Video recommendation model training method, processing equipment and storage medium

Abstract

The embodiment of the invention discloses a training method of a video recommendation model, processing equipment and a storage medium. The method comprises the following steps: in a second training stage, using a first sample set to carry out iterative training on a fine arrangement model after initial training, and adjusting parameters of the fine arrangement model based on a first loss value; and performing iterative training on the recall model after initial training by using a second sample set, and adjusting parameters of the recall model based on the third loss value and the second loss value, and obtaining a trained video recommendation model based on the final trained fine arrangement model and recall model until a third convergence condition is met. According to the embodiment of the invention, the prediction accuracy of the trained model can be improved, and better video recommendation experience is brought to the user.

Description

Technical Field

[0001] The present invention relates to but is not limited to the field of video recommendation, and in particular to a training method, processing device and storage medium for a video recommendation model. Background Art

[0002] With the rapid development of artificial intelligence technology, video recommendation technology can recommend videos that users like to users. For video recommendation scenarios, the model training platform will train the target models of completion and duration based on a large number of users to fit the interests of users from different dimensions. When providing recommendation services to users, the recommendation system engine will go through the stages of recall (for example, obtaining thousands of materials from millions of materials), rough sorting (for example, retaining thousands of materials after model screening), fine sorting (for example, scoring and sorting thousands of materials) and re-sorting (for example, selecting the top 10 materials based on diversity and various strategies), and finally display the materials that users may be interested in to users. However, the trained model still has the problem of inaccurate prediction. Summary of the invention

[0003] In view of this, the embodiments of the present disclosure disclose a training method, processing device and storage medium for a video recommendation model, so as to at least achieve: compared with the training method of the recommendation model that does not consider the prediction results of the refined ranking model when training the recall model, it is possible to improve the prediction accuracy of the trained model, achieve recommendations that are more in line with user behavior, and bring users a better experience of recommending videos.

[0004] According to a first aspect of an embodiment of the present disclosure, a method for training a video recommendation model is provided, the method comprising:

[0005] In the first training stage, the initial refined ranking model in the video recommendation model is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material;

[0006] In the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after initial training; wherein the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: material features and user features;

[0007] In the second training stage, the first sample set is used to iteratively train the refined ranking model after the initial training, and the parameters of the refined ranking model are adjusted based on the first loss value; and the second sample set is used to iteratively train the recall model after the initial training, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the finally trained refined ranking model and the recall model; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

[0008] According to a second aspect of an embodiment of the present disclosure, a training device for a video recommendation model is provided, the device comprising a training module, and the training module is configured to:

[0009] In the first training stage, the initial refined ranking model in the video recommendation model is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material;

[0010] In the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after initial training; wherein the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: the characteristics of the material and the characteristics of the user;

[0011] In the second training stage, the first sample set is used to iteratively train the refined ranking model after the initial training, and the parameters of the refined ranking model are adjusted based on the first loss value; and the second sample set is used to iteratively train the recall model after the initial training, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the finally trained refined ranking model and the recall model; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

[0012] According to a third aspect of an embodiment of the present disclosure, a processing device is provided, the processing device comprising:

[0013] A memory for storing executable programs;

[0014] The processor is used to implement any method described in the embodiments of the present disclosure when executing the executable program stored in the memory.

[0015] According to a fourth aspect of the embodiments of the present disclosure, a computer storage medium is provided, wherein the computer storage medium stores an executable program, and when the executable program is executed by a processor, the method described in any one of the embodiments of the present disclosure is implemented.

[0016] In the disclosed embodiment, in the first training stage, the first sample set is used to iteratively train the initial refined ranking model in the video recommendation model, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein, the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: the characteristics of the material, the characteristics of the user, and the fusion characteristics of the user and the material. In this way, the refined ranking model after initial training obtained by training based on the first loss value and the first convergence condition in the first training stage can be directly used for continued training in the second training stage, thereby shortening the duration of continued training of the refined ranking model in the second stage and improving training efficiency; and the first prediction value of the refined ranking model is applied in the second training stage to realize the transfer of knowledge of the refined ranking model.

[0017] In the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after initial training; wherein the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: the characteristics of the material and the characteristics of the user. In this way, the recall model after initial training obtained by training based on the second loss value and the second convergence condition in the first training stage can be directly used for continued training in the second training stage, thereby shortening the time for continued training of the recall model in the second training stage and improving the overall training efficiency of the video recommendation model.

