Material recommendation method and device, computer equipment and storage medium

Abstract

The invention relates to a material recommendation method and device, computer equipment and a storage medium, and relates to the technical field of information recommendation. The method comprises the following steps: acquiring user attribute information of a target user and material attribute information of candidate materials; inputting the user attribute information and the material attribute information into a task network of a recommendation model to predict recommendation targets, and obtaining predicted values of the candidate materials corresponding to the recommendation targets; inputting the material attribute information of the candidate material and the predicted value of the candidate material corresponding to each recommendation target into a first fusion network of a recommendation model, and carrying out nonlinear weighted fusion on the material granularity to obtain a first fusion predicted value of the user behavior of the target user on the candidate material; recommending the candidate materials based on the first fusion predicted value; through the method, the accuracy of model prediction can be guaranteed while the influence of deviation characteristics is reduced, and the overall performance and recommendation effect of a recommendation system are improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of information recommendation, and in particular to a material recommendation method, device, computer equipment and storage medium. Background Art

[0002] Nowadays, large-scale recommendation ranking systems have been widely used. In recommendation ranking, it is usually necessary to predict each material to be recommended in terms of different recommendation targets, so as to make material recommendations based on the predicted values ​​of the material to be recommended corresponding to different recommendation targets.

[0003] In related technologies, recommendation targets are usually predicted through recommendation models. In order to eliminate biased features in materials, in the actual application stage of recommendation models, the biased features are often uniformly set to the same value to achieve the purpose of eliminating deviations.

[0004] However, the features used in the model training process are often biased. The above method of eliminating bias will cause inconsistencies in the feature input during model training and application. In addition, in multi-objective models, since there is a certain correlation between multiple recommendation targets, the above differences will lead to poor accuracy of model prediction, thereby affecting the overall performance and recommendation effect of the recommendation system. Summary of the invention

[0005] The embodiments of the present application provide a material recommendation method, device, computer equipment and storage medium, which can reduce the impact of deviation characteristics while ensuring the accuracy of model prediction and improving the overall performance and recommendation effect of the recommendation system. The technical solution is as follows.

[0006] In one aspect, a material recommendation method is provided, the method comprising: Obtain user attribute information of the target user and material attribute information of the candidate material; Inputting the user attribute information of the target user and the material attribute information of the candidate material into the task network of the recommendation model to predict the recommendation target, and obtaining the prediction value of the candidate material corresponding to each recommendation target; the prediction value of each recommendation target is used to represent the prediction information of the user behavior of the target user on the candidate material corresponding to the recommendation target; Inputting the material attribute information of the candidate material and the predicted value of the candidate material corresponding to each recommendation target into the first fusion network of the recommendation model to perform nonlinear weighted fusion at the material granularity to obtain the first fusion prediction value of the user behavior of the target user on the candidate material; The candidate material is recommended based on the first fused prediction value.

[0007] In another aspect, a material recommendation device is provided, the device comprising: An information acquisition module is used to acquire user attribute information of a target user and material attribute information of a candidate material; A prediction module is used to input the user attribute information of the target user and the material attribute information of the candidate material into the task network of the recommendation model to predict the recommendation target, and obtain the prediction value of the candidate material corresponding to each recommendation target; the prediction value of each recommendation target is used to represent the prediction information of the user behavior of the target user on the candidate material corresponding to the recommendation target; A first fusion module is used to input the material attribute information of the candidate material and the predicted value of the candidate material corresponding to each recommendation target into the first fusion network of the recommendation model to perform nonlinear weighted fusion at the material granularity to obtain a first fusion prediction value of the user behavior of the target user on the candidate material; A material recommendation module is used to recommend the candidate material based on the first fusion prediction value.

[0008] In a possible implementation manner, the recommendation model further includes a second fusion network, and the device further includes: The second fusion module is used to input the first user feature and the first material feature corresponding to each recommendation target, and the predicted value of the candidate material corresponding to each recommendation target into the second fusion network to perform nonlinear weighted fusion at the user granularity to obtain the second fusion prediction value of the user behavior of the target user on the candidate material; wherein the first user feature and the first material feature corresponding to each recommendation target are obtained after the task network extracts the features corresponding to each recommendation target from the user attribute information and the material attribute information; The material recommendation module is used to recommend the candidate material based on the first fusion prediction value and the second fusion prediction value.

[0009] In a possible implementation, the material recommendation module is used to: Performing weighted fusion on the first fused prediction value and the second fused prediction value to obtain a comprehensive prediction value of the candidate material; Determining a recommended order of the candidate materials based on the comprehensive predicted values ​​of the candidate materials; The candidate materials are recommended based on the recommendation order.

[0010] In a possible implementation manner, the device further includes: A behavior information acquisition module, used to acquire the target user behavior information of the target user on the candidate material; the target user behavior information includes the behavior attribute values ​​of the target user on the candidate material corresponding to each recommendation target; The first updating module is used to update the parameters of the corresponding fusion network based on the target user behavior information and the fusion prediction value; wherein, when the fusion prediction value is the first fusion prediction value, the fusion network is the first fusion network; when the fusion prediction value is the second fusion prediction value, the fusion network is the second fusion network.

[0011] In a possible implementation, the first updating module is used to: Acquire multiple user behavior information corresponding to the candidate material; Based on the multiple user behavior information corresponding to the candidate materials, calculating the mean and variance of the behavior attribute values ​​of the candidate materials corresponding to each recommendation target; A loss function is calculated based on the behavior attribute value of each recommended target in the target user behavior information, the mean and variance value of the behavior attribute value of the candidate material corresponding to each recommended target, and the fusion prediction value to obtain a function value of the loss function; Based on the function value of the loss function, parameters of the corresponding fusion network are updated.

[0012] In a possible implementation, the task network includes a shared network shared by each recommendation target and an independent tower network corresponding to each recommendation target; the prediction module is used to: Inputting the user attribute information and the material attribute information into a shared network shared by each recommendation target to perform feature extraction, and obtaining a first user feature and a first material feature corresponding to each recommendation target; The first user feature and the first material feature corresponding to each recommendation target are respectively input into the independent tower network corresponding to each recommendation target to perform recommendation target prediction, so as to obtain the prediction value of the candidate material corresponding to each recommendation target.

