An information processing method, device and computer readable storage medium

Abstract

Embodiments of the present application disclose an information processing method and device and a computer readable storage medium. Embodiments of the present application obtain a user feature vector converted from user feature information, a video feature vector converted from video feature information, and a promotion feature vector converted from promotion feature information. The user feature vector is subjected to attention fusion processing based on the video feature vector and the promotion feature vector, respectively, to obtain a user-video fusion vector and a user-promotion fusion vector. The user feature vector, the video feature vector, the promotion feature vector, the user-video fusion vector, and the user-promotion fusion vector are spliced to obtain a joint vector. The preset multi-task learning model is subjected to feature weighting training according to different task types based on the joint vector and label information, to obtain the preset multi-task learning model. The target video information and target promotion information having a correlation with the target video information greater than a preset threshold are displayed. In this way, the accuracy of information processing is improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to an information processing method, apparatus, and computer-readable storage medium. Background Technology

[0002] With the continuous development of artificial intelligence, recommendation systems are becoming increasingly intelligent. For example, when recommending advertisements or videos, the system can intelligently recommend the types of advertisements or videos that users are interested in based on their usage habits, thus achieving accurate recommendations.

[0003] However, traditional CTR (Click-Through-Rate) models only infer whether a user will click on an ad, without any correlation inference about whether they like the video associated with the ad. If inference is required, a separate video preference model needs to be created, which is cumbersome and the two cannot be processed together. Therefore, how to make accurate multi-dimensional recommendations for multi-task recommendation types still needs to be solved. Summary of the Invention

[0004] This application provides an information processing method, apparatus, and computer-readable storage medium, which can improve the accuracy of information processing.

[0005] To address the aforementioned technical problems, this application provides the following technical solutions: An information processing method, comprising: Obtain user feature vectors after converting user feature information, video feature vectors after converting video feature information, and promotion feature vectors after converting promotion feature information; Based on the video feature vector and the promotion feature vector, attention fusion processing is performed on the user feature vector to obtain the user video fusion vector and the user promotion fusion vector. The user feature vector, video feature vector, promotion feature vector, user video fusion vector, and user promotion fusion vector are concatenated to obtain a joint vector. Based on the joint vector and label information, the preset multi-task learning model is trained with feature weighting according to different task types to obtain the trained preset multi-task learning model. The display includes target video information and target promotional information whose correlation with the target video information is greater than a preset threshold. The target video information and target promotional information are obtained by pushing them to users through the trained preset multi-task learning model.

[0006] An information processing device, comprising: The acquisition unit is used to acquire the user feature vector after the conversion of user feature information, the video feature vector after the conversion of video feature information, and the promotion feature vector after the conversion of promotion feature information. An attention processing unit is used to perform attention fusion processing on the user feature vector based on the video feature vector and the promotion feature vector respectively, to obtain a user video fusion vector and a user promotion fusion vector. The splicing unit is used to splice the user feature vector, video feature vector, promotion feature vector, user video fusion vector, and user promotion fusion vector to obtain a joint vector; The training unit is used to perform feature-weighted training on the preset multi-task learning model according to different task types based on the joint vector and label information, so as to obtain the trained preset multi-task learning model. The display unit is used to display target video information and target promotional information whose correlation with the target video information is greater than a preset threshold. The target video information and target promotional information are obtained by pushing them to users through the trained preset multi-task learning model.

[0007] In some embodiments, the attention processing unit includes: The dimensionality reduction subunit is used to reduce the dimensionality of the video feature vector and the generalized feature vector through a preset fully connected layer to obtain a target video feature vector and a target generalized feature vector of a preset size. The processing subunit is used to perform attention fusion processing on the user feature vector according to the target video feature vector and the target promotion feature vector respectively, to obtain the user video fusion vector and the user promotion fusion vector.

[0008] In some embodiments, the processing subunit is configured to: Multiply each user feature domain vector in the user feature vector by the first preset matrix vector to obtain the corresponding number of transition vectors; Each transition vector is multiplied by the transpose of the target video feature vector to obtain the first weight value corresponding to each user feature domain vector; The user video fusion vector is obtained by weighting each user feature domain vector and its corresponding first weight value, and then averaging the weighted user feature domain vectors. Each transition vector is multiplied by the transpose of the target promotion feature vector to obtain the second weight value corresponding to each user feature domain vector. The user feature domain vector is weighted according to the corresponding second weight value, and the weighted user feature domain vectors are then averaged to obtain the user promotion fusion vector.

