain the expert vector. In the embodiment, the correlation between each auxiliary sub-network and the main sub-network can be learned and induced by the general recommendation model in the historical data. Generally speaking, for some attributes with strong commonality, the correlation between the auxiliary sub-network and the main sub-network will be relatively high, and for some attributes with weak commonality, the correlation between the auxiliary sub-network and the main sub-network will be relatively low.

[0087] In addition, it should be noted that the dimensions of the fused aggregation auxiliary expert vector and the weighted fused expert vector and the sub-expert vector output by each sub-network are still the same, only the proportion of feature importance will change. In this embodiment, the importance of the main sub-expert vector in the expert vector is high, and the importance of the aggregation auxiliary expert vector is low. In this way, both the difference learning representation between different attribute values and the common features under the same attribute are retained, thereby effectively improving the accuracy of information representation and the accuracy of task prediction.

[0088] 203. Map the feature vector of the recommendation information to a general expert vector through a general expert network, wherein the general expert network corresponds to the overall representation of the recommendation information.

[0089] Specifically, the general expert network is a multi-layer neural network, which can output a general representation corresponding to the overall representation of the recommendation information independent of each attribute of the recommendation information. In this embodiment, it is not necessary to limit the dimensions of the expert vector and the general expert vector to be the same. Specifically, the target vector dimension of the general expert vector can be determined according to the vector dimension of each expert vector and the preset feature proportion of the general expert vector, and then the vector dimension of the general expert vector is expanded to the target vector dimension, wherein the target vector dimension is greater than the vector dimension of each expert vector, and specifically, the target vector dimension can be several times of the vector dimension of the expert vector. In this way, the dimension of the general expert vector can be expanded to improve the feature proportion of the general expert vector in the fused vector, thereby improving the accuracy of task prediction. This embodiment outputs a general expert vector corresponding to the overall representation of the recommendation information, which can further enhance the common mining of difference representation, thereby improving the accuracy of the recommendation information prediction.

[0090] 204. The multiple expert vectors and the general expert vector are weighted and connected through a gating network to obtain a fused vector, and the fused vector is predicted through a tower network to obtain a prediction result of the recommendation information.

[0091] In the embodiment, one or more prediction tasks can be set for the recommendation information, and different prediction tasks correspond to respective independent gating networks and tower networks. The gating network determines different weights of each expert network set for different prediction tasks, and the tower network is used to independently output prediction results of different prediction tasks. Specifically, for each prediction task, the weight of each expert network set and the weight of the general expert vector can be obtained through the gating network corresponding to the prediction task, and then each expert vector and the general expert vector are weightedly connected according to the weight of each expert network set and the weight of the general expert vector to obtain a fusion vector of the prediction task. Then, the fusion vector of the prediction task is predicted through the tower network corresponding to the prediction task to obtain the prediction result of the prediction task. Finally, the prediction result of each prediction task can be obtained, and the prediction result of the recommendation information can be obtained.

[0092] In the embodiment, the prediction task set for the recommendation information can be at least one of a click rate prediction task, a conversion rate prediction task and a preference prediction task. Correspondingly, the prediction result of the recommendation information can include at least one of a click rate prediction result, a conversion rate prediction result and a preference prediction result of the user for the recommended content.

[0093] 205、Based on the prediction result of the recommendation information, it is determined whether the recommended content meets the preset recommendation condition. If yes, the recommended content is pushed to the client where the user information is located.

[0094] Specifically, the recommendation threshold of each prediction task can be determined according to the recommendation condition, and then the prediction result of each prediction task is obtained. The prediction task can include at least one of a click rate prediction task, a conversion rate prediction task and a preference prediction task. Further, for each prediction task, the prediction result of the prediction task can be compared with the recommendation threshold of the prediction task. If the prediction result of each prediction task is greater than the recommendation threshold of the prediction task, the recommended content can be pushed to the client where the user information is located.