[0013] In a possible implementation, the task network includes a shared network shared by each recommendation target and an independent tower network corresponding to each recommendation target; the independent tower network includes a first sub-network and a second sub-network; the prediction module is used to: Inputting the user attribute information and the material attribute information into a shared network shared by each recommendation target to perform feature extraction, and obtaining a second user feature and a second material feature corresponding to each recommendation target; Inputting the second user features and the second material features corresponding to each recommendation target into the first sub-network of the independent tower network corresponding to each recommendation target respectively for feature extraction, and obtaining the first user features and the first material features corresponding to each recommendation target; The first user feature and the first material feature corresponding to each recommendation target are respectively input into the second subnetwork of the independent tower network corresponding to each recommendation target to perform recommendation target prediction, so as to obtain the prediction value of the candidate material corresponding to each recommendation target.

[0014] In a possible implementation manner, the device further includes: The second updating module is used to update the parameters of the task network based on the behavior attribute value of each recommended target in the target user behavior information obtained by the behavior information obtaining module and the predicted value of the candidate material corresponding to each recommended target.

[0015] On the other hand, a computer device is provided, which includes a processor and a memory, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the above-mentioned material recommendation method.

[0016] On the other hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium, and the computer program is loaded and executed by a processor to implement the above-mentioned material recommendation method.

[0017] On the other hand, a computer program product is provided, the computer program product comprising at least one computer program, the computer program being loaded and executed by a processor to implement the material recommendation method provided in the above-mentioned various optional implementations.

[0018] The technical solution provided by this application may have the following beneficial effects: The material recommendation method provided in the embodiment of the present application, when recommending materials for multiple recommendation targets, predicts the recommendation targets based on the user attribute information of the target user and the material attribute information of the candidate materials through the task network of the recommendation model, obtains the prediction values ​​of the candidate materials corresponding to each recommendation target, and performs nonlinear weighted fusion on the material granularity of the material attribute information of the candidate materials and the prediction values ​​of the candidate materials corresponding to each recommendation target through the first fusion network of the recommendation model, obtains the first fusion prediction value of the user behavior of the target user on the candidate materials, and recommends the candidate materials based on the first fusion prediction value; the first fusion network combines the material attribute information and the prediction information of the user's user behavior on the materials corresponding to each recommendation target, so as to fuse the prediction values ​​of multiple recommendation targets into one fusion prediction value on the basis of the material granularity through nonlinear weighting to achieve the purpose of debiasing the materials; compared with the method of uniformly setting the biased features to the same value, the method provided in the present application fully considers the correlation between the material characteristics and the multiple recommendation targets, maintains the consistency of the input content during model training and model application, and reduces the influence of the biased features through operations such as nonlinear weighted fusion, while ensuring the accuracy of the model prediction, thereby improving the overall performance and recommendation effect of the recommendation system.

[0019] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0021] Figure 1 A flow chart of a material recommendation method provided by an exemplary embodiment of the present application is shown; Figure 2 A flow chart of a material recommendation method provided by another exemplary embodiment of the present application is shown; Figure 3 A schematic diagram of a recommendation model provided by an exemplary embodiment of the present application is shown; Figure 4 A block diagram of a material recommendation device provided by an exemplary embodiment of the present application is shown; Figure 5 A structural block diagram of a computer device shown in an exemplary embodiment of the present application is shown; Figure 6 A structural block diagram of a computer device shown in another exemplary embodiment of the present application is shown. DETAILED DESCRIPTION

[0022] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0023] In the information flow recommendation system, users may have various behaviors towards materials, such as browsing, clicking, interacting, complaining, etc. The recommendation system can predict the possibility of users performing various behaviors towards materials, comprehensively rank and recommend materials, thereby improving the material recommendation effect and user experience. Among them, the user's behavior towards materials can be evaluated through recommendation goals, and recommendation goals may include CTR (Click-through Rate), Interaction Rate (interaction rate, such as sharing, liking, following, commenting, etc.) and Watch Time (consumption time), etc., among which the click-through rate refers to the probability of users clicking on the material, the interaction rate refers to the probability of users interacting with the material, and the consumption time refers to the length of time users interact with the material. The recommendation system can construct a recommendation model based on multiple recommendation goals to predict the recommendation target corresponding to the material through the recommendation model. In order to reduce the impact of biased features on material recommendations, a material recommendation method provided in an embodiment of the present application, Figure 1 A flow chart of a material recommendation method provided by an exemplary embodiment of the present application is shown. The method can be executed by a computer device, and the computer device can be implemented as a server or a terminal, such as Figure 1 As shown, the material recommendation method may include the following steps.

[0024] Step 110: Obtain user attribute information of the target user and material attribute information of the candidate material.

[0025] Among them, the target user may be the user for whom material recommendation is to be made, and the user attribute information may include basic information (such as age, gender, region, etc.), behavior history information (such as historical browsing records, search records, etc.), social relationship information (such as follow lists, friend interactions, etc.) and interest preference information (such as categories of collection information and categories of evaluation content, etc.).

[0026] For candidate materials, their material attribute information may include material category (such as video, image, text, link, etc.), material label, material source and material historical performance data (such as historical click rate, interaction rate, etc.).

[0027] Step 120, input the user attribute information of the target user and the material attribute information of the candidate material into the task network of the recommendation model to predict the recommendation target, and obtain the prediction value of the candidate material corresponding to each recommendation target; the prediction value of each recommendation target is used to characterize the prediction information of the target user's user behavior on the candidate material corresponding to the recommendation target.

[0028] The various recommendation goals can be click-through rate, interaction rate, consumption time, etc. Based on different actual needs, the recommendation goals can have different quantities and types of settings. For example, the interaction rate can be further set to like rate, comment rate, forwarding rate, collection rate, etc.

[0029] In an embodiment of the present application, a recommendation model configured in a computer device includes a task network, which may include independent tower networks corresponding to various recommendation targets and a shared network shared by various recommendation targets. In this case, the computer device may input user attribute information of the target user and material attribute information of the candidate material into the task network of a pre-trained recommendation model, and predict the input user attribute information and material attribute information corresponding to each recommendation target through the task network to obtain the predicted value of the candidate material corresponding to each recommendation target; the predicted value of each recommendation target is used to quantify the possibility of the target user performing user behavior corresponding to the recommendation target on the candidate material, for example, the predicted value of the click-through rate indicates the probability of the target user clicking on the candidate material, the predicted value of the interaction rate indicates the probability of the target user performing interactive behavior on the candidate material, and the predicted value of the consumption time is used to indicate the predicted interaction time of the target user on the candidate material.

[0030] Step 130, input the material attribute information of the candidate material and the predicted value of the candidate material corresponding to each recommendation target into the first fusion network of the recommendation model, perform nonlinear weighted fusion at the material granularity, and obtain the first fusion prediction value of the target user's user behavior on the candidate material.

