Methods for obtaining recommendation information, model training methods, devices, electronic equipment, and storage media

By using multimedia and user feature cross-network modules for feature stitching in electronic devices, the problem of inaccurate recommendations in single-objective models is solved, the accuracy of video recommendations and the challenge of new material distribution are improved, and effective coordination of multi-objective learning is achieved.

CN117312630BActive Publication Date: 2026-11-10VIVO MOBILE COMM CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311283657.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-09-28
Publication Date
2026-11-10
Estimated Expiration
2043-09-28

AI Technical Summary

Technical Problem

When existing electronic devices recommend videos using a single-objective model, inaccurate recommendations are common, especially for long videos, which do not meet user needs and result in poor accuracy.

Method used

The recommendation model is trained by inputting multimedia feature description information and user feature description information. It performs feature cross-concatenation through N cross-network modules to obtain multiple feature vectors, and outputs recommendation information based on these feature vectors. The cross-network modules are used to strengthen feature cross-concatenation to improve the generalization of the model.

Benefits of technology

It improves the accuracy of video recommendations for electronic devices, resolves the problems of multi-objective learning conflicts and difficulties in cold start of new materials and distribution of long-tail materials, and enhances the overall accuracy of the recommendation system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117312630B_ABST
    Figure CN117312630B_ABST
Patent Text Reader

Abstract

This application discloses a method, apparatus, electronic device, and storage medium for obtaining recommendation information, belonging to the field of artificial intelligence technology. The method includes: inputting multimedia feature description information and user feature description information into a trained recommendation model, the trained recommendation model including N cross-network modules, each corresponding to a different feature dimension, to obtain at least two first feature vectors; inputting the at least two first feature vectors into a first cross-network module, and cross-concatenating the at least two first feature vectors through the first feature dimension corresponding to the first cross-network module to obtain at least one second feature vector, the first cross-network module being one of the N cross-network modules; and outputting first recommendation information based on the at least one second feature vector, where N is a positive integer.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a method for obtaining recommendation information, a model training method, a device, an electronic device, and a storage medium. Background Technology

[0002] Currently, electronic devices can recommend videos of interest to users based on their historical behavior during use. In related technologies, electronic devices can recommend videos of interest to users using a single-objective model.

[0003] However, electronic devices, using single-objective models, can only satisfy single-objective needs. For example, if the objective is duration, it easily leads to long videos, meaning the videos recommended by the electronic device may not necessarily meet the user's needs. Thus, the accuracy of video recommendations by electronic devices is poor. Summary of the Invention

[0004] The purpose of this application is to provide a method for obtaining recommendation information, a model training method, an apparatus, an electronic device, and a storage medium, which can improve the accuracy of video recommendations by electronic devices.

[0005] In a first aspect, embodiments of this application provide a method for obtaining recommendation information. The method includes: inputting multimedia feature description information and user feature description information into a trained recommendation model, the trained recommendation model including N cross-network modules, each cross-network module corresponding to a different feature dimension, to obtain at least two first feature vectors; inputting the at least two first feature vectors into a first cross-network module, and cross-concatenating the at least two first feature vectors through the first feature dimension corresponding to the first cross-network module to obtain at least one second feature vector, the first cross-network module being one of the N cross-network modules; and outputting first recommendation information based on the at least one second feature vector, where N is a positive integer.

[0006] Secondly, embodiments of this application provide a device for acquiring recommendation information, comprising an input module, an extraction module, and an output module. The input module is used to input multimedia feature description information and user feature description information into a trained recommendation model. The trained recommendation model includes N cross-network modules, each corresponding to a different feature dimension, resulting in at least two first feature vectors. The extraction module inputs the at least two first feature vectors acquired by the input module into a first cross-network module, and cross-concatenates the at least two first feature vectors using the first feature dimension corresponding to the first cross-network module to obtain at least one second feature vector. The first cross-network module is one of the N cross-network modules. The output module is used to output first recommendation information based on at least one second feature vector, where N is a positive integer.

[0007] Thirdly, embodiments of this application provide a model training method, which includes: calculating a first value corresponding to a first cross-network module based on multimedia feature description information and user feature description information, wherein the first value is used to characterize the recommendation probability of each preset video under a first feature dimension, and the first feature dimension is the feature dimension corresponding to the first cross-network module; calculating a second value based on a fourth feature vector, wherein the second value is used to characterize the recommendation probability of each preset video; determining a first loss value corresponding to the first cross-network module based on the first value and the second value; weighting the first loss value using the feature dimension weights of the first cross-network module to obtain a second loss value corresponding to the first cross-network module; and training a recommendation model based on the second loss value corresponding to each cross-network module to obtain a trained recommendation model; wherein the sum of the feature dimension weights of each cross-network module is 1.

[0008] Fourthly, embodiments of this application provide a model training apparatus, which includes: a calculation module, a determination module, a weighting module, and a training module. The calculation module is used to calculate a first value corresponding to a first cross-network module based on multimedia feature description information and user feature description information. The first value represents the recommendation probability of each preset video under a first feature dimension, where the first feature dimension is the feature dimension corresponding to the first cross-network module. It also calculates a second value based on a fourth feature vector, where the second value represents the recommendation probability of each preset video. The determination module is used to determine a first loss value corresponding to the first cross-network module based on the first and second values ​​obtained by the calculation module. The weighting module is used to weight the first loss value using the feature dimension weights of the first cross-network module to obtain a second loss value corresponding to the first cross-network module. The training module is used to train a recommendation model based on the second loss value corresponding to each cross-network module to obtain a trained recommendation model; wherein the sum of the feature dimension weights of each cross-network module is 1.

[0009] Fifthly, embodiments of this application provide an electronic device including a processor and a memory, the memory storing a program or instructions executable on the processor, the program or instructions, when executed by the processor, implementing the steps of the method described in the first aspect or the method described in the third aspect.

[0010] In a sixth aspect, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect or the steps of the method described in the third aspect.

[0011] In a seventh aspect, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the steps of the method described in the first aspect or the method described in the third aspect.

[0012] Eighthly, embodiments of this application provide a computer program product stored in a storage medium, which is executed by at least one processor to implement the steps of the method described in the first aspect or the method described in the third aspect.