[0018] In the second training stage, the first sample set is used to iteratively train the refined ranking model after initial training, and the parameters of the refined ranking model are adjusted based on the first loss value; and the second sample set is used to iteratively train the recall model after initial training, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the refined ranking model and the recall model after final training; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model. In this way, in the second training stage, since the parameters of the recall model are adjusted based on the third loss value and the second loss value until the third convergence condition is met, and since the third loss value is determined based on the training result of the refined ranking model, the knowledge transfer from the refined ranking model to the recall model is essentially realized, so that the prediction result of the trained recall model will be more accurate, thereby improving the accuracy of the video recommendation model in executing the prediction. Compared with the training method of the recommendation model without considering the training result of the refined ranking model, the prediction accuracy of the trained model can be improved, and recommendations that are more in line with user behavior can be achieved, bringing users a better experience of recommended videos. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a schematic diagram of a double-tower recall model according to an exemplary embodiment;

[0020] Figure 2 is a flowchart of a video recommendation method according to an exemplary embodiment;

[0021] Figure 3 is a schematic diagram of a video recommendation method according to an exemplary embodiment;

[0022] Figure 4 is a schematic diagram of a distillation model according to an exemplary embodiment;

[0023] Figure 5 is a schematic diagram of an online and offline part according to an exemplary embodiment;

[0024] Figure 6 The figure is a schematic diagram of a video recommendation model training device according to an exemplary embodiment. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present invention. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.

[0026] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0027] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein are only for the purpose of describing the embodiments of the present invention and are not intended to limit the present invention.

[0029] In order to better understand the embodiments of the present disclosure, the following describes the scenarios in the related art through exemplary embodiments:

[0030] In some embodiments, in the recall phase, the user vector and the material vector are mostly constructed based on the double-tower structure in the candidate generation model of deep learning. During training, the user features are input into the user tower to obtain the user vector, and the item features are input into the material tower to obtain the material vector. After the training is completed, the vectors of all materials are calculated and stored in the material vector library. Based on the artificial neural network (ANN, Artificial Neural Network) algorithm, the material index is constructed.

[0031] In some embodiments, for a given user vector during online prediction, the user vector and the material vector library retrieve material data (e.g., the faiss library) and return the top K (Top K) materials closest to the user vector as candidate materials to the rough ranking stage.

[0032] In some embodiments, a dual-tower model is constructed separately for the completion and duration targets during training. During online prediction, the user's completion vector and duration vector are concatenated together and the concatenated vector is used to retrieve the Top K candidate materials closest to the user vector from the material's faiss library. The most critical part of this process is the proportional fusion of the user's completion vector and the user's duration vector, which requires a large number of online experiments to verify and determine which fusion formula is better designed.

[0033] For example: user vector = a×duration vector+b×completion vector. Parameters a and b represent the weights of duration and completion respectively, which are fixed values, and their magnitudes express the inclination towards different targets.

[0034] In some embodiments, see Figure 1 , shows a dual-tower recall model structure, which includes a first tower model (corresponding to a material tower model) and a second tower model (corresponding to a user tower model). The first tower model is used to process material features, and the second tower model is used to process user features.

[0035] In some embodiments, in the above dual-tower model structure, the user tower and the material tower are modeled separately, and the user and material features are not fully interacted. To extract effective feature information, only the dot product operation is performed on the user vector and the material vector in the final stage, which lacks more effective feature interaction. In addition, the dual-tower structure does not use the cross features (or fusion features) of users and materials, sacrificing more accurate precision in exchange for the speed of online recall.

[0036] In some embodiments, to address the above-mentioned deficiencies, knowledge distillation techniques may be used to transfer the knowledge of the refinement model to the recall model.

[0037] In some embodiments, the refined ranking model can process all features, including user features, material features, cross-features of users and materials (or fusion features), etc., and can extract effective features in each hidden layer to predict the final target.

[0038] In some embodiments, in addition to the original hard labels (completed broadcast labels {0, 1}, duration labels) of the dual-tower recall model, the knowledge extracted from the refined ranking model (i.e., soft labels) can also be used to supervise the training of the recall model, which further improves the performance of the recall model.

[0039] In some embodiments, when serving online, only the recall model part is used, and cross features are not relied on as input, which ensures the consistency of recall model training and prediction. On the one hand, cross features are combined in a way that is more suitable for the prediction task. Generally speaking, adding more cross features will achieve more accurate predictions. On the other hand, no matter how many cross features there are in the material, the recall model structure will only introduce an additional distillation loss, which is easier to balance.

[0040] like Figure 2 As shown, a method for training a video recommendation model is provided in an embodiment of the present disclosure, and the method includes:

[0041] Step S21: in a first training phase, using a first sample set to iteratively train an initial refined ranking model in a video recommendation model, adjusting parameters of the refined ranking model based on a first loss value until the number of training times meets a first convergence condition, thereby obtaining a refined ranking model after initial training;

[0042] Step S22: in the first training phase, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after the initial training;

[0043] Step S23, in the second training stage, use the first sample set to iteratively train the refined ranking model after the initial training, and adjust the parameters of the refined ranking model based on the first loss value; and use the second sample set to iteratively train the recall model after the initial training, and adjust the parameters of the recall model based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the refined ranking model and the recall model after the final training.