[0031] The material attribute information of the candidate material and the predicted value of each recommended target are used as input vectors and input into the first fusion network of the recommendation model, which can also be called a debiased fusion network; inside the first fusion network, nonlinear weighted fusion is performed on different inputs according to the pre-trained network parameters (i.e., the weight parameters corresponding to each input vector) to obtain the first fusion predicted value of the candidate material. The network parameters in the first fusion network can be continuously optimized and adjusted during the model training process to adjust the degree of influence of different material attribute information and recommended target predicted values ​​on the final fusion result.

[0032] By combining the material attribute information of the candidate materials and the user's behavior prediction information on the candidate materials through the first fusion network to perform information fusion, the predicted values ​​of multiple recommendation targets can be unified into one value through nonlinear weighting on the basis of material granularity, that is, the first fusion prediction value that is finally output; through the above fusion method, the material's own characteristics and applicable scope can be clarified through material attributes, and the dynamic interaction situation can be reflected with the help of the user's behavior prediction information on the candidate materials (that is, the prediction value of each recommendation target). The combination of the two can accurately match materials and user needs and improve the accuracy of recommendations; in addition, the recommendation system can dynamically adjust its strategy and enhance adaptability based on the dynamic changes of user behavior prediction information while the material attributes are relatively stable in different situations; the debiased fusion network can also be used to perform weighted fusion of the predicted values ​​of each recommendation target for specific materials and specific users, reduce the deviation caused by material characteristics and individual differences of users, achieve personalized debiased recommendations, optimize user experience, and meet diverse recommendation needs.

[0033] The recommendation model is obtained by training based on a training sample set, which includes user attribute information of sample users, material attribute information of sample materials, and user behavior information labels of sample materials, and the user behavior labels include behavioral attribute labels of sample users for sample materials corresponding to each recommendation target; during model training, the user attribute information of sample users and material attribute information of sample materials are input into the task network for recommendation target prediction, and the prediction value of sample materials corresponding to each recommendation target is obtained, and the material attribute information of sample materials and the prediction value of sample materials corresponding to each recommendation target are input into the first fusion network of the recommendation model for nonlinear weighted fusion at the material granularity, and the first fusion prediction value of user behavior of the target user for the candidate material is obtained, and the parameters of the recommendation model are updated based on the first fusion prediction value of the candidate material, the prediction value of the sample material corresponding to each recommendation target, and the user behavior information label of the sample material, until the training completion condition is met, a trained recommendation model is obtained, and applied to the process of screening and recommending candidate materials for the target user. That is to say, in the embodiment of the present application, the model input is consistent when the model is trained and applied.

[0034] Step 140: recommend candidate materials based on the first fusion prediction value.

[0035] A recommendation sorting mechanism based on the first fused prediction value of each candidate material can be set up in the computer device, and all candidate materials are sorted according to the first fused prediction value of each candidate material. The candidate material with a higher prediction value has a greater potential value to the target user and is ranked higher in the recommendation list. When the target user requests recommended content, the computer device can recommend candidate materials based on the recommendation order indicated in the recommendation list.

[0036] In summary, the material recommendation method provided in the embodiment of the present application, when recommending materials for multiple recommendation targets, predicts the recommendation targets based on the user attribute information of the target user and the material attribute information of the candidate materials through the task network of the recommendation model, and obtains the prediction values ​​of the candidate materials corresponding to each recommendation target. The material attribute information of the candidate materials and the prediction values ​​of the candidate materials corresponding to each recommendation target are nonlinearly weighted fused at the material granularity through the first fusion network of the recommendation model to obtain the first fusion prediction value of the user behavior of the target user on the candidate materials, so as to recommend the candidate materials based on the first fusion prediction value. The first fusion network combines the material Attribute information, user behavior prediction information on materials corresponding to each recommendation target, and then, based on the material granularity, the prediction values ​​of multiple recommendation targets are fused into a fused prediction value through nonlinear weighting to achieve the purpose of debiasing the material; compared with the method of setting biased features to the same value, the method provided in the present application fully considers the material characteristics and the correlation between multiple recommendation targets, maintains the consistency of the input content during model training and model application, and reduces the impact of biased features through operations such as nonlinear weighted fusion, while ensuring the accuracy of model prediction, thereby improving the overall performance and recommendation effect of the recommendation system.

[0037] In a possible implementation, the recommendation model further includes a second fusion network, which is used to fuse the predicted values ​​of multiple recommendation targets based on user granularity, and to determine the recommendation order of candidate materials in combination with the first fusion network that fuses the predicted values ​​of multiple recommendation targets based on material granularity. Figure 2 A flow chart of a material recommendation method provided by another exemplary embodiment of the present application is shown. The method can be executed by a computer device, and the computer device can be implemented as a server or a terminal, such as Figure 2 As shown, the recommended method may include the following steps.

[0038] Step 210: Acquire user attribute information of the target user and material attribute information of the candidate material.

[0039] The user attribute information of the target user may include attribute information uploaded by the user, and may also include attribute information obtained by a computer device through statistical extraction based on historical behaviors of the target user.

[0040] Step 220, input the user attribute information of the target user and the material attribute information of the candidate material into the task network of the recommendation model to predict the recommendation target, and obtain the prediction value of the candidate material corresponding to each recommendation target; the prediction value of each recommendation target is used to characterize the prediction information of the target user's user behavior on the candidate material corresponding to the recommendation target.

[0041] The processing of input information by the task network includes two stages: feature extraction and recommendation target prediction; when performing feature extraction, the computer device can extract feature information related to material recommendation from user attribute information and material attribute information as the first user feature and the first material feature. Furthermore, corresponding to different recommendation targets, the computer device can extract feature information corresponding to different recommendation targets respectively as the first user feature and the first material feature corresponding to each of the different recommendation targets. The feature information related to material recommendation or the feature information related to each recommendation target can be set by relevant personnel based on actual needs, or can be determined by clustering and extracting the features of historical recommendation information and the features of the recommended users corresponding to the historical recommendation information, and this application does not limit this.

[0042] The task network in the recommendation model may include a shared network shared by each recommendation target and an independent tower network corresponding to each recommendation target. In a possible implementation, the computer device may perform feature extraction through the shared network and perform recommendation target prediction through the independent tower network. The process may be implemented as follows: Inputting user attribute information and material attribute information into a shared network shared by each recommendation target to perform feature extraction, and obtaining first user features and first material features corresponding to each recommendation target; The first user feature and the first material feature corresponding to each recommendation target are input into the independent tower network corresponding to each recommendation target to perform recommendation target prediction, and the prediction value of the candidate material corresponding to each recommendation target is obtained.