[0013] In this embodiment, the electronic device can input multimedia feature description information and user feature description information into a trained recommendation model. The trained recommendation model includes N cross-network modules, each corresponding to a different feature dimension, resulting in at least two first feature vectors. The at least two first feature vectors are input into a first cross-network module, and through the first feature dimension corresponding to the first cross-network module, the at least two first feature vectors are cross-concatenated to obtain at least one second feature vector. The first cross-network module is one of the N cross-network modules. Based on the at least one second feature vector, first recommendation information is output, where N is a positive integer. In this scheme, since the electronic device can cross-concatenate at least two first feature vectors through the first cross-network module to obtain at least one second feature vector, the electronic device can strengthen feature cross-concatenation through the first cross-network module, improve the model's generalization, and thus improve the accuracy of video recommendation by the electronic device using the at least two cross-concatenated first feature vectors. Attached Figure Description

[0014] Figure 1 This is one of the flowcharts of a method for obtaining recommendation information provided in the embodiments of this application;

[0015] Figure 2 This is a second flowchart of a method for obtaining recommendation information provided in an embodiment of this application;

[0016] Figure 3 This is one of the flowcharts of a model training method provided in the embodiments of this application;

[0017] Figure 4 This is a second flowchart of a model training method provided in an embodiment of this application;

[0018] Figure 5 This is one of the schematic diagrams illustrating a recommendation model provided in the embodiments of this application;

[0019] Figure 6 This is a schematic diagram of the structure of a device for obtaining recommendation information provided in an embodiment of this application;

[0020] Figure 7 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application;

[0021] Figure 8 This is one of the hardware structure diagrams of an electronic device provided in the embodiments of this application;

[0022] Figure 9 This is a second schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0024] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0025] The following explains the technical terms used in the recommended methods, apparatus, electronic devices, and storage media provided in the embodiments of this application.

[0026] Multi-task learning model (Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts, MMoe): A neural network model capable of multi-objective learning.

[0027] Expert: Multi-objective sharing expert network in MMoe models

[0028] Material cold start: refers to newly generated materials that are difficult or inaccurate to distribute by the recommendation system due to a lack of or very little historical behavior.

[0029] L2 normalization: A commonly used data preprocessing method, also known as Euclidean normalization. It scales the feature vectors by dividing each sample's feature vector by its L2 norm, thus giving the feature vectors a unit norm.

[0030] Nonnegativity constraint: A nonnegativity constraint is a restriction that requires a set of parameters or variables to have values ​​greater than or equal to zero. In machine learning and optimization problems, nonnegativity constraints are often used to ensure the nonnegativity of model parameters or variables, such as in linear programming problems and nonnegative matrix factorization.

[0031] Multilayer Perceptron (MLP): Also known as an artificial neural network, it can contain multiple hidden layers in addition to the input and output layers. The simplest MLP contains only one hidden layer, i.e., a 3-layer structure. Moreover, the layers of an MLP are fully connected. The bottom layer of an MLP is the input layer, the middle layers are the hidden layers, and the last layer is the output layer.

[0032] The following description, in conjunction with the accompanying drawings, details the method for obtaining recommendation information, the model training method, the apparatus, the electronic device, and the storage medium provided in the embodiments of this application through specific implementations and application scenarios.

[0033] With the rapid development of internet technology, the cost of obtaining information is decreasing, and how to acquire content that is both interesting and valuable to oneself has become an increasingly important issue. Typically, electronic devices can use recommendation systems to suggest personalized content to each user, thereby significantly reducing the cost of information acquisition.

[0034] In related technologies, when electronic devices model user preferences through recommendation systems, single-objective modeling, such as focusing on clicks, tends to produce clickbait titles, focus on duration, and generate long videos; conversely, focusing on completion time tends to generate short videos. Therefore, clicks, duration, and completion time are generally considered simultaneously to model user interests from different dimensions. Early solutions used different models for different objectives because these three objectives are strongly correlated. To save resources and reduce online inference time, subsequent developments have moved to modeling all three objectives together, i.e., multi-objective models. In multi-objective modeling, different objectives share underlying parameters but also have their own corresponding upper-level parameters. For example, in the MMoe model structure, multiple objectives share multiple expert network modules, and each objective has different weight parameters for the experts. This improvement in network structure greatly saves resources and ensures diverse user experiences. However, it also suffers from low recommendation accuracy and difficulties in distributing cold-start and long-tail materials. Furthermore, multiple objectives experience learning conflicts in the shared parameters, meaning that optimizing one objective leads to a decline in the performance of other objectives.

[0035] Existing multi-objective modeling schemes for recommendation systems generally suffer from the following problems:

[0036] 1. General multi-objective models, such as the MMoe model, are difficult to achieve good results in specific scenarios, and the accuracy of recommendations needs to be improved.

[0037] 2. There are problems with the cold start of new materials and the difficulty in distributing long-tail materials. These materials are easily ignored by the system due to their low activity.

[0038] 3. There is a problem of learning conflicts between multiple objectives. Some works have mitigated these conflicts by optimizing shared and unique modules among objectives from the perspective of model structure. However, these approaches cannot completely resolve learning conflicts between multiple objectives.

[0039] 4. Common Pareto optimization strategies for resolving conflicts in multi-objective learning mostly employ heuristic algorithms. The searched solutions may have low accuracy or require a large number of iterations, making execution time-consuming and unfavorable for the rapid iteration of recommendation models.

[0040] In the recommendation method, apparatus, electronic device and storage medium provided in this application, since the electronic device can cross-concatenate at least two first feature vectors through the first cross-network module to obtain at least one second feature vector, the electronic device can strengthen feature cross-concatenation through the first cross-network module, improve the generalization of the model, and thus improve the accuracy of the electronic device in video recommendation by outputting first recommendation information through at least two cross-concatenated first feature vectors.

[0041] The execution entity of the recommendation information acquisition method provided in this application embodiment can be a recommendation information acquisition device, which can be an electronic device or a functional module in an electronic device. The following uses an electronic device as an example to illustrate the technical solution provided in this application embodiment.

[0042] This application provides a method for obtaining recommendation information. Figure 1 A flowchart illustrating a method for obtaining recommendation information provided in an embodiment of this application is shown. Figure 1 As shown, the method for obtaining recommendation information provided in this application embodiment may include the following steps 201 to 203.

[0043] Step 201: The electronic device inputs the multimedia feature description information and the user feature description information into the trained recommendation model to obtain at least two first feature vectors.

[0044] In this embodiment of the application, the trained recommendation model includes N cross-network modules, each of which corresponds to a different feature dimension, where N is a positive integer.

[0045] Optionally, in this embodiment of the application, the multimedia feature description information mentioned above can be the feature description information of a video.

[0046] For example, the feature description information of the video may include at least one of the following: video identity document (ID), video category, such as: movie, TV series, animation, video historical click rate, video content information, such as text information and image information, and video playback timestamp information.

[0047] For example, the user characteristic description information mentioned above may include at least one of the following: user gender, user age, video sequence clicked by the user within a historical time period, and timestamp information of the videos clicked by the user.

[0048] Optionally, in this embodiment of the application, the electronic device can obtain user characteristic description information through the user's historical behavior.

[0049] For example, electronic devices can save the actions a user performs while browsing videos, such as the number of times a video is clicked, the number of videos clicked, and the types of videos clicked.

[0050] Furthermore, electronic devices can obtain video content information of the video corresponding to the user's historical behavior, such as the timestamp information of the video played by the electronic device, the text information of the video, and the image information of the video.