[0044] The video recommendation model training method disclosed in the present invention can be applied to a terminal. The terminal involved in the present invention may be, but is not limited to, a computer, a mobile phone, a wearable device, a vehicle-mounted terminal, a road side unit (RSU, Road Side Unit), a smart home terminal, an industrial sensor device and / or a medical device, etc.

[0045] It should be noted that the above method corresponds to the training process of the model for initial video recommendation, and the training process can be performed on a dedicated device. After training the initial video recommendation model, the trained video recommendation model can be obtained, and the trained video recommendation model can be transplanted to an electronic device for video recommendation model application, and the trained video recommendation model is applied. Of course, the training process of the initial video recommendation model and the trained video recommendation model can also be performed on the same electronic device, which is not limited here. When the training process of the initial video recommendation model is performed on a dedicated device, after the training of the initial video recommendation model is completed, the trained video recommendation model can be transplanted from the dedicated device to the electronic device that applies the model for execution.

[0046] In some embodiments, the refined ranking model is used to predict the user's interest level in the material for the first sample in the input first sample set, and output a first prediction value.

[0047] In some embodiments, the first loss value is determined based on a cross entropy between the first predicted value and a preset first target value.

[0048] In some embodiments, the first target value may be a preset value, and the first target value may be determined according to training requirements.

[0049] In some embodiments, the first convergence condition may be the number of training times, for example, 5000 times.

[0050] In some embodiments, the first convergence condition may be the training duration, for example, 30 minutes.

[0051] In some embodiments, the first convergence condition may be that the corresponding convergence function converges.

[0052] In some embodiments, the first sample includes: material features, user features, and fusion features of the user and the material.

[0053] In some embodiments, the material features may be features of a video sample, such as a duration feature and / or a completion feature, etc. The material may be understood as a video sample.

[0054] In some embodiments, the user feature may be a feature associated with user behavior. For example, the feature of a user performing a predetermined operation, which may be an operation of a user sliding a video, a user commenting on a video, a user clicking to vote for a video, a user liking a video, a user collecting a video, a user sharing a video, a user forwarding a video, and / or a user clicking to enter a homepage, etc., without limitation herein.

[0055] In some embodiments, the fusion feature of the user and the material may be a feature obtained by fusing the user feature with the material feature, and may also be referred to as a cross feature of the user and the material.

[0056] In some embodiments, the recall model is used to predict whether the user is interested in the material for a second sample in the input second sample set, and output a second prediction value.

[0057] In some embodiments, the second loss value is determined based on a cross entropy between the second predicted value and a preset second target value.

[0058] In some embodiments, the second target value may be a preset value, and the second target value may be determined according to training requirements.

[0059] In some embodiments, the second convergence condition may be at least one of the following:

[0060] The number of training times reaches a predetermined number (for example, it can be the same as the first convergence condition, both are 5000 times, but not limited thereto);

[0061] The training duration reaches the predetermined duration, for example, 1 hour;

[0062] The loss function converges.

[0063] In some embodiments, the second sample includes: material characteristics and user characteristics.

[0064] In some embodiments, the third loss value is determined based on a cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refinement model.

[0065] In some embodiments, the third convergence condition may be at least one of the following:

[0066] The training times reach the predetermined times;

[0067] The training duration reaches the scheduled duration;

[0068] The loss function converges.

[0069] In some embodiments, the positive samples in the first sample set and the positive samples in the second sample set are selected in a different manner.

[0070] In some embodiments, the negative samples in the first sample set and the negative samples in the second sample set are selected in a different manner.

[0071] In some embodiments, the first samples in the first sample set include positive samples and negative samples.

[0072] In some embodiments, a positive sample is a sample in which the user's interaction time with the material or the material's playback time is higher than a threshold.

[0073] In some embodiments, negative samples are samples in which the user's interaction time with the material or the material's playback time is lower than a threshold, and / or samples obtained based on probabilistic random negative sampling.

[0074] In some embodiments, the second samples in the second sample set include positive samples and negative samples.

[0075] In some embodiments, positive samples are samples in which the user's interaction time with the material or the material's playback time is higher than a threshold, and / or samples in which the material's playback time is lower than a threshold.

[0076] In some embodiments, negative samples are samples obtained based on probabilistic random sampling.

[0077] In some embodiments, the first feature set and the second feature set may share material features and / or user features.

[0078] In some embodiments, the refined ranking model and the recall model have shared sample features; the shared sample features are used by the refined ranking model and / or the recall model to correct the sample features in the corresponding sample set.

[0079] In some embodiments, the refined ranking model and the recall model have shared sample features; the shared sample features are sample features used in the training process of the refined ranking model; the shared sample features are used by the recall model to correct sample features in the corresponding sample set.

[0080] In some embodiments, the refined ranking model and the recall model have shared sample features; the shared sample features are sample features used in the training process of the recall model; the shared sample features are used by the refined ranking model to correct sample features in the corresponding sample set.