[0043] Among them, the internal structure of the shared network can be a three-layer fully connected neural network, and the internal structure of the independent tower network can also be a three-layer fully connected neural network; the number of shared networks can be one or more, and when there are multiple shared networks, the output of each shared network contains feature information corresponding to each recommendation target (including user features and material features), and the feature information corresponding to the recommendation target in the output of each shared network serves as the input of the corresponding independent tower network, and the output of the independent tower network is the predicted value of the corresponding recommendation target; schematically, when the number of shared networks is 3 and the number of recommendation targets is 3 and they are click-through rate, interaction rate and consumption time respectively, the task network can include shared network 1, shared network 2 and shared network 3 shared with the three recommendation targets, and independent tower network 1, independent tower network corresponding to the three recommendation targets respectively. 2 and independent tower network 3; the inputs of the three shared networks are all user attribute information and material attribute information. Each shared network extracts features to obtain first user features and first material features corresponding to each recommendation target and inputs them into each independent tower network. That is to say, the input of each independent tower network is the feature information corresponding to the recommendation target in the outputs of the three shared networks. Schematically, the input of independent tower network 1 comes from the three shared networks and is the feature information corresponding to the click-through rate in the outputs of the three shared networks. Each independent tower network predicts the recommendation target based on the first user features and first material features it receives, and obtains the predicted value of the corresponding recommendation target. For example, the output of independent tower network 1 is the predicted value of the click-through rate, the output of independent tower network 2 is the predicted value of the interaction rate, and the output of independent tower network 3 is the predicted value of the consumption time.

[0044] Alternatively, in another possible implementation, the task network includes a shared network shared by each recommendation target and an independent tower network corresponding to each recommendation target, and the independent tower network includes a first subnetwork and a second subnetwork, the first subnetwork is used for feature extraction, and the second subnetwork is used for recommendation target prediction, and the internal structure of each subnetwork may also be a three-layer fully connected neural network; in this case, the computer device may perform feature extraction through the shared network of the recommendation model and the first subnetwork of the independent tower network, and perform recommendation target prediction through the second subnetwork of the independent tower network, and the process may be implemented as follows: Inputting the user attribute information and the material attribute information into a shared network shared by each recommendation target to perform feature extraction, and obtaining a second user feature and a second material feature corresponding to each recommendation target; Inputting the second user features and the second material features corresponding to each recommendation target into the first sub-network of the independent tower network corresponding to each recommendation target respectively for feature extraction, thereby obtaining the first user features and the first material features corresponding to each recommendation target; The first user feature and the first material feature corresponding to each recommendation target are respectively input into the second sub-network of the independent tower network corresponding to each recommendation target to perform recommendation target prediction, and obtain the prediction value of the candidate material corresponding to each recommendation target.

[0045] In this case, the recommendation information corresponding to each recommendation target in the output of each shared network is used as the input of the corresponding first sub-network (i.e., the independent tower network front network). That is to say, the input of each first sub-network comes from the output of multiple shared networks, and the output of each first sub-network is used as the input of the corresponding second sub-network (i.e., the independent tower network back network), and the output of the second sub-network is the predicted value of the corresponding recommendation target; schematically, when the number of shared networks is 3 and the number of recommendation targets is 3 and they are click-through rate, interaction rate and consumption time respectively, the task network can include shared network 1, shared network 2 and shared network 3 shared by the three recommendation targets, and independent tower networks corresponding to the three recommendation targets respectively. 1. Independent tower network 2 and independent tower network 3, each of which includes a first subnetwork and a second subnetwork; the inputs of the three shared networks are user attribute information and material attribute information, each shared network inputs the second user features and second material features corresponding to each recommendation target obtained after feature extraction into the first subnetwork of the corresponding independent tower network, each first subnetwork further extracts features based on the second user features and second material features received, and inputs the extracted first user features and first material features into the corresponding second subnetwork, each second subnetwork predicts the recommendation target based on the first user features and first material features received, and obtains the predicted values ​​of the corresponding recommendation targets.

[0046] Step 230, input the material attribute information of the candidate material and the predicted value of the candidate material corresponding to each recommendation target into the first fusion network of the recommendation model, perform nonlinear weighted fusion at the material granularity, and obtain the first fusion prediction value of the target user's user behavior on the candidate material.

[0047] When training the first fusion network, the material attribute information of the sample material is input into the first fusion network so that the first fusion network can learn the material deviation information, and the predicted values ​​of each recommended target are input into the first fusion network so that the first fusion network can learn and train a set of parameters, which is convenient for weighted fusion of the predicted values ​​of each recommended target and debiasing the material. When applied, the first fusion network performs nonlinear weighted fusion based on the input material attribute information and the predicted values ​​of the candidate materials corresponding to each recommended target through the trained network parameters to obtain the first fusion prediction value.

[0048] Step 240, input the first user features and the first material features corresponding to each recommendation target, and the predicted value of the candidate material corresponding to each recommendation target into the second fusion network for nonlinear weighted fusion at the user granularity to obtain the second fusion prediction value of the target user's user behavior on the candidate material; wherein the first user features and the first material features corresponding to each recommendation target are obtained after the task network extracts the features corresponding to each recommendation target from the user attribute information and the material attribute information.

[0049] Corresponding to different task network settings, the first user feature and the first material feature corresponding to each recommendation target can be feature information output after feature extraction by the shared network; or, the first user feature and the first material feature corresponding to each recommendation target are feature information output after feature extraction by the first subnetwork of the independent tower network.

[0050] Taking the first user feature and the first material feature as an example, which are feature information output after feature extraction by the first subnetwork of the independent tower network, the input of the second fusion network is the output of each first subnetwork and the output of each second subnetwork, and the output of the first subnetwork is the result of further refining the feature information output by the shared network; when training the first fusion network, by inputting the output result of the first subnetwork into the second fusion network, the second fusion network can learn the user and material information, and by inputting the output result of the second subnetwork into the second fusion network, the second fusion network can learn and train a set of parameters, which is convenient for weighted fusion of the predicted values ​​of each recommendation target.

[0051] When applied, the second fusion network performs nonlinear weighted fusion based on the input first user features, first material features and the predicted values ​​of the candidate materials corresponding to each recommendation target through the trained network parameters to obtain the second fusion prediction value.

[0052] Step 250: recommending candidate materials based on the first fused prediction value and the second fused prediction value.