[0009] In some embodiments, the training unit includes: The input subunit is used to input the joint vector into a preset multi-task learning model to train multiple expert networks in the preset multi-task learning model, thereby obtaining multiple trained expert networks. Determine the sub-unit, which is used to determine the third weight value for each expert network under different task types; The weighted subunit is used to weight and connect the outputs of each expert network according to the third weight value corresponding to each task type. The output subunit is used to load the corresponding weighted connection output into the corresponding task training network in the preset multi-task learning model according to the task type, and output the target output result corresponding to each task type. The comparison subunit is used to compare the target output of each task type with the corresponding label information to obtain the difference value; The adjustment subunit is used to adjust the network parameters of the task training network according to the difference value until the difference value converges, thereby obtaining the trained preset multi-task learning model.

[0010] In some embodiments, the determining subunit is configured to: The joint vector is concatenated with the target video feature vector to obtain the first concatenation vector; Multiply the first connection vector by the second preset matrix vector to obtain the third weight value of each expert network under the video task type; The joint vector is concatenated with the target generalization feature vector to obtain the second connection vector; Multiply the second connection vector by the second preset matrix vector to obtain the third weight value corresponding to each expert network under the promotion task type.

[0011] In some embodiments, the weighting subunit is used for: The outputs of each expert network are weighted and connected according to the third weight value corresponding to the video task type. The outputs of each expert network are weighted and connected according to the third weight value corresponding to the promotion task type.

[0012] In some embodiments, the output subunit is configured to: According to the video task type, the output after the corresponding weighted connection is loaded into the first task training network in the preset multi-task learning model, and the first output result corresponding to the video task type is output. Obtain the first low-order cross feature corresponding to the video task type, input the first low-order cross feature into the factorization machine model, and output the second output result corresponding to the first low-order cross feature. According to the promotion task type, the output after the corresponding weighted connection is loaded into the second task training network in the preset multi-task learning model, and the third output result corresponding to the promotion task type is output. Obtain the second low-order cross feature corresponding to the promotion task type, input the second low-order cross feature into the factorization machine model, and output the fourth output result corresponding to the second low-order cross feature; Add the first output result and the second output result to obtain the target output result corresponding to the video task type; Add the third and fourth output results to obtain the target output result corresponding to the promotion task type.

[0013] In some embodiments, the acquisition unit is configured to: Acquire user characteristic information, video characteristic information, and promotion characteristic information; The feature identifier corresponding to each feature domain in the user feature information is vectorized to obtain the user feature domain vector corresponding to each feature domain. The user feature vector is obtained by concatenating the user feature domain vectors corresponding to each feature domain. The feature identifier corresponding to each feature domain in the video feature information is vectorized to obtain the video feature domain vector corresponding to each feature domain. The video feature vector is obtained by concatenating the video feature vectors corresponding to each feature domain. The feature identifier corresponding to each feature domain in the promotion feature information is vectorized to obtain the promotion feature domain vector corresponding to each feature domain. The generalized feature vector is obtained by concatenating the generalized feature vectors corresponding to each feature domain.

[0014] A computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to perform the steps in the above-described information processing method.

[0015] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the information processing method described above.

[0016] A computer program product or computer program includes computer instructions stored in a storage medium. A processor of a computer device reads the computer instructions from the storage medium and executes the computer instructions, causing the computer to perform the steps of the aforementioned information processing method.

[0017] This application embodiment obtains user feature vectors transformed from user feature information, video feature vectors transformed from video feature information, and promotion feature vectors transformed from promotion feature information; it performs attention fusion processing on the user feature vectors based on the video feature vectors and promotion feature vectors respectively to obtain user video fusion vectors and user promotion fusion vectors; it concatenates the user feature vectors, video feature vectors, promotion feature vectors, user video fusion vectors, and user promotion fusion vectors to obtain a joint vector; it trains a preset multi-task learning model according to different task types based on the joint vectors and tag information to obtain a trained preset multi-task learning model; and it displays target video information and target promotion information whose correlation with the target video information is greater than a preset threshold. This target video information and target promotion information are obtained by pushing information to users through the trained preset multi-task learning model. Therefore, by utilizing an attention mechanism, user feature vectors are fused with video feature vectors and promotion feature vectors respectively, resulting in user video fusion vectors and user promotion fusion vectors. These vectors represent the features that video and promotion tasks are more interested in from the user feature vectors, preserving the commonalities of multiple tasks while capturing the information relevant to each task. Furthermore, the user feature vectors, video feature vectors, promotion feature vectors, user video fusion vectors, and user promotion fusion vectors are concatenated to obtain a joint vector. This joint vector is then used to train a pre-defined multi-task learning model with feature weighting according to different task types, resulting in a trained pre-defined multi-task learning model. The model outputs target video information and target promotion information with a correlation greater than a preset threshold, which are then displayed. Compared to traditional CTR prediction methods, the pre-defined multi-task learning model trained in this application can learn by focusing on the features required by the task through a self-attention mechanism and can simultaneously predict multiple tasks, resulting in higher accuracy of the model output and significantly improving the accuracy of information processing. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of a scenario for the information processing system provided in an embodiment of this application; Figure 2 This is a flowchart illustrating the information processing method provided in an embodiment of this application; Figure 3This is another flowchart illustrating the information processing method provided in the embodiments of this application; Figure 4a A product schematic diagram illustrating the information processing method provided in the embodiments of this application; Figure 4b A schematic diagram of the architecture of the multi-task learning model provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the information processing device provided in the embodiments of this application; Figure 6 This is a schematic diagram of the server structure provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application provides an information processing method, apparatus, and computer-readable storage medium.