[0081] In some embodiments, the shared sample features are determined based on the prediction results of the refinement model and / or the recall model.

[0082] In some embodiments, the number of shared sample features determined by the refinement model is greater than the number of shared sample features determined by the recall model.

[0083] In some embodiments, in the first training stage, the initial refined ranking model in the video recommendation model is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material; in the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after initial training; Among them, the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: material characteristics and user characteristics; in the second training stage, the first sample set is used to iteratively train the refined ranking model after initial training, and the parameters of the refined ranking model are adjusted based on the first loss value; and, the second sample set is used to iteratively train the recall model after initial training, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the final trained refined ranking model and the recall model; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

[0084] In some embodiments, the refined model includes a fully connected neural network (DNN, Deep Neural Network) layer and an activation layer.

[0085] In some embodiments, in the first training stage, the material features, user features, and fusion features of the user and the material in the first sample are input into the DNN for feature crossover operation to obtain the first cross feature; the first cross feature is input into the activation layer for activation to obtain the first prediction value of the refined ranking model; the cross entropy is determined based on the first prediction value and the first target value; the first loss value of the refined ranking model is determined based on the cross entropy; in the first training stage, the initial refined ranking model in the video recommendation model is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output the first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and the preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material; in the first training stage, the video recommendation model is iteratively trained using the second sample set The initial recall model in the recommendation model is iteratively trained, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, so as to obtain the recall model after the initial training; wherein, the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output the second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: material characteristics and user characteristics; in the second training stage, the refined ranking model after the initial training is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value; and, the recall model after the initial training is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the refined ranking model and the recall model after the final training; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

[0086] In some embodiments, the recall model includes a first tower model and a second tower model.

[0087] In some embodiments, in the first training stage, the first sample set is used to iteratively train the initial refined ranking model in the video recommendation model, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and the preset first target value; the first sample includes: material characteristics, user characteristics, and fusion characteristics of the user and the material; in the first training stage, the material characteristics of the second sample are input into the first tower model for feature cross operation to obtain a second cross feature; the user characteristics of the second sample are input into the second tower model for feature cross operation to obtain a third cross feature; an inner product operation of the vector is performed on the second cross feature and the third cross feature to obtain a second prediction value; the cross entropy is determined based on the second prediction value and the second target value; the second loss value is determined based on the cross entropy. In the first training stage, the first sample set is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy. The two sample sets are used to iteratively train the initial recall model in the video recommendation model, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after the initial training; wherein, the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output the second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: material features and user features; in the second training stage, the first sample set is used to iteratively train the refined ranking model after the initial training, and the parameters of the refined ranking model are adjusted based on the first loss value; and, the second sample set is used to iteratively train the recall model after the initial training, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the refined ranking model and the recall model after the final training; wherein, the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

[0088] In some embodiments, in the first training stage, the material features, user features and fusion features of the user and the material in the first sample are input into the DNN for feature crossover operation to obtain the first cross feature; the first cross feature is input into the activation layer for activation to obtain the first prediction value of the refined ranking model; the cross entropy is determined based on the first prediction value and the first target value; the first loss value of the refined ranking model is determined based on the cross entropy; in the first training stage, the initial refined ranking model in the video recommendation model is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output the first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and the preset first target value; the first sample includes: material features, user features and fusion features of the user and the material; in the first training stage, the material features of the second sample are input into the first tower model for feature crossover operation to obtain the second cross feature; the user features of the second sample are input into the second tower model for feature crossover operation to obtain the third cross feature; for the second cross feature and the third cross The inner product operation of the cross feature vector is performed to obtain a second prediction value; the cross entropy is determined based on the second prediction value and the second target value; the second loss value is determined based on the cross entropy. In the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after initial training; wherein the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output the second prediction value; the second loss value is based on the cross entropy between the second prediction value and the preset second target value Determine; the second sample includes: material features and user features; in the second training stage, the first sample set is used to iteratively train the refined ranking model after the initial training, and the parameters of the refined ranking model are adjusted based on the first loss value; and, the second sample set is used to iteratively train the recall model after the initial training, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the finally trained refined ranking model and the recall model; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

[0089] In some embodiments, the weight of negative samples obtained based on probabilistic random negative sampling is reduced when determining the first loss value of the refined ranking model; in the first training stage, the initial refined ranking model in the video recommendation model is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material; in the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition. Convergence condition, obtain the recall model after initial training; wherein, the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: material characteristics and user characteristics; in the second training stage, the refined ranking model after initial training is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value; and, the recall model after initial training is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the refined ranking model and the recall model after final training; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

[0090] In some embodiments, the weight of positive samples based on the playback time of the material being lower than the threshold is reduced when determining the second loss value of the recall model; in the first training stage, the initial refined ranking model in the video recommendation model is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material; in the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition. Convergence condition, obtain the recall model after initial training; wherein, the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: material characteristics and user characteristics; in the second training stage, the refined ranking model after initial training is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value; and, the recall model after initial training is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the refined ranking model and the recall model after final training; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