[0051] Optionally, in the embodiments of this application, the multimedia feature description information and user feature description information mentioned above can be discrete features or floating-point features.

[0052] Optionally, in this embodiment of the application, the trained recommendation model may include an input feature module. The electronic device may input the multimedia feature description information and user feature description information into the recommendation model based on the input feature module to obtain at least two first feature vectors.

[0053] In this embodiment of the application, the electronic device can extract at least two first feature vectors based on the first multimedia feature description information and the first user feature description information.

[0054] In this embodiment of the application, the trained recommendation model may further include a vector embedding layer. The electronic device can obtain the feature vectors corresponding to the multimedia feature description information and the user feature description information through the vector embedding layer, so as to extract at least two first feature vectors from the feature vectors.

[0055] For example, the trained recommendation model can concatenate multimedia feature description information and user feature description information in the vector embedding layer and perform vector embedding, so that the recommendation model can generate feature vectors corresponding to the multimedia feature description information and user feature description information.

[0056] In this embodiment of the application, after obtaining the feature vectors corresponding to the multimedia feature description information and the user feature description information, the electronic device can input the feature vectors into at least two expert network modules to obtain at least two first feature vectors.

[0057] In this embodiment of the application, the feature dimensions of each of the at least two first feature vectors are different.

[0058] Step 202: The electronic device inputs the above-mentioned at least two first feature vectors into the first cross-network module, and cross-concatenates the at least two first feature vectors through the first feature dimension corresponding to the first cross-network module to obtain at least one second feature vector.

[0059] In this embodiment of the application, the first cross-network module is one of N cross-network modules.

[0060] It should be noted that feature cross-structuring is one of the most effective methods for optimizing recommendation models, demonstrating good performance across various scenarios and tasks. Typical multi-objective network models, such as the MMoe network model, aggregate multiple expert embeddings with different weights for each objective, but lack cross-information among these expert embeddings. Therefore, this application adds a cross-network module to the MMoe network model to perform feature cross-structuring on at least two of the aforementioned first feature vectors, thereby improving the generalization and accuracy of the MMoe network model.

[0061] It should be noted that the above targets are predicted videos that meet user needs.

[0062] For example: taking the click-through rate (CTR) target as an example, and the CTR target corresponds to 3 gate networks; the gate network corresponding to the CTR target generates the weights of 3 expert modules corresponding to the CTR target, wherein the weights of the 3 expert modules are as follows: the weight of expert module 1 under the CTR target is ctr_expert1_weight, the weight of expert module 2 under the CTR target is ctr_expert2_weight, and the weight of expert module 3 under the CTR target is ctr_expert3_weight. The input to the expert network cross module can be set as: [the weight of expert module 1 under the click-through rate target, ctr_expert1_weight * the first feature vector of expert module 1, expert1 embedding; the weight of expert module 2 under the click-through rate target, ctr_expert2_weight * the first feature vector of expert module 2, expert2 embedding; ctr_expert1_weight * expert1 embedding - ctr_expert2_weight * expert2 embedding; ctr_expert1_weight * expert1 embedding ⊙ ctr_expert2_weight * expert2 embedding]. The expert network cross module can be implemented as a multilayer perceptron, and the dimension of the output embedding can be determined according to the project. Thus, the cross output of expert1 and expert2 required for the CTR task, ctr_expert_1_2, is obtained. Similarly, the cross output of expert2 and expert3, ctr_expert_2_3, and the cross output of expert1 and expert3, ctr_expert_1_3, can be obtained. The second feature vector is obtained by concatenating ctr_expert1_weight*expert1embedding + ctr_expert2_weight*expert2 embedding + ctr_expert3_weight*expert3 embedding.

[0063] Optionally, in this embodiment of the application, the trained recommendation model includes M expert network modules, each of the M expert network modules corresponds to at least one feature dimension, and the feature dimensions corresponding to the M expert network modules include the feature dimensions corresponding to the N cross-network modules, where M is an integer greater than 1.

[0064] For example, combined Figure 1 ,like Figure 2 As shown, step 202 above can be specifically implemented through step 202a below.

[0065] Step 202a: The electronic device inputs the above multimedia feature description information and user feature description information into each of the M expert network modules to obtain M first feature vectors.

[0066] In this embodiment of the application, one of the M expert modules, the expert network module, outputs a first feature vector.

[0067] For example, each expert network module in the embodiments of this application is a multilayer perceptron module.

[0068] It should be noted that the feature dimensions mentioned above refer to the focus of each expert network module. For example, taking two expert modules as an example, expert module 1 focuses on the video's click-through rate, while expert module 2 focuses on the video's completion rate.

[0069] In this embodiment, after the electronic device inputs the feature vectors corresponding to the multimedia feature description information and the user feature description information into each of the M expert network modules, each expert network can perform feature extraction processing on the feature vectors corresponding to the first multimedia feature description information and the first user feature description information according to its own corresponding feature dimension, thereby obtaining M first feature vectors.

[0070] Optionally, in this embodiment of the application, the above-trained recommendation model further includes a gate network module. After the electronic device extracts at least two first feature vectors, the electronic device can obtain the weights corresponding to each expert network module through the gate network module. Then, the electronic device can concatenate the weights corresponding to each expert network module with the first feature vectors output by each expert network module to obtain a first feature vector containing the weights.

[0071] For example, the above-mentioned gate network module can be one or more.

[0072] For example, the above-mentioned gate network module can correspond one-to-one with M expert network modules.

[0073] Step 203: The electronic device outputs first recommendation information based on at least one of the above-mentioned second feature vectors.

[0074] In this embodiment of the application, after obtaining at least one second feature vector, the electronic device can input the at least one second feature vector into the target tower corresponding to the target to obtain the prediction result corresponding to the target. Then, the electronic device can output the first recommendation information based on the prediction result.

[0075] Optionally, in the embodiments of this application, the above prediction result can be a score or a probability value.

[0076] For example, after obtaining the above-mentioned at least one second feature vector, the electronic device can input the at least one second feature vector into the target tower corresponding to the target to obtain the Logit value corresponding to the target. The Logit value is used to calculate the prediction result corresponding to the target, and the Logit value is input into the target prediction module to obtain the score value.

[0077] It's understandable that each video corresponds to a score.

[0078] It should be noted that if there are multiple targets in the recommendation model, the scores corresponding to the multiple targets need to be added together, and the scores corresponding to each video need to be sorted from largest to smallest. The top K videos are then selected as the first recommendation information to be recommended to the user.