[0022] Please see Figure 1 , Figure 1 This is a schematic diagram of an information processing system provided in an embodiment of this application, including: terminal A and a server (the information processing system may also include other terminals besides terminal A; the specific number of terminals is not limited here). Terminal A and the server can be connected via a communication network, which may include wireless networks and wired networks. The wireless network includes one or more combinations of wireless wide area networks, wireless local area networks, wireless metropolitan area networks, and wireless personal networks. The network includes network entities such as routers and gateways, which are not shown in the diagram. Terminal A can interact with the server through the communication network; for example, terminal A can send user characteristic information, video characteristic information, and promotional characteristic information to the server through an instant messaging application.

[0023] It should be noted that during the technical implementation process of this application, all user characteristic information obtained through legal means has undergone strict anonymization processing to ensure that it does not contain any personally identifiable data (such as name, ID number, contact information, etc.). This information is used only for technical function implementation and data analysis, complying with the requirements of laws and regulations such as the Personal Information Protection Law and the Cybersecurity Law, and fundamentally eliminating the risk of privacy leakage through anonymization, de-identification, and other technical means. All data processing flows follow the principle of minimum necessity, and data security is ensured through security measures such as encrypted storage and access control isolation.

[0024] The information processing system may include an information processing device, which may be integrated into a server. This server may be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Figure 1 As shown, the server receives user feature information, video feature information, and promotion feature information sent by terminal A. It obtains the user feature vector (converted from user feature information), the video feature vector (converted from video feature information), and the promotion feature vector (converted from promotion feature information). Based on the video feature vector and promotion feature vector, it performs attention fusion processing on the user feature vector to obtain a user video fusion vector and a user promotion fusion vector. It then concatenates these vectors to obtain a joint vector. Based on the joint vector and label information, it trains a preset multi-task learning model with feature weighting according to different task types, resulting in a trained preset multi-task learning model. Based on this trained model, it pushes target video information and target promotion information with a correlation greater than a preset threshold to the terminal, enabling the terminal to display both the target video information and the promotion information with a correlation greater than the preset threshold, thus achieving simultaneous recommendation of video and promotional information.

[0025] In this information processing system, terminal A can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal A can install various applications needed by the user, such as video applications. Terminal A can collect user characteristic information, video characteristic information, and promotional characteristic information when the user is using the video application, and send them to the server. In addition, it can also receive target video information and target promotional information with a correlation greater than a preset threshold pushed to the user by the server based on a pre-trained multi-task learning model, and display them.

[0026] It should be noted that, Figure 1 The schematic diagram of the information processing system shown is merely an example. The information processing system and scenario described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of information processing systems and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0027] The following sections will provide detailed explanations.

[0028] In this embodiment, the description will be from the perspective of an information processing device, which can be integrated into a server that has storage units and is equipped with a microprocessor and has computing capabilities.

[0029] Please see Figure 2 , Figure 2 This is a flowchart illustrating the information processing method provided in an embodiment of this application. The information processing method includes: In step 101, the user feature vector after conversion of user feature information, the video feature vector after conversion of video feature information, and the promotion feature vector after conversion of promotion feature information are obtained.

[0030] In this embodiment, user feature information, video feature information, and promotion feature information can be collected in real time. The user feature information can consist of feature information from multiple user feature domains, such as user preference for videos, user gender, and recently clicked tags. The video feature information can consist of feature information from multiple video feature domains, such as video category and video duration. The promotion feature information can consist of feature information from multiple promotion feature domains, such as promotion information category and promotion information price range; this promotion information can be advertising information.