[0091] In the disclosed embodiment, in the first training stage, the first sample set is used to iteratively train the initial refined ranking model in the video recommendation model, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein, the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material. In this way, the refined ranking model after initial training obtained by training based on the first loss value and the first convergence condition in the first training stage can be directly used for continued training in the second training stage, thereby shortening the duration of continued training of the refined ranking model in the second stage and improving training efficiency; and the first prediction value of the refined ranking model is applied in the second training stage to realize the transfer of knowledge of the refined ranking model. In the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after initial training; wherein the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: material features and user features. In this way, the recall model after initial training obtained by training based on the second loss value and the second convergence condition in the first training stage can be directly used for continued training in the second training stage, thereby shortening the time for continued training of the recall model in the second training stage and improving the overall training efficiency of the video recommendation model. In the second training stage, the first sample set is used to iteratively train the refined ranking model after the initial training, and the parameters of the refined ranking model are adjusted based on the first loss value; and the second sample set is used to iteratively train the recall model after the initial training, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the finally trained refined ranking model and the recall model; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.In this way, in the second training stage, since the parameters of the recall model will be adjusted based on the third loss value and the second loss value until the third convergence condition is met, and since the third loss value is determined based on the training results of the refined ranking model, the knowledge transfer from the refined ranking model to the recall model is actually realized, so that the prediction results of the trained recall model will be more accurate, thereby improving the accuracy of the video recommendation model's execution prediction. Compared with the training method of the recommendation model that does not consider the training results of the refined ranking model, the prediction accuracy of the trained model can be improved, and recommendations that are more in line with user behavior can be achieved, bringing users a better experience of recommending videos.

[0092] It should be noted that those skilled in the art can understand that the method provided in the embodiments of the present disclosure can be executed alone or together with some methods in the embodiments of the present disclosure or some methods in related technologies.

[0093] like Figure 3 As shown, a method for video recommendation is provided in an embodiment of the present disclosure, and the method includes:

[0094] Step S31, obtaining multiple candidate video samples;

[0095] Step S32: input multiple candidate video samples into the trained video recommendation model to obtain recommended videos recommended to the user.

[0096] In some embodiments, in the first training stage, the first sample set is used to iteratively train the initial refined ranking model in the video recommendation model, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material; in the first training stage, the second sample set is used to iteratively train the initial recall model in the video recommendation model, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after initial training; wherein the recall model is used to predict the user's interest in the material for the second sample in the input second sample set Measure whether the user is interested in the material and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and a preset second target value; the second sample includes: material features and user features; in the second training stage, use the first sample set to iteratively train the refined ranking model after initial training, and adjust the parameters of the refined ranking model based on the first loss value; and, use the second sample set to iteratively train the recall model after initial training, and adjust the parameters of the recall model based on the third loss value and the second loss value; until the third convergence condition is met, a trained video recommendation model is obtained based on the final trained refined ranking model and the recall model; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model; obtain multiple candidate video samples; input the multiple candidate video samples into the trained video recommendation model to obtain recommended videos recommended to the user.

[0097] It should be noted that those skilled in the art can understand that the method provided in the embodiments of the present disclosure can be executed alone or together with some methods in the embodiments of the present disclosure or some methods in related technologies.

[0098] In order to better understand the embodiments of the present disclosure, the technical solution of the present disclosure is further described below through the following exemplary embodiments:

[0099] In some embodiments, the video recommendation process will go through the steps of recall, rough ranking and reranking in sequence. The technical solution of the present disclosure is implemented in the recall stage. Figure 4 , showing a schematic diagram of the structure of a knowledge distillation model. Based on the original dual-tower model recall, a refined ranking model is added in parallel, and the knowledge transfer is realized through the distillation loss (loss, corresponding to the third loss value in this disclosure), the link consistency of the recall and refined ranking models is achieved, and the ability of the dual-tower recall model is improved.

[0100] In some embodiments, the feature input part includes: user features, material features, and user-material cross-features (corresponding to the user-material fusion features disclosed in the present invention).

[0101] In some embodiments, for the dual-tower recall model part: user features are input into the user tower, and the features are cross-combined through multiple layers of DNN to obtain user embedding (corresponding to the second cross feature). Material features are input into the material tower, and the features are cross-combined through multiple layers of DNN to obtain material embedding (corresponding to the first cross feature). For the completion target (the playback threshold is greater than a certain value, which is 1, otherwise it is 0), the user embedding and the material embedding are inner-producted, and the predicted value is cross-entropy with the completion target to obtain the loss of the dual-tower recall model.