[0079] In the method for obtaining recommendation information provided in this application embodiment, the electronic device can input multimedia feature description information and user feature description information into a trained recommendation model. The trained recommendation model includes N cross-network modules, each corresponding to a different feature dimension, to obtain at least two first feature vectors. The at least two first feature vectors are input into a first cross-network module, and through the first feature dimension corresponding to the first cross-network module, the at least two first feature vectors are cross-concatenated to obtain at least one second feature vector. The first cross-network module is one of the N cross-network modules. Based on the at least one second feature vector, first recommendation information is output, where N is a positive integer. In this scheme, since the electronic device can cross-concatenate at least two first feature vectors through the first cross-network module to obtain at least one second feature vector, the electronic device can strengthen feature cross-concatenation through the first cross-network module, improve the model's generalization, and thus improve the accuracy of video recommendation by the electronic device through the at least two cross-concatenated first feature vectors.

[0080] Optionally, in the embodiments of this application, after step 201 above, the recommended method provided in the embodiments of this application further includes steps 301 to 303 as described below.

[0081] Step 301: The electronic device performs feature concatenation on the above multimedia feature description information and user feature description information to obtain a third feature vector.

[0082] It should be noted that the introduction of step 301 is mainly to solve the problem of difficulty in distributing new materials and medium-to-long-tail materials in the recommendation system. Due to the lack of sufficient historical behavior, the relevant sparse features of such materials are often submerged in the recommendation model, resulting in low prediction accuracy.

[0083] In this application, the memory of sparse features can be enhanced by performing shallow cross-referencing of multimedia feature description information and user feature description information, and then directly inputting the cross-referencing result into the output layer of the target tower.

[0084] Optionally, in this embodiment of the application, after processing the third feature vector, the electronic device can perform one-dimensional scalar embedding processing on the third feature vector to obtain the final crossover result.

[0085] Optionally, in this embodiment of the application, the electronic device can fuse the cross results according to different objectives and the weights corresponding to those objectives, thereby obtaining cross results that are more closely aligned with the objectives.

[0086] For example, suppose the user-side first-level category profile features are {1:0.2, 7:0.4, 10:0.4}, where 1, 7, and 10 are the first-level category codes of the short videos the user has watched, and 0.2, 0.4, and 0.4 are the degree of the user's preference for these categories. If the material's ID feature is id1, then the cross feature between the material ID and the user-side first-level category profile feature is {1_id1:0.2,7_id1:0.4,10_id1:0.4}. Encoding 1_id1, 7_id1, and 10_id1 and performing one-dimensional scalar embedding processing yields the final cross result as {x_1_id1*0.2,x_7_id1*0.4,x_10_id1*0.4}. For different objectives, the cross values ​​are fused with different weights. For example, for the click-through rate (CTR) task, the fused cross result is: logit_ctr_cross=w1_ctr*x_1_id1*0.2+w2_ctr*x_7_id1*0.4+w3_ctr*x_10_id1*0.4. Furthermore, by performing one-dimensional scalar embedding on the material information to obtain logit_ctr_id and logit_ctr_up, the performance of the recommendation model is further improved.

[0087] Step 302: The electronic device performs feature concatenation on at least one second feature vector and the third feature vector to obtain a fourth feature vector.

[0088] Optionally, in this embodiment of the application, the electronic device can obtain at least one logit value corresponding to a second feature vector and a logit value corresponding to a third feature vector, and then add the logit value corresponding to at least one second feature vector and the logit value corresponding to the third feature vector to obtain a total logit value, which is used to calculate the prediction result corresponding to the target.

[0089] For example, the recommendation model can add the final cross-entropy result to the logit output value of the target tower to obtain the total logit value for each task. For instance, the total logit value for click-through rate (CTR) is logit_ctr_f = logit_ctr + logit_ctr_cross + logit_ctr_id + logit_ctr_up, and the loss for the CTR target is ctr_loss = cross_entropy. Similarly, the dur_loss for the duration target can be obtained as mse(logit_dur_f), and the cv_loss for the completion target can be obtained as mse(logit_cv_f).

[0090] Step 303: The electronic device outputs recommendation information based on the fourth feature vector mentioned above.

[0091] In this embodiment of the application, after obtaining the fourth feature vector, the electronic device can input the fourth feature vector into the target tower corresponding to the target to obtain the prediction result corresponding to the target. Then, the electronic device can output recommendation information based on the prediction result.

[0092] For example, after obtaining the fourth feature vector, the electronic device can input the fourth feature vector into the target tower corresponding to the target to obtain the Logit value corresponding to the target. The Logit value is used to calculate the prediction result corresponding to the target, and the Logit value is input into the target prediction module to obtain the score value. Then, the electronic device can output recommendation information based on the prediction result.

[0093] In this embodiment, the electronic device can perform feature concatenation using the multimedia feature description information and user feature description information to obtain a third feature vector, and then concatenate the at least one second feature vector with the third feature vector to obtain a fourth feature vector, thereby solving the problem of difficult distribution of new materials and mid-to-long-tail materials in the recommendation system and improving the accuracy of prediction by the electronic device.

[0094] This application provides a model training method. Figure 3 A flowchart of a model training method provided in an embodiment of this application is shown. Figure 3 As shown, the model training method provided in this application embodiment may include the following steps 401 to 405.

[0095] Step 401: The electronic device calculates the first value corresponding to the first cross-network module based on the multimedia feature description information and the user feature description information.

[0096] In this embodiment of the application, the first numerical value is used to characterize the recommendation probability of each preset recommended video under the first feature dimension, and the first feature dimension is the feature dimension corresponding to the first cross-network module.

[0097] In this embodiment of the application, the electronic device can input multimedia feature description information and user feature description information into the expert network to obtain the above-mentioned at least two first feature vectors, and input the at least two first feature vectors into the first cross network module to obtain the first value corresponding to the first cross network module.

[0098] Optionally, in the embodiments of this application, the number of the first value mentioned above can be one or more.

[0099] Optionally, in this embodiment of the application, the number of the first cross-network modules can be one or more.

[0100] For example, the number of the first numerical values ​​mentioned above corresponds to the number of the first cross-network modules mentioned above. That is, assuming there are 3 first cross-network modules, then there are also 3 first numerical values.

[0101] Step 402: The electronic device calculates the second value based on the fourth feature vector mentioned above.

[0102] In this embodiment of the application, the second value is used to characterize the recommendation probability of each preset recommended video.

[0103] In this embodiment of the application, after obtaining the fourth feature vector, the electronic device can obtain the aforementioned second value through the target tower in the training model.

[0104] Optionally, in the embodiments of this application, the number of the second value mentioned above can be one or more.

[0105] For example, the quantity of the second value mentioned above corresponds to the quantity of the target towers mentioned above. That is, assuming there are 3 target towers, the quantity of the second value is also 3.

[0106] Step 403: Based on the first and second values ​​mentioned above, the electronic device determines the first loss value corresponding to the first cross-network module.

[0107] In this embodiment of the application, the electronic device can determine the sum of the first value and the second value as the first loss value corresponding to the first cross-network module.

[0108] It should be noted that when there are multiple first cross-network modules, each first cross-network module corresponds to a first value and a second value.