[0095] In the embodiment, the recommendation threshold can be a probability threshold or a ranking threshold. In addition, if the recommendation system considers that a certain task is more important, the recommendation threshold of the prediction result of the prediction task can be set to be higher, and the recommendation threshold of the prediction result of the other prediction task can be set to be lower. Alternatively, the calculation result (such as the sum or dot product value of multiple prediction results) of the prediction results of multiple prediction tasks can be compared with the recommendation threshold to determine whether the recommended content is pushed to the user.

[0096] The information recommendation method provided in this embodiment transforms the feature vector of recommendation information into sub-expert vectors corresponding to different attribute values ​​through multiple sub-networks. By using multiple expert network sets to obtain expert vectors corresponding to each attribute, the recommendation information can be differentiated in its learning representation under different attributes and attribute values. Simultaneously, by fusing the sub-expert vectors within each expert network set to obtain expert vectors, and outputting a general expert vector independent of each attribute through a general expert network, and by weighted concatenation of multiple expert vectors to obtain a fusion vector, the method predicts recommendation information based on the fusion vector. This enables the mining of commonalities among the differentiated representations, thereby improving the accuracy of recommendation information prediction. Based on this, the above method can effectively mine the differences and similarities between recommendation information under different attributes and attribute values, thus effectively improving the accuracy of content recommendation.

[0097] Furthermore, as Figures 1 to 4 The specific implementation of the method shown in this embodiment provides an information recommendation device, such as... Figure 5 As shown, the device includes: an information acquisition module 31, an information processing module 32, an information prediction module 33, and an information push module 34.

[0098] The information acquisition module 31 can be used to acquire the feature vector of the recommendation information in response to the information recommendation request. The recommendation information includes user information, recommendation content and context information. The recommendation information also includes multiple attributes and the attribute value of each attribute.

[0099] Information processing module 32 can be used to obtain multiple expert vectors based on the feature vector of the recommendation information through an expert network set corresponding to each attribute, wherein the expert vectors are obtained by fusing the sub-expert vectors output by each sub-network in the expert network set, and each sub-network corresponds to an attribute value of the attribute;

[0100] Information prediction module 33 can be used to perform weighted connection of the multiple expert vectors through a gating network to obtain a fusion vector, and to predict the fusion vector through a tower network to obtain the prediction result of the recommendation information;

[0101] The information push module 34 can be used to determine whether the recommended content meets the preset recommendation conditions based on the prediction results of the recommended information. If it does, the recommended content is pushed to the client where the user information is located.

[0102] In a specific application scenario, the information obtaining module 31 can be specifically configured to obtain user information, recommended content and context information in response to an information recommendation request; determine attributes of recommended information and attribute values of each attribute according to the user information, the recommended content and the context information, wherein the attributes include at least one of user gender, user age, user state, user behavior, commodity category, time, place and channel; perform feature processing on the user information, the recommended content and the context information, and connect the user information, the recommended content, the context information and the attribute values of each attribute in the recommended information after the feature processing to obtain a feature vector of the recommended information.

[0103] In a specific application scenario, the information processing module 32 can be specifically configured to determine a main sub-network and an auxiliary sub-network corresponding to the recommended information in each expert network set according to the attribute values of each attribute in the recommended information, wherein the attribute value corresponding to the main sub-network is the same as the attribute value of the recommended information, and the attribute value corresponding to the auxiliary sub-network is different from the attribute value of the recommended information; map the feature vector of the recommended information into a main expert vector through the main sub-network, and map the feature vector of the recommended information into an auxiliary expert vector through the auxiliary sub-network; aggregate the auxiliary expert vectors for each expert network set, and perform weighted fusion on the aggregated auxiliary expert vectors and the main expert vector to obtain an expert vector.

[0104] In a specific application scenario, the information processing module 32 can be specifically configured to obtain the correlation between each auxiliary sub-network and the main sub-network for each expert network set; determine the weight of each auxiliary expert vector according to the correlation between each auxiliary sub-network and the main sub-network; aggregate the auxiliary expert vectors according to the weight of each auxiliary expert vector to obtain an aggregated auxiliary expert vector; and perform weighted fusion on the main expert vector and the aggregated auxiliary expert vector through an update gate network to obtain an expert vector.