[0102] In some embodiments, for the refined ranking model part: user features, material features, and user-material cross features are input together into a multi-layer DNN for full feature crossover, the activation function of the last layer is a sigmoid function, the predicted value of the refined ranking model is input, and cross entropy is performed with the completion rate target to obtain the loss of the refined ranking model (corresponding to the first loss value).

[0103] In some embodiments, in the early stage of model training, the loss of refinement (corresponding to the first loss value) only updates the parameters of the refinement model, and the loss of recall (corresponding to the second loss value) only updates the parameters of the dual-tower recall model. In the later stage of model training, the distillation loss (corresponding to the third loss value) and the recall loss together update the parameters of the recall layer.

[0104] In some embodiments, in the recommendation system, the sample selection space of the refined ranking model is inconsistent with the sample selection space of the double-tower recall model.

[0105] In some embodiments, the positive samples of the refined ranking model are samples that have been exposed to interaction or whose playback time is higher than a threshold, and the negative samples are samples that have been exposed to playback time lower than a threshold.

[0106] In some embodiments, the positive samples of the dual-tower recall model are consistent with the positive samples of the fine ranking model, and the negative samples are samples randomly negatively sampled with probability from the entire material library.

[0107] Since the dual-tower recall model based on knowledge distillation needs to train both the refinement model and the recall model, it involves the design of the sample space. Therefore, the overall sample design space is designed as follows:

[0108] In some embodiments, the refined model sample space is:

[0109] Positive samples are: samples that have been exposed to interaction or whose playback time is higher than the threshold are positive samples;

[0110] Negative samples are: samples whose exposure duration is lower than the threshold are negative samples, and samples randomly sampled with probability are weak negative samples.

[0111] In some embodiments, the double-tower recall model sample space is:

[0112] Positive samples: samples that have been exposed to interaction or whose playback time is higher than the threshold are positive samples, and samples that have been exposed to playback time lower than the threshold are weak positive samples;

[0113] Negative samples: Samples randomly sampled with probability are negative samples.

[0114] In some embodiments, for the refined ranking model part, the newly added samples with probabilistic random negative sampling are weak negative samples. When calculating the sample loss, this batch of samples is downgraded (for example, the weight is set to 0.5), and the other parts remain unchanged.

[0115] In some embodiments, for the dual-tower recall model, newly added samples with exposure and playback time below the threshold are weak positive samples, and this batch of samples is downgraded (for example, the weight is set to 0.2), while other parts remain unchanged.

[0116] In some embodiments, please refer again to Figure 4 , the refined ranking model and the recall model share some features at the bottom layer, such as user features and item features. When the loss reverse gradient is derived, the question of which part of the underlying embedding corresponding to the user features and item features is updated involves selection. Since the refined ranking model is more fully trained and features are more fully crossed, the embedding is learned better. Therefore, the update strategy can be: the embeddings of user features and item features are mainly updated by the refined ranking loss and the recall loss, but the update amplitude of the refined ranking loss is greater than that of the recall part. In this way, the underlying embedding can be learned more fully.

[0117] In some embodiments, three parts of loss are designed in model training, namely, recall model loss, refinement model loss, and distillation loss (first loss value). During the training process, the variable λ is used to control whether the distillation loss participates in the parameter update of the recall model. Among them, the loss value loss is:

[0118]

[0119] Among them, L represents the cross entropy loss; y represents the true label; f(,) represents the prediction function; Ls, Ld, Lt represent the loss of the recall model, the distillation loss, and the loss of the refined ranking model respectively; X represents user features and material features; X* represents the cross-features of users and materials; W represents the model parameters.

[0120] In some embodiments, at the beginning of model training, the loss of refinement (Ls) only updates the parameters of the refinement model (Wt), and the loss of recall (Lt) only updates the parameters of the recall model (Wt). When the number of training rounds reaches a certain stage (for example, 5000 rounds), the parameters of the recall model (Wt) are composed of two parts, namely the loss of recall (Lt) and the distillation loss (Ld). The distillation loss is obtained by calculating the cross entropy between the predicted value of the refinement model (f(X, X*; Wt)) as the soft completion target of the recall model and the predicted value of the recall model f(X; Ws).

[0121] In some embodiments, see Figure 5 , each material in the material library will be Figure 2 The model recalls the material tower of the model, calculates the embedding vector of each material, and then uses the ANN algorithm to establish its index for each material.

[0122] In some embodiments, when an online user requests a connection, Figure 5 The model recalls the user tower of the model, calculates the embedding vector of the user, and then goes to the material's faiss library to retrieve the TopK materials closest to the user vector and returns them to the coarse ranking model.

[0123] In the above example, each indicator has improved, e.g.

[0124] Scale indicators: Recommended vv3: +0.76%, recommended time3: +0.65%;

[0125] Efficiency indicators: average vv3 per time: +0.15%, average time3 per time: +0.18%, average vv3 per person: +0.66%, average reading per person: +0.69%, average reading per time: +0.05%.