[0109] Step 404: The electronic device uses the feature dimension weights of the first cross-network module to weight the first loss value, thereby obtaining the second loss value corresponding to the first cross-network module.

[0110] In this embodiment of the application, the sum of the feature dimension weights of each of the above N cross-network modules is 1.

[0111] It should be noted that the specific implementation process can be found in the above embodiments. To avoid repetition, it will not be repeated here.

[0112] Step 405: The electronic device trains a recommendation model based on the second loss value corresponding to each of the above-mentioned cross-network modules to obtain the trained recommendation model.

[0113] In this embodiment of the application, the sum of the feature dimension weights of each of the above N cross-network modules is 1.

[0114] In this embodiment of the application, the electronic device can perform backpropagation training on the recommendation model based on the second loss value corresponding to each of the above-mentioned cross-network modules to obtain the trained recommendation model.

[0115] In this embodiment of the application, after the model training is completed, the electronic device can use online inference to determine the user's click rate corresponding to the short video score pCtr, the duration corresponding to the score pDur, and the completion score pCv. The results of the fusion of the three scores according to certain rules are sorted from largest to smallest, and the top K is selected and pushed to the user.

[0116] Optionally, in the embodiments of this application, combined with Figure 3 ,like Figure 4 As shown, prior to step 404 above, the recommended method provided in this application embodiment further includes steps 501 and 502 as described below.

[0117] Step 501: The electronic device normalizes the first gradient vector corresponding to each of the above-mentioned cross-network modules to obtain the second gradient vector corresponding to each of the cross-network modules.

[0118] In this embodiment, the first gradient vector corresponding to each of the above N cross-network modules is determined based on the second loss value and the first feature vector corresponding to the cross-network module.

[0119] For example, for each cross-network module, the electronic device can calculate the gradient of each target loss with respect to the shared variables based on backpropagation. Abbreviated as

[0120] Step 502: According to Pareto theory, the electronic device obtains the feature dimension weight vector based on the second gradient vector corresponding to each cross-network module, and determines the feature dimension weight corresponding to each cross-network module based on the feature dimension weight vector.

[0121] For example, the Pareto theory described above can be used as...

[0122] stα ctr ≥0,α dur ≥0,

[0123] α cr ≥0,α ctr +α dur +α cr =1

[0124] Then the electronic device can obtain a set of weighting coefficients α = [α ctr α dur α cr ].

[0125] In this embodiment of the application, the electronic device improves the Pareto-based multi-objective parameter solving method, which facilitates tensor quantization, distributed computing, and solves the learning conflict problem of multiple objectives.

[0126] Optionally, in the embodiments of this application, step 502 can be implemented by steps 501a and 501b as described below.

[0127] Step 501a: When the feature dimension weight vector is non-negative, the electronic device calculates the feature dimension weight corresponding to each of the cross-network modules based on the feature dimension weight vector.

[0128] In this embodiment of the application, after obtaining the first gradient corresponding to each target, the electronic device can perform gradient L2 normalization on each first gradient and construct a matrix. And we obtain M = CC T Then, electronic devices can... A weight vector is obtained, which contains the feature dimension weights corresponding to each of the cross-network modules.

[0129] Step 501b: When the above feature dimension weight vector is negative, the electronic device calculates the feature dimension weight corresponding to each cross network module according to the Pareto rule based on the feature dimension weight vector.

[0130] In this embodiment of the application, the Pareto rule includes: if a Pareto feasible solution falls on the Pareto boundary corresponding to the second feature dimension, then the feature dimension weight of the cross network module corresponding to the second feature dimension is 0, and the second feature dimension is the feature dimension of at least one cross network module among N cross network modules.

[0131] For example, the Pareto target weights are located on the non-negative constraint boundary of the target weights. The solutions can be discussed separately for different cases, with a total of 6 possible cases.

[0132] Case 1: The feasible solution lies on the boundary of the CTR target, i.e., α ctr =0;

[0133] Scenario 2: The feasible solution lies on the boundary of the duration objective, i.e., α. dur =0;

[0134] Scenario 3: The feasible solution lies on the boundary of the complete broadcast objective, i.e., α. cr =0;

[0135] Scenario 4: The feasible solution lies on the boundary between the CTR objective and the duration objective, i.e., α. ctr =α dur =0;

[0136] Case 5: The feasible solution lies on the boundary between the CTR objective and the completion objective, i.e., α. ctr =α cr =0;

[0137] Scenario 6: The feasible solution lies on the boundary between the duration objective and the completion objective, i.e., α. dur =α cr =0.

[0138] For each case, the optimal solution can be directly calculated. The solution with the smallest loss among the six possible solutions is the optimal solution under the original objective. Cases one, two, and three differ only in their boundary objectives and have similar solutions. Cases four, five, and six also have similar solutions.

[0139] The feasible solution lies on the boundary of the CTR objective, degenerating into a bi-objective optimization problem. The solution at this point is: From the normalization of the gradients in the above steps, it is easy to know that α ctr ≥0andα dur If ≤1, then α ctr =1-α dur .

[0140] From the above method embodiments, it is easy to see that in situation four, α cr =1, and the corresponding loss is 1, saving the calculation of cases four, five and six.

[0141] In this embodiment, the electronic device improves the Pareto-based multi-objective parameter solving method, which facilitates tensor quantization and distributed computing, and resolves the learning conflict problem of multiple objectives.

[0142] For example, such as Figure 5 As shown, the recommendation model 10 may include: a feature input module, a vector embedding module, Gate network 1, Gate network 2, Gate network 3, expert network module 1, expert network module 2, expert network module 3, expert cross-network module 1, expert cross-network module 2, expert cross-network module 3, and target tower 1. Figure 1 The middle is represented by the click-through rate (Ctr) tower, and the target tower is 2. Figure 2 The time-based tower represents the target tower (3). Figure 5 The system consists of a completion tower, a CTR prediction module, a duration prediction module, a completion prediction module, and a shallow cross-network. The feature input module is connected to a vector embedding module, which in turn is connected to Gate network 1, Gate network 2, Gate network 3, and three expert network modules. These three expert network modules are connected to three expert cross-network modules, which in turn are connected to three target towers. The CTR tower is connected to CTR prediction module 14, the duration tower is connected to duration prediction module 15, and the completion tower is connected to completion prediction module 16. The feature input module is connected to a shallow cross-network, which in turn is connected to the CTR prediction module, duration prediction module, and completion prediction module.

[0143] The model training method provided in this application allows electronic devices to improve the prediction accuracy of multi-objective models and solve the problems of material cold start and mid-to-long-tail material distribution by designing expert network cross sub-networks, shallow cross sub-structures of user-side features and material IDs, and up. Furthermore, it improves the Pareto-based multi-objective parameter solving method to resolve multi-objective learning conflicts.