[0053] In a possible implementation, the computer device may sort the candidate materials based on the first fusion prediction value or the second fusion prediction value as a sorting basis, thereby determining a recommended order for the candidate materials, and recommend the candidate materials according to the recommended order.

[0054] In another possible implementation, the computer device may determine the recommended order of the candidate materials based on the first fused prediction value and the second fused prediction value. The process may be implemented as follows: Performing weighted fusion on the first fusion prediction value and the second fusion prediction value to obtain a comprehensive prediction value of the candidate material; Determine the recommended order of candidate materials based on the comprehensive predicted values ​​of the candidate materials; Recommend candidate materials based on the recommendation order.

[0055] Among them, the first fusion prediction value and the second fusion prediction value have their own corresponding weights, and the weight value of each weight can be set based on actual needs, or the weight value of each weight can be dynamically adjusted based on the actual recommendation effect. This application does not impose any restrictions on this.

[0056] In summary, the material recommendation method provided in the embodiment of the present application, when recommending materials for multiple recommendation targets, predicts the recommendation targets based on the user attribute information of the target user and the material attribute information of the candidate materials through the task network of the recommendation model, and obtains the prediction values ​​of the candidate materials corresponding to each recommendation target. The material attribute information of the candidate materials and the prediction values ​​of the candidate materials corresponding to each recommendation target are nonlinearly weighted fused at the material granularity through the first fusion network of the recommendation model to obtain the first fused prediction value of the user behavior of the target user on the candidate materials. The first user feature of the target user, the first material feature of the candidate materials, and the prediction values ​​of the candidate materials corresponding to each recommendation target are nonlinearly weighted fused at the user granularity through the second fusion network of the recommendation model. The first fused prediction value and the second fused prediction value are combined to obtain a second fused prediction value of the target user's user behavior on the candidate material, and the candidate material is recommended based on the first fused prediction value and the second fused prediction value, so that the prediction values ​​of multiple recommendation targets are fused into corresponding fused prediction values ​​on the basis of material granularity and user granularity respectively through nonlinear weighting, so as to comprehensively consider the influence at different granularities; compared with the method of uniformly setting biased features to the same value, the method provided in the present application fully considers the correlation between material characteristics, user characteristics and multiple recommendation targets, maintains the consistency of input content during model training and model application, and eliminates the influence of biased features while ensuring the accuracy of model prediction through operations such as nonlinear weighted fusion, thereby improving the overall performance and recommendation effect of the recommendation system.

[0057] Since the user behavior towards materials will change over time, training samples can be collected during the model application process and the recommendation model can be updated. The training process of the recommendation model is the same as the updating process of the recommendation model. The embodiment of the present application takes the updating model of the recommendation model as an example to illustrate the training process of the recommendation model.

[0058] In the embodiment of the present application, different functional networks in the recommendation model can be trained separately, wherein the training process of the task network can be implemented as follows: Obtaining target user behavior information of the target user on the candidate material; the target user behavior information includes the behavior attribute values ​​of the target user on the candidate material corresponding to each recommendation target; Based on the behavior attribute values ​​of each recommended target in the target user behavior information and the predicted values ​​of the candidate materials corresponding to each recommended target, the parameters of the task network are updated; the task network includes: a shared network shared by each recommended target and an independent tower network corresponding to each recommended target.

[0059] The target user behavior information is the actual user behavior information of the target user on the candidate materials recommended to the user, such as whether the user clicks on the candidate material, the user's consumption time on the candidate material, whether the user interacts with the candidate material, etc. The target user behavior information can be used as the user behavior information label corresponding to the candidate material. When the user clicks on or interacts with the candidate material, the training sample corresponding to the candidate material is a positive sample, and the behavior attribute label corresponding to the click-through rate in the user behavior information label is 1, or the behavior attribute label corresponding to the interaction rate is 1; when the user does not click on or interact with the candidate material, the training sample corresponding to the candidate material is a negative sample, and the behavior attribute label corresponding to the click-through rate in the user behavior information label is 0, or the behavior attribute label corresponding to the interaction rate is 0; since the user's consumption time on the material is a value greater than 0 and the unit is seconds, in order to unify the data dimension, in an embodiment of the present application, the log function can be used to compress the original consumption time, and the compressed consumption time value is used as the behavior attribute label for the user's consumption time on the candidate material.

[0060] When updating the parameters of the independent tower network and / or shared network in the task network, the computer device can perform corresponding loss function calculations based on the behavioral attribute labels of each recommended target and the predicted values ​​of each recommended target output by the task network, obtain the function value of the loss function corresponding to each recommended target, and update the parameters of the task network based on the function value of the loss function corresponding to each recommended target.

[0061] Among them, when the parameters of the task network are updated based on the function value of the loss function corresponding to each recommended target, in a possible implementation method, the computer device can calculate the update gradient of the corresponding recommended target based on the function value of the loss function corresponding to each recommended target, so as to update the parameters of the corresponding independent tower network and the shared network shared by each recommended target. Schematically, after the function value 1 of the loss function is calculated based on the predicted value of recommended target 1 and the behavioral attribute label 1 corresponding to recommended target 1, the update gradient of recommended target 1 is calculated, and the network parameters of the independent tower network corresponding to recommended target 1 in the task network and the shared network shared by each recommended target are updated through gradient backpropagation.

[0062] In another possible implementation, the computer device may perform weighted summation on the function values ​​of the loss function corresponding to each recommended target to obtain the function value of the comprehensive loss function, calculate the update gradient based on the function value of the comprehensive loss function, and update the parameters of each independent tower network and / or each shared network in the task network through the update gradient to train the task network, wherein, in the update gradient obtained by calculating the function value of the comprehensive loss function, the gradient corresponding to each recommended target acts on the corresponding independent tower network respectively, and the gradient of each recommended target acts together on each shared network.

[0063] The process of training each fusion network can be implemented as follows: Obtaining target user behavior information of the target user on the candidate material; the target user behavior information includes the behavior attribute values ​​of the target user on the candidate material corresponding to each recommendation target; Based on the target user behavior information and the fusion prediction value, the parameters of the corresponding fusion network are updated; wherein, when the fusion prediction value is the first fusion prediction value, the fusion network is the first fusion network; when the fusion prediction value is the second fusion prediction value, the fusion network is the second fusion network.

[0064] That is to say, when the computer device updates the first fusion network based on the training samples corresponding to the candidate materials, the computer device calculates the function value of the first loss function based on the target user behavior information and the first fusion prediction value, so as to update the first fusion network based on the function value of the first loss function; when the computer device updates the second fusion network based on the training samples corresponding to the candidate materials, the computer device calculates the function value of the second loss function based on the target user behavior information and the second fusion prediction value, so as to update the second fusion network based on the function value of the second loss function.