[0126] In the disclosed embodiment, distillation learning is used to transfer the knowledge of the refined ranking model to improve the performance of the recall model.

[0127] In the disclosed embodiment, the inconsistency problem between the refined ranking and the recalled sample spaces is fully utilized, and it is proposed to use the samples that have not been “interacted” with the exposure as weak positive samples of the recall model.

[0128] In the disclosed embodiment, the distillation loss is used to learn the interactive capabilities of features in the refined ranking model and extract high-order feature combinations.

[0129] In the disclosed embodiment, when the underlying embedding is updated, the fine ranking loss is used to lead its update, thereby effectively transferring knowledge to the recall model.

[0130] like Figure 6 As shown, in an embodiment of the present disclosure, a training device for a video recommendation model is provided, the device comprising a training module 61, and the training module 61 is configured as follows:

[0131] In the first training stage, the initial refined ranking model in the video recommendation model is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material;

[0132] In the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after initial training; wherein the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: material features and user features;

[0133] In the second training stage, the first sample set is used to iteratively train the refined ranking model after the initial training, and the parameters of the refined ranking model are adjusted based on the first loss value; and the second sample set is used to iteratively train the recall model after the initial training, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the finally trained refined ranking model and the recall model; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

[0134] The training module 61 is further configured to:

[0135] Inputting the material feature, the user feature, and the fusion feature of the user and the material in the first sample into the DNN to perform a feature cross operation to obtain a first cross feature;

[0136] Inputting the first cross feature into the activation layer for activation to obtain the first prediction value of the refined ranking model;

[0137] Determining a cross entropy based on the first predicted value and the first target value;

[0138] A first loss value of the refined ranking model is determined based on the cross entropy.

[0139] The training module 61 is further configured to:

[0140] Inputting the material feature of the second sample into the first tower model to perform feature crossover operation to obtain a second crossover feature;

[0141] Inputting the user feature of the second sample into a second tower model to perform a feature crossover operation to obtain a third crossover feature;

[0142] Performing an inner product operation of a vector on the second cross feature and the third cross feature to obtain the second prediction value;

[0143] Determine a cross entropy based on the second predicted value and the second target value;

[0144] The second loss value is determined based on the cross entropy.

[0145] The training module 61 is further configured to:

[0146] The positive samples and / or negative samples in the first sample set are selected in a different manner from the positive samples and / or negative samples in the second sample set.

[0147] The training module 61 is further configured to:

[0148] The first samples in the first sample set include positive samples and negative samples; the positive samples are samples in which the user's interaction time with the material or the material's playback time is higher than a threshold; the negative samples are samples in which the user's interaction time with the material or the material's playback time is lower than a threshold, and / or samples obtained based on probabilistic random negative sampling.

[0149] The training module 61 is further configured to:

[0150] The second samples in the second sample set include positive samples and negative samples; the positive samples are samples in which the user's interaction time with the material or the material's playback time is higher than a threshold, and / or samples in which the material's playback time is lower than a threshold; the negative samples are samples obtained based on probability random sampling.

[0151] The training module 61 is further configured to:

[0152] The method further comprises at least one of the following:

[0153] Reducing the weight of negative samples obtained based on probabilistic random negative sampling when determining the first loss value of the refined ranking model;

[0154] The weight of positive samples whose playing time of the material is lower than the threshold is reduced when determining the second loss value of the recall model.

[0155] The training module 61 is further configured to:

[0156] The refined ranking model and the recall model have shared sample features; the shared sample features are used for the refined ranking model and / or the recall model to correct sample features in corresponding sample sets.

[0157] The training module 61 is further configured to:

[0158] The shared sample features are determined based on the prediction results of the refined ranking model and / or the recall model; the number of shared sample features determined by the refined ranking model is greater than the number of shared sample features determined by the recall model.

[0159] The present disclosure provides a processing device, the processing device comprising:

[0160] A memory for storing executable programs;

[0161] The processor is used to implement any method described in the embodiments of the present disclosure when executing the executable program stored in the memory.

[0162] It can be understood that the memory can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM, SyncLink Dynamic Random Access Memory), and direct RAM bus random access memory (DRRAM, Direct Rambus Random Access Memory).The memories described in the embodiments of the present application are intended to include, but are not limited to, these and any other suitable types of memories.

[0163] Among them, the method disclosed in the present invention can be applied to the processor or implemented by the processor. The processor can be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the voice conversion method can be completed by the hardware integrated logic circuit or software instructions in the processor. The above-mentioned processor can be a general-purpose processor, a digital signal processor (DSP, Digital Signal Processor), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The processor can implement or execute the methods, steps and logic block diagrams disclosed in the present invention. The general-purpose processor can be a microprocessor or any conventional processor, etc. In combination with the steps of the method disclosed in the present invention, it can be directly embodied as a hardware decoding processor to execute, or it can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium, which is located in a memory, and the processor reads the information in the memory, and completes the steps of the voice conversion method provided in the embodiment of the present application in combination with its hardware.