[0144] It should be noted that the method for obtaining recommendation information provided in this application embodiment can be executed by a recommendation information acquisition device, an electronic device, or a functional module or entity within an electronic device. This application embodiment uses a recommendation information acquisition device executing the recommendation method as an example to illustrate the recommendation information acquisition device provided in this application embodiment.

[0145] Figure 6 A schematic diagram of a possible structure of the device for obtaining recommendation information involved in an embodiment of this application is shown. For example... Figure 6 As shown, the device 70 for obtaining recommendation information may include an input module 71, an extraction module 72, and an output module 73.

[0146] The input module 71 is used to input multimedia feature description information and user feature description information into the trained recommendation model. The trained recommendation model includes N cross-network modules, each corresponding to a different feature dimension, to obtain at least two first feature vectors. The extraction module 72 is used to input the at least two first feature vectors obtained by the input module into a first cross-network module, and cross-concatenate the at least two first feature vectors through the first feature dimension corresponding to the first cross-network module to obtain at least one second feature vector. The first cross-network module is one of the N cross-network modules. The output module 73 is used to output first recommendation information based on the at least one second feature vector extracted by the extraction module, where N is a positive integer.

[0147] In one possible implementation, the device for obtaining recommendation information provided in this application embodiment further includes a splicing module. The splicing module is used to, after the input module inputs multimedia feature description information and user feature description information into the recommendation model, splice the multimedia feature description information and user feature description information to obtain a third feature vector, and then splice at least one second feature vector and the third feature vector to obtain a fourth feature vector. The output module 73 is further used to output recommendation information based on the fourth feature vector.

[0148] In one possible implementation, the trained recommendation model includes M expert network modules, each expert network module corresponding to at least one feature dimension, and the feature dimensions corresponding to the M expert network modules include the feature dimensions corresponding to the N cross-network modules, where M is an integer greater than 1; the input module 71 is specifically used to input multimedia feature description information and user feature description information into each expert network module respectively to obtain M first feature vectors, and each expert network module outputs one first feature vector.

[0149] This application provides a device for obtaining recommendation information. Since the device can cross-concatenate at least two first feature vectors through a first cross-network module to obtain at least one second feature vector, the device can strengthen feature cross-concatenation through the first cross-network module, improve the generalization of the model, and thus improve the accuracy of the device in recommending videos by outputting first recommendation information through at least two cross-concatenated first feature vectors.

[0150] Figure 7 A schematic diagram of a possible structure of the model training apparatus involved in an embodiment of this application is shown. For example... Figure 7 As shown, the model training device 80 may include: a calculation module 81, a determination module 82, a weighting module 83, and a training module 84.

[0151] The system includes the following modules: Calculation module 81 calculates a first value corresponding to the first cross-network module based on multimedia feature description information and user feature description information. This first value represents the recommendation probability of each video to be recommended under the first feature dimension. It also calculates a second value based on the fourth feature vector, which represents the recommendation probability of each video to be recommended. Determination module 82 determines a first loss value corresponding to the first cross-network module based on the first and second values. Weighting module 83 weights the first loss value using the feature dimension weights of the first cross-network module to obtain a second loss value corresponding to the first cross-network module. Training module 84 trains the recommendation model based on the second loss value corresponding to each cross-network module, obtaining the trained recommendation model. The sum of the feature dimension weights of each cross-network module is 1.

[0152] In one possible implementation, the model training apparatus provided in this application embodiment further includes a processing module. The processing module is configured to, before weighting the first loss value using the feature dimension weights of the first cross-network module to obtain the second loss value corresponding to the first cross-network module, normalize the first gradient vector corresponding to each cross-network module to obtain the second gradient vector corresponding to each cross-network module. The first gradient vector corresponding to each cross-network module is determined based on the second loss value and the first feature vector corresponding to the cross-network module. Furthermore, according to Pareto theory, based on the second gradient vector corresponding to each cross-network module, a feature dimension weight vector is obtained, and based on the feature dimension weight vector, the feature dimension weights corresponding to each cross-network module are determined.

[0153] In one possible implementation, the aforementioned calculation module 81 is specifically used to calculate the feature dimension weights corresponding to each cross-network module based on the feature dimension weight vector when the feature dimension weight vector is non-negative; and to calculate the feature dimension weights corresponding to each cross-network module based on the feature dimension weight vector according to the Pareto rule when the feature dimension weight vector is negative; wherein the Pareto rule includes: if the Pareto feasible solution falls on the Pareto boundary corresponding to the second feature dimension, then the feature dimension weight of the cross-network module corresponding to the second feature dimension is 0, and the second feature dimension is the feature dimension of at least one of the N cross-network modules.

[0154] This application provides a model training device that can improve the prediction accuracy of multi-objective models and solve the problems of cold start and distribution of mid-to-long-tail materials by designing expert network cross sub-networks, shallow cross sub-structures of user-side features and material IDs, and up values. Furthermore, it improves the Pareto-based multi-objective parameter solving method to resolve multi-objective learning conflicts.

[0155] The device for obtaining recommendation information in this application embodiment can be an electronic device or a component within an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal or other devices besides a terminal. For example, a mobile electronic device can be a mobile phone, tablet computer, laptop computer, PDA, in-vehicle electronic device, mobile internet device (MID), augmented reality (AR) / virtual reality (VR) device, robot, wearable device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc. It can also be a server, network attached storage (NAS), personal computer (PC), television set (TV), ATM, or self-service machine, etc. This application embodiment does not specifically limit the device.

[0156] The device for obtaining recommendation information in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit the specific operating system used.

[0157] The device for obtaining recommendation information provided in this application embodiment can implement the various processes implemented in the above method embodiments. To avoid repetition, it will not be described again here.

[0158] Optionally, such as Figure 8 As shown, this application embodiment also provides an electronic device 90, including a processor 91 and a memory 92. The memory 92 stores a program or instructions that can run on the processor 91. When the program or instructions are executed by the processor 91, they implement the various steps of the above-mentioned recommended method embodiment and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0159] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.

[0160] Figure 9 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of this application.

[0161] The electronic device 100 includes, but is not limited to, components such as: radio frequency unit 101, network module 102, audio output unit 103, input unit 104, sensor 105, display unit 106, user input unit 107, interface unit 108, memory 109, and processor 110.

[0162] Those skilled in the art will understand that the electronic device 100 may also include a power supply (such as a battery) for supplying power to various components. The power supply may be logically connected to the processor 110 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system. Figure 9 The electronic device structure shown does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements, which will not be elaborated here.

[0163] The input unit 104 is used to input multimedia feature description information and user feature description information into the trained recommendation model. The trained recommendation model includes N cross-network modules, each corresponding to a different feature dimension, to obtain at least two first feature vectors. The processor 110 is used to input at least two first feature vectors into the first cross-network module, and cross-concatenate the at least two first feature vectors through the first feature dimension corresponding to the first cross-network module to obtain at least one second feature vector. The first cross-network module is one of the N cross-network modules. Based on at least one second feature vector, the processor outputs first recommendation information, where N is a positive integer.