[0105] In a specific application scenario, the information processing module 32 can be further configured to map the feature vector of the recommended information into a general expert vector through a general expert network, wherein the general expert network corresponds to the overall representation of the recommended information; and the information prediction module 33 can be further configured to perform weighted connection on the expert vector output by each expert network set and the general expert vector through the gate network to obtain the fusion vector.

[0106] In a specific application scenario, the information processing module 32 is further configured to determine a target vector dimension of the general expert vector according to a vector dimension of each of the expert vectors and a preset characteristic proportion of the general expert vector; and expand the vector dimension of the general expert vector to the target vector dimension, where the target vector dimension is greater than the vector dimension of each of the expert vectors.

[0107] In a specific application scenario, the number of the gating networks and the tower networks is one or more, and each of the gating networks and each of the tower networks corresponds to a prediction task; the information prediction module 33 is specifically configured to, for each of the prediction tasks, acquire the weight of each of the expert network sets through the gating network corresponding to the prediction task, and perform weighted connection on the expert vectors output by each of the expert network sets according to the weight of each of the expert network sets to obtain a fusion vector of the prediction task; for each of the prediction tasks, perform prediction on the fusion vector of the prediction task through the tower network corresponding to the prediction task to obtain a prediction result of the prediction task; and acquire the prediction result of the recommendation information according to the prediction result of each of the prediction tasks.

[0108] In a specific application scenario, the information pushing module 34 is specifically configured to determine a recommendation threshold of each of the prediction tasks according to the recommendation condition; acquire the prediction result of each of the prediction tasks, where the prediction task includes at least one of a click rate prediction task, a conversion rate prediction task and a preference degree prediction task; for each of the prediction tasks, compare the prediction result of the prediction task with the recommendation threshold of the prediction task; and if the prediction result of each of the prediction tasks is greater than the recommendation threshold of the prediction task, push the recommendation content to a client where the user information is located.

[0109] In a specific application scenario, the information recommendation method is applied to a pre-trained general recommendation model, the general recommendation model includes a plurality of expert network sets, at least one gating network and at least one tower network, where each of the expert network sets corresponds to an attribute of the recommendation information, the expert network set includes a plurality of sub-networks and a mask attention mechanism, the sub-network is configured to output a sub-expert vector corresponding to an attribute value of the attribute, and the mask attention mechanism is configured to fuse each of the sub-expert vectors into the expert vector; each of the gating networks corresponds to a prediction task, and is configured to perform weighted connection on the expert vectors output by each of the expert network sets to obtain a fusion vector of the prediction task; and each of the tower systems corresponds to a prediction task, and is configured to perform prediction on the fusion vector of the prediction task to obtain a prediction result of the prediction task.

[0110] In a specific application scenario, the mask attention mechanism comprises a Boolean mask, an attention operator and an update gate network, wherein the Boolean mask is configured to generate an indication vector comprising elements corresponding to the plurality of sub-networks respectively, wherein an element corresponding to a main sub-network is a first preset value, and an element corresponding to an auxiliary sub-network is a second preset value, the attribute value corresponding to the main sub-network is the same as the attribute value of the recommendation information, and the attribute value corresponding to the auxiliary sub-network is different from the attribute value of the recommendation information; the attention operator is configured to aggregate an auxiliary expert vector output by the auxiliary sub-network; and the update gate network is configured to perform weighted fusion on the aggregated auxiliary expert vector and a main expert vector output by the main sub-network to obtain an expert vector.

[0111] In a specific application scenario, the general recommendation model further comprises a general expert network, wherein the general expert network corresponds to the overall representation of the recommendation information and is configured to output a general expert vector corresponding to the overall representation of the recommendation information; and the gate network is further configured to perform weighted connection on the expert vector output by each expert network set and the general expert vector to obtain a fusion vector of the prediction task.

[0112] It should be noted that other corresponding descriptions of the functions of the information recommendation device provided in the embodiment can be referred to the corresponding descriptions in the Figures 1 to 4 , which will not be described here.