[0065] Among them, based on the target user behavior information and the fusion prediction value, the process of updating the parameters of the corresponding fusion network can be implemented as follows: Obtain multiple user behavior information corresponding to the candidate materials; Based on multiple user behavior information corresponding to the candidate materials, calculate the mean and variance of the behavior attribute values ​​of the candidate materials corresponding to each recommendation target; The loss function is calculated based on the behavior attribute value of each recommended target in the target user behavior information, the mean and variance value of the behavior attribute value of the candidate material corresponding to each recommended target, and the fusion prediction value to obtain the function value of the loss function; Based on the function value of the loss function, the parameters of the corresponding fusion network are updated.

[0066] Among them, the multiple user behavior information can be the user behavior information generated in the historical recommendation process corresponding to the candidate material, and the multiple user behavior information includes the target user behavior information, so as to calculate the mean and variance value of the behavior attribute values ​​of the candidate material corresponding to each recommendation target based on the multiple user behavior information; schematically, taking the number of recommendation targets as 3 as an example, the loss function for updating the fusion network can be expressed as:

[0067] in, That is the loss value of the loss function, is the predicted value of the fusion network, It is behavior attribute value, that is, the behavior attribute value of the i-th recommended target in the target user behavior information, It is The mean of the behavioral attribute values, It is The variance of the behavior attribute values.

[0068] The above loss function is used to standardize classification problems (such as whether to click or interact) and regression problems (i.e., consumption time) for easy unified calculation and solves the problem of inconsistent dimensions of different recommendation targets. The gradient is calculated through the function value of the loss function, and the parameters of the corresponding fusion network are updated after the gradient is returned to realize the training of each fusion network.

[0069] Take the recommendation targets as click-through rate, interaction rate and consumption time, and the independent tower network includes the first sub-network and the second sub-network as an example. Figure 3 A schematic diagram of a recommendation model provided by an exemplary embodiment of the present application is shown. Figure 3 As shown, the recommendation model is composed of a task network 310, a first fusion network 320 and a second fusion network 330, wherein the task network 310 includes an independent tower network corresponding to each recommendation target and a shared network shared by each recommendation target; the process of model training can be implemented as follows: Obtain a training sample set, which includes multiple training samples, each of which includes user attribute information of a sample user, material attribute information of a sample material, and a user behavior information label of the sample material, wherein the user behavior label includes a behavior attribute label of the sample user for each recommendation target of the sample material; The recommendation model is iteratively trained based on the training sample set until the training completion condition is met, and a trained recommendation model is obtained; during each iterative training process: Input the user attribute information of the sample user and the material attribute information of the sample material into the task network of the recommendation model to predict the recommendation target, and obtain the prediction value of the training sample corresponding to each recommendation target; Inputting the material attribute information of the sample material and the predicted value of the sample material corresponding to each recommendation target into the first fusion network of the recommendation model, performing nonlinear weighted fusion at the material granularity, and obtaining the first fusion prediction value of the user behavior of the sample user on the sample material; The first user features and the first material features corresponding to each recommendation target, as well as the predicted values ​​of the sample materials corresponding to each recommendation target, are input into the second fusion network for nonlinear weighted fusion at the user granularity to obtain the second fusion predicted value of the user behavior of the sample user on the sample material; wherein the first user features and the first material features corresponding to each recommendation target are obtained after the task network extracts the features corresponding to each recommendation target from the user attribute information of the sample user and the material attribute information of the sample material; Based on the prediction values ​​of the training samples corresponding to the recommended targets, the first fusion prediction values, the second fusion prediction values ​​and the user behavior information labels, the parameters of each functional network of the recommendation model are updated.

[0070] When the task network predicts the recommendation target based on the user attribute information of the sample user and the material attribute information of the sample material, the shared network corresponding to each recommendation target is used to extract features, and the second user features and second material features corresponding to each recommendation target are output by each shared network respectively; Input the second user features and the second material features corresponding to each recommendation target respectively output by each shared network into the first sub-network of the independent tower network corresponding to each recommendation target respectively for feature extraction, so as to obtain the first user features and the first material features corresponding to each recommendation target; The first user feature and the first material feature corresponding to each recommendation target are respectively input into the second sub-network of the independent tower network corresponding to each recommendation target to perform recommendation target prediction, and obtain the prediction value of the sample material corresponding to each recommendation target.

[0071] Among them, the process of updating the parameters of each functional network of the recommendation model based on the prediction value of the training samples corresponding to each recommendation target, the first fusion prediction value, the second fusion prediction value and the user behavior information label can refer to the relevant content of model updating based on the training samples corresponding to the candidate materials, which will not be repeated here.

[0072] By using the above-mentioned model training method to train the model, the recommendation model obtained by training can perform feature learning and feature fusion from both the user granularity and the material granularity, fully considering user needs and material characteristics, while improving the prediction accuracy, it can reduce the impact of material characteristics and user individual differences on the prediction results, thereby improving the overall performance and recommendation effect of the recommendation system.

[0073] Figure 4 FIG. 1 shows a block diagram of a material recommendation device provided by an exemplary embodiment of the present application. The device can be applied to a computer device to perform the following steps: Figure 1 or Figure 2 For all or part of the steps of the embodiment shown, the computer device can be implemented as a server or a terminal, such as Figure 4 As shown, the device comprises: The information acquisition module 410 is used to acquire the user attribute information of the target user and the material attribute information of the candidate material; The prediction module 420 is used to input the user attribute information of the target user and the material attribute information of the candidate material into the task network of the recommendation model to perform recommendation target prediction, and obtain the prediction value of the candidate material corresponding to each recommendation target; the prediction value of each recommendation target is used to represent the prediction information of the user behavior of the target user on the candidate material corresponding to the recommendation target; The first fusion module 430 is used to input the material attribute information of the candidate material and the predicted value of the candidate material corresponding to each recommendation target into the first fusion network of the recommendation model to perform nonlinear weighted fusion at the material granularity to obtain the first fusion prediction value of the user behavior of the target user on the candidate material; The material recommendation module 440 is used to recommend the candidate material based on the first fusion prediction value.

[0074] In a possible implementation manner, the recommendation model further includes a second fusion network, and the device further includes: The second fusion module is used to input the first user feature and the first material feature corresponding to each recommendation target, and the predicted value of the candidate material corresponding to each recommendation target into the second fusion network to perform nonlinear weighted fusion at the user granularity to obtain the second fusion prediction value of the user behavior of the target user on the candidate material; wherein the first user feature and the first material feature corresponding to each recommendation target are obtained after the task network extracts the features corresponding to each recommendation target from the user attribute information and the material attribute information; The material recommendation module 440 is used to recommend the candidate material based on the first fusion prediction value and the second fusion prediction value.