[0164] This application provides an electronic device that, by cross-concatenating at least two first feature vectors through a first cross-network module to obtain at least one second feature vector, enhances feature cross-concatenation, improves model generalization, and thus improves the accuracy of video recommendation by enabling the electronic device to output first recommendation information from the at least two cross-concatenated first feature vectors.

[0165] Optionally, in this embodiment of the application, the processor 110 is further configured to input multimedia feature description information and user feature description information into the trained recommendation model, then concatenate the multimedia feature description information and user feature description information to obtain a third feature vector; and concatenate at least one second feature vector and the third feature vector to obtain a fourth feature vector; and output recommendation information based on the fourth feature vector.

[0166] Optionally, in this embodiment of the application, the trained recommendation model includes M expert network modules, each expert network module corresponding to at least one feature dimension, the feature dimensions corresponding to the M expert network modules include the feature dimensions corresponding to the N cross network modules, and M is an integer greater than 1; the processor 110 is specifically used to input multimedia feature description information and user feature description information into each expert network module respectively to obtain M first feature vectors, and one expert network module outputs one first feature vector.

[0167] The processor 110 is configured to: calculate a first value corresponding to the first cross-network module based on multimedia feature description information and user feature description information, wherein the first value represents the recommendation probability of each preset recommended video under the first feature dimension, and the first feature dimension is the feature dimension corresponding to the first cross-network module; calculate a second value based on the fourth feature vector, wherein the second value represents the recommendation probability of each preset recommended video; determine a first loss value corresponding to the first cross-network module based on the first and second values; weight the first loss value using the feature dimension weights of the first cross-network module to obtain a second loss value corresponding to the first cross-network module; and train a recommendation model based on the second loss value corresponding to each cross-network module to obtain the trained recommendation model; wherein the sum of the feature dimension weights of each cross-network module is 1.

[0168] This application provides an electronic device that can improve the prediction accuracy of multi-objective models and solve the problems of cold start and distribution of mid-to-long-tail materials by designing expert network cross sub-networks, shallow cross sub-structures of user-side features and material IDs, and up. Furthermore, it improves the Pareto-based multi-objective parameter solving method to resolve multi-objective learning conflicts.

[0169] Optionally, in this embodiment of the application, the processor 110 is further configured to, before weighting the first loss value using the feature dimension weights of the first cross-network module to obtain the second loss value corresponding to the first cross-network module, normalize the first gradient vector corresponding to each cross-network module to obtain the second gradient vector corresponding to each cross-network module. The first gradient vector corresponding to each cross-network module is determined based on the second loss value and the first feature vector corresponding to the cross-network module. According to Pareto theory, based on the second gradient vector corresponding to each cross-network module, the feature dimension weight vector is obtained, and based on the feature dimension weight vector, the feature dimension weight corresponding to each cross-network module is determined.

[0170] Optionally, in this embodiment of the application, the processor 110 is specifically used to calculate the feature dimension weight corresponding to each cross-network module based on the feature dimension weight vector when the feature dimension weight vector is non-negative; and to calculate the feature dimension weight corresponding to each cross-network module according to the Pareto rule when the feature dimension weight vector is negative; wherein, the Pareto rule includes: if the Pareto feasible solution falls on the Pareto boundary corresponding to the second feature dimension, then the feature dimension weight of the cross-network module corresponding to the second feature dimension is 0, and the second feature dimension is the feature dimension of at least one cross-network module among the N cross-network modules.

[0171] The electronic device provided in this application embodiment can implement the various processes implemented in the above method embodiments and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0172] The beneficial effects of the various implementation methods in this embodiment can be found in the beneficial effects of the corresponding implementation methods in the above method embodiments. To avoid repetition, they will not be repeated here.

[0173] It should be understood that, in this embodiment, the input unit 104 may include a graphics processing unit (GPU) 1041 and a microphone 1042. The GPU 1041 processes image data of still images or videos obtained by an image capture device (such as a camera) in video capture mode or image capture mode. The display unit 106 may include a display panel 1061, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like. The user input unit 107 includes at least one of a touch panel 1071 and other input devices 1072. The touch panel 1071 is also called a touch screen. The touch panel 1071 may include a touch detection device and a touch controller. Other input devices 1072 may include, but are not limited to, a physical keyboard, function keys (such as volume control buttons, power buttons, etc.), a trackball, a mouse, and a joystick, which will not be described in detail here.

[0174] The memory 109 can be used to store software programs and various data. The memory 109 may primarily include a first storage area for storing programs or instructions and a second storage area for storing data. The first storage area may store the operating system, application programs or instructions required for at least one function (such as sound playback, image playback, etc.). Furthermore, the memory 109 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), static random access memory (SRAM), 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), and direct memory bus RAM (DRRAM). The memory 109 in the embodiments of this application includes, but is not limited to, these and any other suitable types of memory.

[0175] Processor 110 may include one or more processing units; optionally, processor 110 integrates an application processor and a modem processor, wherein the application processor mainly handles operations involving the operating system, user interface, and applications, and the modem processor mainly handles wireless communication signals, such as a baseband processor. It is understood that the aforementioned modem processor may also not be integrated into processor 110.

[0176] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above method embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0177] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0178] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above method embodiments and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0179] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0180] This application provides a computer program product stored in a storage medium. The program product is executed by at least one processor to implement the various processes of the above-described recommended method embodiments and achieve the same technical effects. To avoid repetition, further details are omitted here.

[0181] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0182] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0183] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for obtaining recommendation information, characterized in that, The method includes: Multimedia feature description information and user feature description information are input into the trained recommendation model. The trained recommendation model includes N cross-network modules, each of which corresponds to a different feature dimension, to obtain at least two first feature vectors. The at least two first feature vectors are input into the first cross-network module. The at least two first feature vectors are cross-concatenated through the first feature dimension corresponding to the first cross-network module to obtain at least one second feature vector. The first cross-network module is one of the N cross-network modules. Based on the at least one second feature vector, output first recommendation information, where N is a positive integer; The trained recommendation model includes M expert network modules, each expert network module corresponds to at least one feature dimension, and the feature dimensions corresponding to the M expert network modules include the feature dimensions corresponding to the N cross network modules, where M is an integer greater than 1; After inputting multimedia feature description information and user feature description information into the trained recommendation model to obtain at least two first feature vectors, the method further includes: The multimedia feature description information and the user feature description information are concatenated to obtain a third feature vector; The at least one second feature vector and the third feature vector are concatenated to obtain a fourth feature vector; Based on the fourth feature vector, the recommendation information is output.