[0113] Based on the above method as shown in Figures 1 to 4 , accordingly, the embodiment also provides a storage medium having a computer program stored thereon, which is executed by a processor to implement the information recommendation method as shown in Figures 1 to 4 .

[0114] Based on such understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.), and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various implementation scenarios of the present application.

[0115] Based on the above method as shown in Figures 1 to 4 , and Figure 5 the information recommendation device embodiment, in order to achieve the above purpose, the embodiment further provides a computer device for information recommendation, which can be a personal computer, a server, a smart phone, a tablet computer, a smart watch, or other network devices, etc. The computer device comprises a storage medium and a processor; the storage medium is configured to store a computer program and an operating system; and the processor is configured to execute the computer program to implement the above method as shown in Figures 1 to 4The method shown.

[0116] Optionally, the computer device can further include an internal memory, a communication interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, a display, an input device such as a keyboard, and the like. Optionally, the communication interface can further include a USB interface, a card reader interface, and the like. The network interface can optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), and the like.

[0117] Those skilled in the art can understand that the computer device structure for operation action recognition provided by the embodiment does not constitute a limitation on the computer device, and can include more or fewer components, or combine certain components, or different component arrangements.

[0118] The storage medium can further include an operating system and a network communication module. The operating system is a program for managing the hardware of the computer device and the to-be-identified software resources, supporting the running of the information processing program and other to-be-identified software and / or programs. The network communication module is used to realize the communication between the components in the storage medium, and the communication with other hardware and software in the information processing computer device.

[0119] Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software and necessary general hardware platforms, or by hardware. The present application responds to an information recommendation request, obtains a feature vector of recommendation information, wherein the recommendation information includes user information, recommendation content and context information, the recommendation information further includes a plurality of attributes and attribute values of each attribute, then, based on the feature vector of the recommendation information, a plurality of expert vectors are obtained through an expert network set corresponding to each attribute, wherein the expert vector is obtained by fusing a sub-expert vector output by each sub-network in the expert network set, and each sub-network corresponds to an attribute value of the attribute, further, the plurality of expert vectors are weighted and connected through a gate network to obtain a fusion vector, the fusion vector is predicted through a tower network to obtain a prediction result of the recommendation information, finally, based on the prediction result of the recommendation information, it is judged whether the recommendation content meets a preset recommendation condition, if so, the recommendation content is pushed to a client where the user information is located. Compared with the prior art, the above method can effectively mine the difference and similarity between the recommendation information in different attributes and different attribute values, and compared with the traditional recommendation method, the above method can effectively improve the accuracy of content recommendation.

[0120] Those skilled in the art can understand that the modules or flows in the drawings are not necessarily required for implementing the present application. Those skilled in the art can understand that the modules in the devices in the implementation scenarios can be distributed in the devices in the implementation scenarios according to the description of the implementation scenarios, or can be changed to be located in one or more devices different from the implementation scenarios. The modules in the above implementation scenarios can be combined into one module, or can be further split into multiple sub-modules.

[0121] The above application numbers are only for description, and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only some specific implementation scenarios of the present application, but the present application is not limited thereto, and any variations that can be thought of by those skilled in the art should fall within the protection scope of the present application.

Claims

1. An information recommendation method, characterized in that, The method includes: In response to an information recommendation request, a feature vector of the recommendation information is obtained, wherein the recommendation information includes user information, recommended content and context information, and the recommendation information also includes multiple attributes and the attribute value of each attribute; Based on the feature vector of the recommendation information, multiple expert vectors are obtained through expert network sets corresponding to each attribute. Each expert vector is obtained by fusing the sub-expert vectors output by each sub-network in the expert network set, where each sub-network corresponds to an attribute value of the attribute. Each expert network set includes a masked attention mechanism, which includes a Boolean mask, an attention operator, and an update gate network. For each attribute in the recommendation information, when the recommendation information includes an attribute value, the masked attention mechanism determines the sub-network corresponding to the attribute value as the main sub-network and the other sub-networks as auxiliary sub-networks. The feature vector of the recommendation information is mapped to the main expert vector through the main sub-network, and the feature vector of the recommendation information is mapped to the auxiliary expert vector through the auxiliary sub-networks. For each expert network set, the auxiliary expert vectors are aggregated, and the aggregated auxiliary expert vectors and the main expert vectors are weighted and fused to obtain the expert vector. The multiple expert vectors are weighted and connected by a gating network to obtain a fusion vector. The fusion vector is then used to predict the recommendation information by a tower network. Based on the prediction results of the recommendation information, it is determined whether the recommended content meets the preset recommendation conditions. If it does, the recommended content is pushed to the client where the user information is located.