[0075] In a possible implementation, the material recommendation module 440 is used to: Performing weighted fusion on the first fused prediction value and the second fused prediction value to obtain a comprehensive prediction value of the candidate material; Determining a recommended order of the candidate materials based on the comprehensive predicted values ​​of the candidate materials; The candidate materials are recommended based on the recommendation order.

[0076] In a possible implementation manner, the device further includes: A behavior information acquisition module, used to acquire the target user behavior information of the target user on the candidate material; the target user behavior information includes the behavior attribute values ​​of the target user on the candidate material corresponding to each recommendation target; The first updating module is used to update the parameters of the corresponding fusion network based on the target user behavior information and the fusion prediction value; wherein, when the fusion prediction value is the first fusion prediction value, the fusion network is the first fusion network; when the fusion prediction value is the second fusion prediction value, the fusion network is the second fusion network.

[0077] In a possible implementation, the first updating module is used to: Acquire multiple user behavior information corresponding to the candidate material; Based on the multiple user behavior information corresponding to the candidate materials, calculating the mean and variance of the behavior attribute values ​​of the candidate materials corresponding to each recommendation target; A loss function is calculated based on the behavior attribute value of each recommended target in the target user behavior information, the mean and variance value of the behavior attribute value of the candidate material corresponding to each recommended target, and the fusion prediction value to obtain a function value of the loss function; Based on the function value of the loss function, parameters of the corresponding fusion network are updated.

[0078] In a possible implementation, the task network includes a shared network shared by each recommendation target and an independent tower network corresponding to each recommendation target; the prediction module 420 is used to: Inputting the user attribute information and the material attribute information into a shared network shared by each recommendation target to perform feature extraction, and obtaining a first user feature and a first material feature corresponding to each recommendation target; The first user feature and the first material feature corresponding to each recommendation target are respectively input into the independent tower network corresponding to each recommendation target to perform recommendation target prediction, so as to obtain the prediction value of the candidate material corresponding to each recommendation target.

[0079] In a possible implementation, the task network includes a shared network shared by each recommendation target and an independent tower network corresponding to each recommendation target; the independent tower network includes a first sub-network and a second sub-network; the prediction module 420 is used to: Inputting the user attribute information and the material attribute information into a shared network shared by each recommendation target to perform feature extraction, and obtaining a second user feature and a second material feature corresponding to each recommendation target; Inputting the second user features and the second material features corresponding to each recommendation target into the first sub-network of the independent tower network corresponding to each recommendation target respectively for feature extraction, thereby obtaining the first user features and the first material features corresponding to each recommendation target; The first user feature and the first material feature corresponding to each recommendation target are respectively input into the second subnetwork of the independent tower network corresponding to each recommendation target to perform recommendation target prediction, so as to obtain the prediction value of the candidate material corresponding to each recommendation target.

[0080] In a possible implementation manner, the device further includes: The second updating module is used to update the parameters of the task network based on the behavior attribute value of each recommended target in the target user behavior information obtained by the behavior information obtaining module and the predicted value of the candidate material corresponding to each recommended target.

[0081] In summary, the material recommendation device provided in the embodiment of the present application, when recommending materials for multiple recommendation targets, predicts the recommendation targets based on the user attribute information of the target user and the material attribute information of the candidate materials through the task network of the recommendation model, obtains the prediction values ​​of the candidate materials corresponding to each recommendation target, and performs nonlinear weighted fusion on the material attribute information of the candidate materials and the prediction values ​​of the candidate materials corresponding to each recommendation target through the first fusion network of the recommendation model to obtain the first fused prediction value corresponding to the candidate materials, so as to recommend the candidate materials based on the first fused prediction value; thereby, on the basis of the material granularity, the prediction values ​​of multiple recommendation targets are fused into one fused prediction value through a nonlinear weighted method; compared with the method of uniformly setting biased features to the same value, the device provided in the present application fully considers the correlation between material characteristics and multiple recommendation targets, maintains the consistency of the input content during model training and model application, and reduces the influence of biased features through operations such as nonlinear weighted fusion, while ensuring the accuracy of model prediction, thereby improving the overall performance and recommendation effect of the recommendation system.

[0082] Figure 5The structural block diagram of a computer device 500 shown in an exemplary embodiment of the present application is shown. The computer device can be implemented as the server in the above-mentioned solution of the present application. The computer device 500 includes a central processing unit (CPU) 501, a system memory 504 including a random access memory (RAM) 502 and a read-only memory (ROM) 503, and a system bus 505 connecting the system memory 504 and the central processing unit 501. The computer device 500 also includes a large-capacity storage device 506 for storing an operating system 509, an application program 510 and other program modules 511.

[0083] According to various embodiments of the present application, the computer device 500 can also be connected to a remote computer on the network through a network such as the Internet. That is, the computer device 500 can be connected to the network 508 through the network interface unit 507 connected to the system bus 505, or the network interface unit 507 can be used to connect to other types of networks or remote computer systems (not shown).

[0084] The memory also includes at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is stored in the memory. The central processing unit 501 implements all or part of the steps in the material recommendation method shown in the above-mentioned embodiments by executing the at least one instruction, at least one program, code set or instruction set.

[0085] Figure 6 FIG. 6 is a block diagram of a computer device 600 according to another exemplary embodiment of the present application. The computer device 600 may be implemented as the terminal described above.

[0086] Typically, the computer device 600 includes a processor 601 and a memory 602 .

[0087] In some embodiments, the computer device 600 may further optionally include: a peripheral device interface 603 and at least one peripheral device. The processor 601, the memory 602 and the peripheral device interface 603 may be connected via a bus or a signal line. Each peripheral device may be connected to the peripheral device interface 603 via a bus, a signal line or a circuit board. Specifically, the peripheral device includes: at least one of a radio frequency circuit 604, a display screen 605, a camera assembly 606, an audio circuit 607 and a power supply 608.

[0088] In some embodiments, the computer device 600 further includes one or more sensors 609 , including but not limited to: an acceleration sensor 610 , a gyroscope sensor 611 , a pressure sensor 612 , an optical sensor 613 , and a proximity sensor 614 .

[0089] Those skilled in the art will understand that Figure 6 The structure shown in the figure does not constitute a limitation on the computer device 600, and the computer device 600 may include more or less components than those shown in the figure, or combine some components, or adopt a different arrangement of components.