2. The method according to claim 1, characterized in that, Each of the expert network modules corresponds to at least one feature dimension, and the feature dimensions corresponding to the M expert network modules include the feature dimensions corresponding to the N cross-network modules, where M is an integer greater than 1; Based on the multimedia feature description information and the user feature description information, at least two first feature vectors are extracted, including: The multimedia feature description information and the user feature description information are respectively input into each of the expert network modules to obtain M first feature vectors, and each expert network module outputs one first feature vector.

3. A model training method, characterized in that, The method for obtaining recommendation information as described in claim 1 or 2, the method comprising: Based on multimedia feature description information and user feature description information, a first value corresponding to the first cross-network module is calculated. The first value is used to characterize the recommendation probability of each preset video under the first feature dimension. The first feature dimension is the feature dimension corresponding to the first cross-network module. Based on the fourth feature vector, a second value is calculated, which is used to characterize the recommendation probability of each preset video; Based on the first value and the second value, determine the first loss value corresponding to the first cross-network module; The first loss value is weighted by using the feature dimension weights of the first cross-network module to obtain the second loss value corresponding to the first cross-network module; Based on the second loss value corresponding to each of the cross-network modules, a recommendation model is trained to obtain the trained recommendation model. Wherein, the sum of the feature dimension weights of each of the cross-network modules is 1; Before weighting the first loss value using the feature dimension weights of the first cross-network module to obtain the second loss value corresponding to the first cross-network module, the method further includes: The first gradient vector corresponding to each of the cross-network modules is normalized to obtain the second gradient vector corresponding to each of the cross-network modules. The first gradient vector corresponding to each of the cross-network modules is determined based on the target loss value and the first feature vector corresponding to the cross-network module. According to Pareto theory, based on the second gradient vector corresponding to each of the cross-network modules, the feature dimension weight vector is obtained, and based on the feature dimension weight vector, the feature dimension weight corresponding to each of the cross-network modules is determined.

4. The method according to claim 3, characterized in that, The step of determining the feature dimension weights corresponding to each cross-network module based on the feature dimension weight vector includes: If the feature dimension weight vector is non-negative, calculate the feature dimension weight corresponding to each cross-network module based on the feature dimension weight vector. When the feature dimension weight vector is negative, the feature dimension weights corresponding to each cross-network module are calculated based on the feature dimension weight vector according to the Pareto rule. The Pareto rule includes the following: if a Pareto feasible solution falls on the Pareto boundary corresponding to the second feature dimension, then the feature dimension weight of the cross-network module corresponding to the second feature dimension is 0, and the second feature dimension is the feature dimension of at least one of the N cross-network modules.

5. A device for acquiring recommendation information, characterized in that, The device includes: an input module, an extraction module, and an output module; The input module is used to input multimedia feature description information and user feature description information into the trained recommendation model. The trained recommendation model includes N cross-network modules, each of which corresponds to a different feature dimension, and obtains at least two first feature vectors. The extraction module inputs at least two first feature vectors obtained by the input module into a first cross-network module, and performs cross-concatenation of the at least two first feature vectors through the first feature dimension corresponding to the first cross-network module to obtain at least one second feature vector. The first cross-network module is one of the N cross-network modules. The output module is used to output first recommendation information based on at least one second feature vector extracted by the extraction module, where N is a positive integer; The trained recommendation model includes M expert network modules, each expert network module corresponds to at least one feature dimension, and the feature dimensions corresponding to the M expert network modules include the feature dimensions corresponding to the N cross network modules, where M is an integer greater than 1; The device for obtaining recommendation information also includes a splicing module; The splicing module is used to input multimedia feature description information and user feature description information into the trained recommendation model from the input module, obtain at least two first feature vectors, splice the multimedia feature description information and the user feature description information to obtain a third feature vector; and splice the at least one second feature vector and the third feature vector to obtain a fourth feature vector. The output module is also used to output the recommendation information based on the fourth feature vector.

6. The apparatus according to claim 5, characterized in that, Each of the expert network modules corresponds to at least one feature dimension, and the feature dimensions corresponding to the M expert network modules include the feature dimensions corresponding to the N cross-network modules, where M is an integer greater than 1; the input module is specifically used to input the multimedia feature description information and the user feature description information into each of the expert network modules respectively to obtain M first feature vectors, and each expert network module outputs one first feature vector.

7. A model training device, characterized in that, The model training device, which is applied to the recommendation information acquisition device as described in claim 5 or 6, comprises: a calculation module, a determination module, a weighting module, and a training module; The calculation module is used to calculate a first value corresponding to the first cross-network module based on multimedia feature description information and user feature description information. The first value is used to represent the recommendation probability of each preset video under the first feature dimension, where the first feature dimension is the feature dimension corresponding to the first cross-network module. The module is also used to calculate a second value based on the fourth feature vector. The second value is used to represent the recommendation probability of each preset video. The determining module is used to determine the first loss value corresponding to the first cross-network module based on the first value and the second value obtained by the calculation module. The weighting module is used to weight the first loss value using the feature dimension weights of the first cross-network module to obtain the second loss value corresponding to the first cross-network module. The training module is used to train the recommendation model based on the second loss value corresponding to each of the cross-network modules, and obtain the trained recommendation model. Wherein, the sum of the feature dimension weights of each of the cross-network modules is 1; The model training device also includes a processing module; The processing module is configured to, before weighting the first loss value using the feature dimension weights of the first cross-network module to obtain the second loss value corresponding to the first cross-network module, normalize the first gradient vector corresponding to each cross-network module to obtain the second gradient vector corresponding to each cross-network module. The first gradient vector corresponding to each cross-network module is determined based on the target loss value and the first feature vector corresponding to the cross-network module. Furthermore, according to Pareto theory, based on the second gradient vector corresponding to each cross-network module, a feature dimension weight vector is obtained, and based on the feature dimension weight vector, the feature dimension weight corresponding to each cross-network module is determined.

8. The apparatus according to claim 7, characterized in that, The calculation module is specifically used to calculate the feature dimension weights corresponding to each of the cross-network modules based on the feature dimension weight vector when the feature dimension weight vector is non-negative; and to calculate the feature dimension weights corresponding to each of the cross-network modules according to the Pareto rule when the feature dimension weight vector is negative; wherein the Pareto rule includes: if the Pareto feasible solution falls on the Pareto boundary corresponding to the second feature dimension, then the feature dimension weight of the cross-network module corresponding to the second feature dimension is 0, and the second feature dimension is the feature dimension of at least one of the N cross-network modules.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein when the program or instructions are executed by the processor, they implement the steps of the method for obtaining recommendation information as described in claim 1 or 2, or the steps of the model training method as described in claim 3 or 4.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method for obtaining recommendation information as described in claim 1 or 2, or the steps of the model training method as described in claim 3 or 4.

Citation Information

Patent Citations

  • Recommendation model training method and device, electronic equipment and storage medium

    CN113553448A

  • Recommendation method, training method and device

    CN114997412A