2. The method according to claim 1, characterized in that, The step of obtaining the feature vector of the recommendation information in response to the information recommendation request includes: In response to information recommendation requests, obtain user information, recommended content, and contextual information; Based on the user information, recommended content, and context information, the attributes of the recommended information and the attribute value of each attribute are determined, wherein the attributes include at least one of the following: user gender, user age, user status, user behavior, product category, time, location, and channel. The user information, recommended content, and context information are characterized, and the attribute values ​​of each attribute in the characterized user information, recommended content, context information, and recommended information are concatenated to obtain the feature vector of the recommended information.

3. The method according to claim 1, characterized in that, For each of the expert network sets, the auxiliary expert vectors are aggregated, and the aggregated auxiliary expert vectors and the main expert vectors are weighted and fused to obtain expert vectors, including: For each set of expert networks, obtain the correlation between each auxiliary sub-network and the main sub-network; The weight of each auxiliary expert vector is determined based on the correlation between each auxiliary subnetwork and the main subnetwork. Based on the weight of each of the auxiliary expert vectors, the auxiliary expert vectors are aggregated to obtain aggregated auxiliary expert vectors; By updating the gate network, the main expert vector and the aggregated auxiliary expert vector are weighted and fused to obtain the expert vector.

4. The method according to claim 1, characterized in that, Before performing weighted concatenation of the multiple expert vectors through a gating network to obtain the fused vector, the method further includes: The feature vectors of the recommendation information are mapped to general expert vectors through a general expert network, wherein the general expert network corresponds to the overall representation of the recommendation information; The step of weighting and connecting the multiple expert vectors through a gating network to obtain a fused vector includes: The fusion vector is obtained by weighting and connecting the expert vectors output by each expert network set and the general expert vector through the gating network.

5. The method according to claim 4, characterized in that, Before obtaining the fused vector by weighted concatenation of the expert vectors output by each expert network set and the general expert vector through the gating network, the method further includes: The target vector dimension of the general expert vector is determined based on the vector dimension of each expert vector and the feature proportion of the preset general expert vector. The vector dimension of the general expert vector is expanded to the target vector dimension, wherein the target vector dimension is greater than the vector dimension of each expert vector.

6. The method according to claim 1, characterized in that, The number of the gated network and the tower network is one or more, and each gated network and each tower network corresponds to a prediction task; The step of weightedly connecting the multiple expert vectors through a gating network to obtain a fusion vector, and then predicting the fusion vector through a tower network to obtain the prediction result of the recommendation information, includes: For each prediction task, the weights of each expert network set are obtained through the gating network corresponding to the prediction task, and the expert vectors output by each expert network set are weighted and connected according to the weights of each expert network set to obtain the fusion vector of the prediction task. For each prediction task, the fusion vector of the prediction task is predicted through the tower network corresponding to the prediction task to obtain the prediction result of the prediction task. The prediction results of the recommendation information are obtained based on the prediction results of each prediction task.

7. The method according to claim 6, characterized in that, The step of determining whether the recommended content meets preset recommendation conditions based on the prediction result of the recommendation information, and if so, pushing the recommended content to the client where the user information is located, includes: Based on the recommendation criteria, determine the recommendation threshold for each prediction task; Obtain the prediction result for each of the prediction tasks, wherein the prediction tasks include at least one of click-through rate prediction task, conversion rate prediction task, and preference prediction task; For each prediction task, the prediction result of the prediction task is compared with the recommendation threshold of the prediction task; If the prediction result of each prediction task is greater than the recommendation threshold of the prediction task, then the recommended content will be pushed to the client where the user information is located.