[0090] In an exemplary embodiment, a computer-readable storage medium is also provided, in which at least one computer program is stored, and the computer program is loaded and executed by a processor to implement all or part of the steps in the above-mentioned material recommendation method. For example, the computer-readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, etc.

[0091] In an exemplary embodiment, a computer program product is also provided, the computer program product comprising at least one computer program, the computer program being loaded and executed by a processor Figure 1 or Figure 2 All or part of the steps of the material recommendation method shown in any embodiment.

[0092] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0093] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A material recommendation method, characterized in that: The method comprises: Obtain user attribute information of the target user and material attribute information of the candidate materials; Inputting the user attribute information of the target user and the material attribute information of the candidate material into the task network of the recommendation model to predict the recommendation target, and obtaining the prediction value of the candidate material corresponding to each recommendation target; the prediction value of each recommendation target is used to represent the prediction information of the user behavior of the target user on the candidate material corresponding to the recommendation target; Inputting the material attribute information of the candidate material and the predicted value of the candidate material corresponding to each recommendation target into the first fusion network of the recommendation model to perform nonlinear weighted fusion at the material granularity to obtain the first fusion prediction value of the target user's user behavior on the candidate material; The candidate material is recommended based on the first fused prediction value.

2. The method according to claim 1, characterized in that The recommendation model further includes a second fusion network, and the recommending the candidate material based on the first fusion prediction value includes: Inputting the first user feature and the first material feature corresponding to each recommendation target, and the predicted value of the candidate material corresponding to each recommendation target into the second fusion network for nonlinear weighted fusion at the user granularity, to obtain the second fusion prediction value of the user behavior of the target user on the candidate material; wherein the first user feature and the first material feature corresponding to each recommendation target are obtained after the task network extracts the features corresponding to each recommendation target from the user attribute information and the material attribute information; The candidate material is recommended based on the first fused prediction value and the second fused prediction value.

3. The method according to claim 2, characterized in that The recommending the candidate material based on the first fused prediction value and the second fused prediction value includes: Performing weighted fusion on the first fused prediction value and the second fused prediction value to obtain a comprehensive prediction value of the candidate material; Determining a recommended order of the candidate materials based on the comprehensive predicted values ​​of the candidate materials; The candidate materials are recommended based on the recommendation order.

4. The method according to claim 2, characterized in that: The method further comprises: Acquire target user behavior information of the target user on the candidate material; the target user behavior information includes behavior attribute values ​​of the target user on the candidate material corresponding to each recommendation target; Based on the target user behavior information and the fusion prediction value, the parameters of the corresponding fusion network are updated; wherein, when the fusion prediction value is the first fusion prediction value, the fusion network is the first fusion network; when the fusion prediction value is the second fusion prediction value, the fusion network is the second fusion network.

5. The method according to claim 4, characterized in that The updating of parameters of the corresponding fusion network based on the target user behavior information and the fusion prediction value includes: Acquire multiple user behavior information corresponding to the candidate material; Based on the multiple user behavior information corresponding to the candidate materials, calculating the mean and variance of the behavior attribute values ​​of the candidate materials corresponding to each recommendation target; A loss function is calculated based on the behavior attribute value of each recommended target in the target user behavior information, the mean and variance value of the behavior attribute value of the candidate material corresponding to each recommended target, and the fusion prediction value to obtain a function value of the loss function; Based on the function value of the loss function, parameters of the corresponding fusion network are updated.

6. The method according to claim 2, characterized in that The task network includes a shared network shared by each recommended target and an independent tower network corresponding to each recommended target; The step of inputting the user attribute information of the target user and the material attribute information of the candidate material into the task network of the recommendation model to perform recommendation target prediction, and obtaining the prediction value of the candidate material corresponding to each recommendation target, comprises: Inputting the user attribute information and the material attribute information into a shared network shared by each recommendation target to perform feature extraction, and obtaining a first user feature and a first material feature corresponding to each recommendation target; The first user feature and the first material feature corresponding to each recommendation target are respectively input into the independent tower network corresponding to each recommendation target to perform recommendation target prediction, so as to obtain the prediction value of the candidate material corresponding to each recommendation target.

7. The method according to claim 2, characterized in that The task network includes a shared network shared by each recommended target and an independent tower network corresponding to each recommended target; the independent tower network includes a first sub-network and a second sub-network; The step of inputting the user attribute information of the target user and the material attribute information of the candidate material into the task network of the recommendation model to perform recommendation target prediction, and obtaining the prediction value of the candidate material corresponding to each recommendation target, comprises: Inputting the user attribute information and the material attribute information into a shared network shared by each recommendation target to perform feature extraction, and obtaining a second user feature and a second material feature corresponding to each recommendation target; Inputting the second user features and the second material features corresponding to each recommendation target into the first sub-network of the independent tower network corresponding to each recommendation target respectively for feature extraction, thereby obtaining the first user features and the first material features corresponding to each recommendation target; The first user feature and the first material feature corresponding to each recommendation target are respectively input into the second subnetwork of the independent tower network corresponding to each recommendation target to perform recommendation target prediction, so as to obtain the prediction value of the candidate material corresponding to each recommendation target.

8. The method according to claim 6 or 7, characterized in that: The method further comprises: Acquire target user behavior information of the target user on the candidate material; the target user behavior information includes behavior attribute values ​​of the target user on the candidate material corresponding to each recommendation target; Based on the behavior attribute value of each recommended target in the target user behavior information and the predicted value of the candidate material corresponding to each recommended target, the parameters of the task network are updated.

9. A material recommendation device, characterized in that: The device comprises: An information acquisition module is used to acquire user attribute information of a target user and material attribute information of a candidate material; A prediction module is used to input the user attribute information of the target user and the material attribute information of the candidate material into the task network of the recommendation model to predict the recommendation target, and obtain the prediction value of the candidate material corresponding to each recommendation target; the prediction value of each recommendation target is used to represent the prediction information of the user behavior of the target user on the candidate material corresponding to the recommendation target; A first fusion module is used to input the material attribute information of the candidate material and the predicted value of the candidate material corresponding to each recommendation target into the first fusion network of the recommendation model to perform nonlinear weighted fusion at the material granularity to obtain a first fusion prediction value of the user behavior of the target user on the candidate material; A material recommendation module is used to recommend the candidate material based on the first fusion prediction value.

10. A computer device, characterized in that: The computer device includes a processor and a memory, the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the material recommendation method according to any one of claims 1 to 8.

Images