[0164] The present invention also provides a computer storage medium, wherein the computer storage medium stores an executable program, and when the executable program is executed by a processor, the method described in any one of the embodiments of the present disclosure is implemented. Specifically, it can be a computer-readable storage medium, for example, a memory including a computer program, and the computer program can be executed by a processor of a processing device to complete the steps described in the method of the embodiment of the present application. The computer-readable storage medium can be a memory such as a ROM, a PROM, an EPROM, an EEPROM, a Flash Memory, a magnetic surface memory, an optical disk, or a CD-ROM.

[0165] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A video recommendation model training method, characterized in that: The method comprises: In the first training stage, the initial refined ranking model in the video recommendation model is iteratively trained using the first sample set, and the parameters of the refined ranking model are adjusted based on the first loss value until the number of training times meets the first convergence condition, thereby obtaining the refined ranking model after initial training; wherein the refined ranking model is used to predict the user's interest in the material for the first sample in the input first sample set, and output a first prediction value; the first loss value is determined based on the cross entropy between the first prediction value and a preset first target value; the first sample includes: material features, user features, and fusion features of the user and the material; In the first training stage, the initial recall model in the video recommendation model is iteratively trained using the second sample set, and the parameters of the recall model are adjusted based on the second loss value until the number of training times meets the second convergence condition, thereby obtaining the recall model after initial training; wherein the recall model is used to predict whether the user is interested in the material for the second sample in the input second sample set, and output a second prediction value; the second loss value is determined based on the cross entropy between the second prediction value and the preset second target value; the second sample includes: material features and user features; In the second training stage, the first sample set is used to iteratively train the refined ranking model after the initial training, and the parameters of the refined ranking model are adjusted based on the first loss value; and the second sample set is used to iteratively train the recall model after the initial training, and the parameters of the recall model are adjusted based on the third loss value and the second loss value; until the third convergence condition is met, the trained video recommendation model is obtained based on the finally trained refined ranking model and the recall model; wherein the third loss value is determined based on the cross entropy between the second prediction value predicted by the recall model and the first prediction value predicted by the refined ranking model.

2. The method according to claim 1, characterized in that The refined sorting model includes a fully connected neural network DNN layer and an activation layer; the method also includes: Inputting the material feature, the user feature, and the fusion feature of the user and the material in the first sample into the DNN to perform a feature cross operation to obtain a first cross feature; Inputting the first cross feature into the activation layer for activation to obtain the first prediction value of the refined ranking model; Determining a cross entropy based on the first predicted value and the first target value; A first loss value of the refined ranking model is determined based on the cross entropy.

3. The method according to claim 1, characterized in that The recall model includes a first tower model and a second tower model, and the method further includes: Inputting the material feature of the second sample into the first tower model to perform feature crossover operation to obtain a second crossover feature; Inputting the user feature of the second sample into a second tower model to perform a feature crossover operation to obtain a third crossover feature; Performing an inner product operation of a vector on the second cross feature and the third cross feature to obtain the second prediction value; Determine a cross entropy based on the second predicted value and the second target value; The second loss value is determined based on the cross entropy.

4. The method according to any one of claims 1 to 3, characterized in that The positive samples and / or negative samples in the first sample set are selected in a different manner from the positive samples and / or negative samples in the second sample set.

5. The method according to claim 4, characterized in that The first samples in the first sample set include positive samples and negative samples; the positive samples are samples in which the user's interaction time with the material or the material's playback time is higher than a threshold; the negative samples are samples in which the user's interaction time with the material or the material's playback time is lower than a threshold, and / or samples obtained based on probabilistic random negative sampling.

6. The method according to claim 4, characterized in that The second samples in the second sample set include positive samples and negative samples; the positive samples are samples in which the user's interaction time with the material or the material's playback time is higher than a threshold, and / or samples in which the material's playback time is lower than a threshold; the negative samples are samples obtained based on probability random sampling.

7. The method according to claim 5 or 6, characterized in that: The method further comprises at least one of the following: Reducing the weight of negative samples obtained based on probabilistic random negative sampling when determining the first loss value of the refined ranking model; The weight of positive samples whose playing time of the material is lower than the threshold is reduced when determining the second loss value of the recall model.

8. The method according to claim 1, characterized in that The refined ranking model and the recall model have shared sample features; the shared sample features are used for the refined ranking model and / or the recall model to correct sample features in corresponding sample sets.

9. The method according to claim 8, characterized in that The shared sample features are determined based on the prediction results of the refined ranking model and / or the recall model; the number of shared sample features determined by the refined ranking model is greater than the number of shared sample features determined by the recall model.

10. A processing device, characterized in that: The processing equipment comprises: A memory for storing executable programs; A processor, configured to implement the method according to any one of claims 1 to 9 when executing the executable program stored in the memory.

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

Images