8. The method according to any one of claims 1-7, characterized in that, The information recommendation method is applied to a pre-trained general recommendation model, which includes multiple expert network sets, at least one gating network, and at least one tower network, wherein... Each of the expert network sets corresponds to an attribute of the recommendation information. The expert network set includes multiple sub-networks and a masked attention mechanism. The sub-networks are used to output a sub-expert vector corresponding to an attribute value of the attribute. The masked attention mechanism is used to fuse each of the sub-expert vectors into the expert vector. Each of the gated networks corresponds to a prediction task and is used to perform weighted concatenation of the expert vectors output by each of the expert network sets to obtain the fusion vector of the prediction task. Each tower network corresponds to a prediction task, used to predict the fusion vector of the prediction task to obtain the prediction result of the prediction task.

9. The method according to claim 8, characterized in that, In the mask attention mechanism: The Boolean mask is used to generate an indicator vector, which includes elements corresponding to multiple sub-networks respectively. The element corresponding to the main sub-network is a first preset value, and the element corresponding to the auxiliary sub-network is a second preset value. The attribute value corresponding to the main sub-network is the same as the attribute value of the recommendation information, and the attribute value corresponding to the auxiliary sub-network is different from the attribute value of the recommendation information. The attention operator is used to aggregate the auxiliary expert vectors output by the auxiliary sub-network; The update gate network is used to perform weighted fusion of the aggregated auxiliary expert vector and the main expert vector output by the main sub-network to obtain the expert vector.

10. The method according to claim 8, characterized in that, The general recommendation model also includes a general expert network, wherein... The general expert network corresponds to the overall representation of the recommendation information and is used to output a general expert vector corresponding to the overall representation of the recommendation information. The gating network is also used to perform a weighted connection of the expert vectors output by each expert network set and the general expert vector to obtain the fusion vector of the prediction task.

11. An information recommendation device, characterized in that, The device includes: The information acquisition module is used to acquire the feature vector of the recommended information in response to the information recommendation request. The recommended information includes user information, recommended content and context information. The recommended information also includes multiple attributes and the attribute value of each attribute. An information processing module is used to obtain multiple expert vectors based on the feature vectors of the recommendation information through an expert network set corresponding to each attribute. The expert vectors are obtained by fusing the sub-expert vectors output by each sub-network in the expert network set, with each sub-network corresponding to an attribute value of the attribute. Each expert network set includes a masked attention mechanism, which includes a Boolean mask, an attention operator, and an update gate network. For each attribute in the recommendation information, when the recommendation information includes an attribute value, the masked attention mechanism determines the sub-network corresponding to the attribute value as the main sub-network and the other sub-networks as auxiliary sub-networks. The feature vectors of the recommendation information are mapped to main expert vectors through the main sub-networks, and the feature vectors of the recommendation information are mapped to auxiliary expert vectors through the auxiliary sub-networks. For each expert network set, the auxiliary expert vectors are aggregated, and the aggregated auxiliary expert vectors and the main expert vectors are weighted and fused to obtain an expert vector. The information prediction module is used to perform weighted connection of the multiple expert vectors through a gating network to obtain a fusion vector, and then use a tower network to predict the fusion vector to obtain the prediction result of the recommendation information. The information push module is used to determine whether the recommended content meets the preset recommendation conditions based on the prediction results of the recommended information. If it does, the recommended content is pushed to the client where the user information is located.

12. The apparatus according to claim 11, characterized in that, The information acquisition module is specifically used to acquire user information, recommended content, and context information in response to an information recommendation request. Based on the user information, recommended content, and context information, the attributes of the recommended information and the attribute value of each attribute are determined, wherein the attributes include at least one of the following: user gender, user age, user status, user behavior, product category, time, location, and channel. The user information, recommended content, and context information are characterized, and the attribute values ​​of each attribute in the characterized user information, recommended content, context information, and recommended information are concatenated to obtain the feature vector of the recommended information.

13. The apparatus according to claim 11, characterized in that, The information processing module is further configured to obtain the correlation between each auxiliary sub-network and the main sub-network for each expert network set; The weight of each auxiliary expert vector is determined based on the correlation between each auxiliary subnetwork and the main subnetwork. Based on the weight of each of the auxiliary expert vectors, the auxiliary expert vectors are aggregated to obtain aggregated auxiliary expert vectors; By updating the gate network, the main expert vector and the aggregated auxiliary expert vector are weighted and fused to obtain the expert vector.

14. The apparatus according to claim 11, characterized in that, The information processing module is further configured to map the feature vector of the recommendation information into a general expert vector through a general expert network, wherein the general expert network corresponds to the overall representation of the recommendation information; The information prediction module is further configured to perform a weighted connection of the expert vectors output by each expert network set and the general expert vector through the gating network to obtain the fusion vector.

15. The apparatus according to claim 14, characterized in that, The information processing module is further configured to determine the target vector dimension of the general expert vector based on the vector dimension of each expert vector and the feature proportion of the preset general expert vector. The vector dimension of the general expert vector is expanded to the target vector dimension, wherein the target vector dimension is greater than the vector dimension of each expert vector.

16. The apparatus according to claim 11, characterized in that, The number of the gated network and the tower network is one or more, and each gated network and each tower network corresponds to a prediction task; The information prediction module is specifically used to obtain the weights of each expert network set through a gating network corresponding to the prediction task for each prediction task, and to perform weighted connection of the expert vectors output by each expert network set according to the weights of each expert network set to obtain the fusion vector of the prediction task. For each prediction task, the fusion vector of the prediction task is predicted through the tower network corresponding to the prediction task to obtain the prediction result of the prediction task. The prediction results of the recommendation information are obtained based on the prediction results of each prediction task.

17. The apparatus according to claim 16, characterized in that, The information push module is specifically used to determine the recommendation threshold for each prediction task based on the recommendation conditions. Obtain the prediction result for each of the prediction tasks, wherein the prediction tasks include at least one of click-through rate prediction task, conversion rate prediction task, and preference prediction task; For each prediction task, the prediction result of the prediction task is compared with the recommendation threshold of the prediction task; If the prediction result of each prediction task is greater than the recommendation threshold of the prediction task, then the recommended content will be pushed to the client where the user information is located.

18. The apparatus according to any one of claims 11-17, characterized in that, The information recommendation device is applied to a pre-trained general recommendation model, which includes multiple expert network sets, at least one gating network, and at least one tower network, wherein... Each of the expert network sets corresponds to an attribute of the recommendation information. The expert network set includes multiple sub-networks and a masked attention mechanism. The sub-networks are used to output a sub-expert vector corresponding to an attribute value of the attribute. The masked attention mechanism is used to fuse each of the sub-expert vectors into the expert vector. Each of the gated networks corresponds to a prediction task and is used to perform weighted concatenation of the expert vectors output by each of the expert network sets to obtain the fusion vector of the prediction task. Each tower network corresponds to a prediction task, used to predict the fusion vector of the prediction task to obtain the prediction result of the prediction task.

19. The apparatus according to claim 18, characterized in that, In the mask attention mechanism: The Boolean mask is used to generate an indicator vector, which includes elements corresponding to multiple sub-networks respectively. The element corresponding to the main sub-network is a first preset value, and the element corresponding to the auxiliary sub-network is a second preset value. The attribute value corresponding to the main sub-network is the same as the attribute value of the recommendation information, and the attribute value corresponding to the auxiliary sub-network is different from the attribute value of the recommendation information. The attention operator is used to aggregate the auxiliary expert vectors output by the auxiliary sub-network; The update gate network is used to perform weighted fusion of the aggregated auxiliary expert vector and the main expert vector output by the main sub-network to obtain the expert vector.

20. The apparatus according to claim 18, characterized in that, The general recommendation model also includes a general expert network, wherein... The general expert network corresponds to the overall representation of the recommendation information and is used to output a general expert vector corresponding to the overall representation of the recommendation information. The gating network is also used to perform a weighted connection of the expert vectors output by each expert network set and the general expert vector to obtain the fusion vector of the prediction task.

21. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

22. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

Images