[0031] Furthermore, the feature information of each feature domain in the user feature information, video feature information, and promotion feature information can be processed through embedding (dimensionality reduction) to obtain a vector for each feature domain, which serves as the representation of the feature domain. This embedding can transform a large sparse vector into a low-dimensional space that preserves semantic relationships. The vectors of each feature domain in the user feature information are concatenated to obtain the user feature vector. The vectors of each feature domain in the video feature information are concatenated to obtain the video feature vector. The vectors of each feature domain in the promotion feature information are concatenated to obtain the promotion feature vector.

[0032] In some implementations, the step of obtaining the user feature vector, video feature vector, and promotion feature vector converted from user feature information, video feature information, and promotion feature information includes: (1) Obtain user characteristic information, video characteristic information, and promotion characteristic information; (2) Vectorize the feature identifier corresponding to each feature domain in the user feature information to obtain the user feature domain vector corresponding to each feature domain; (3) Concatenate the user feature vectors corresponding to each feature domain to obtain the user feature vector; (4) Vectorize the feature identifier corresponding to each feature domain in the video feature information to obtain the video feature domain vector corresponding to each feature domain. (5) Concatenate the video feature vectors corresponding to each feature domain to obtain the video feature vector; (6) Vectorize the feature identifier corresponding to each feature domain in the promotion feature information to obtain the promotion feature domain vector corresponding to each feature domain; (7) Concatenate the generalized feature vectors corresponding to each feature domain to obtain the generalized feature vector.

[0033] This involves acquiring user feature information, video feature information, and promotion feature information. Then, the feature identifier corresponding to each feature domain in the user feature information is vectorized to obtain the user feature domain vector corresponding to each feature domain. Each feature domain can contain multiple feature identifiers, which are values ​​representing the feature obtained by feature hashing. For example, the gender feature can be hashed to 0 or 1, where 0 can represent female and 1 can represent male. To better understand the embodiments of this application, the following example can be used as a reference: Assume that the user features consist of three feature domains: user's preference for videos (feature domain 1), user's gender (feature domain 2), and user's recently clicked tags (feature domain 3). Feature domain 1 contains 3 feature identifiers (IDs), feature domain 2 contains 1 feature identifier, and feature domain 3 contains 8 feature identifiers. The feature embedding size is set to 8. After embedding, feature domain 1 becomes 3 vectors of 1x8 dimensions, feature domain 2 becomes 1 vector of 1x8 dimensions, and feature domain 3 becomes 8 vectors of 1x8 dimensions. Then, average pooling is performed on the multiple vectors within each feature domain, resulting in a 1x8 dimension vector for each feature domain, representing its respective feature domain. The three feature domain vectors are concatenated to obtain the user feature vector. Similarly, the generation process for video feature vectors and promotion feature vectors is the same, and will not be elaborated here.

[0034] In step 102, attention fusion processing is performed on the user feature vector based on the video feature vector and the promotion feature vector respectively to obtain the user video fusion vector and the user promotion fusion vector.

[0035] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0036] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0037] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0038] The solutions provided in this application involve technologies such as machine learning in artificial intelligence, and are specifically illustrated through the following embodiments: The attention mechanism, in particular, focuses limited attention on key information, thereby conserving resources and quickly obtaining the most effective information. Specifically, the attention operation aims to better identify the unique points of interest related to user characteristics for different task types.

[0039] Based on this, the embodiments of this application perform attention fusion processing on the user feature vector based on the video feature vector and the promotion feature vector respectively, to obtain the user video fusion vector and the user promotion fusion vector. The user video fusion vector is the attention distribution information of the importance of each feature domain in the user feature vector in the video task type. User feature domains that are important to the video task type tend to have larger weights, indicating high attention, while user feature domains that are not important to the video task type tend to have smaller weights, indicating low attention.

[0040] Furthermore, the user promotion fusion vector represents the importance distribution information of each feature domain in the user feature vector within the promotion task type. User feature domains that are important to the promotion task type tend to have larger weights, indicating high attention, while user feature domains that are not important to the promotion task type tend to have smaller weights, indicating low attention.

[0041] Therefore, by using self-attention processing to represent the more important features in the user feature vector for video tasks and promotion tasks, we can achieve user video fusion vector and user promotion fusion vector. This can not only preserve the commonalities of multiple tasks, but also capture the information of interest for each task, making the subsequent training for multi-task processing more accurate.

[0042] In some implementations, the step of performing attention fusion processing on the user feature vector based on the video feature vector and the promotion feature vector respectively to obtain the user video fusion vector and the user promotion fusion vector may include: (1) The video feature vector and the generalized feature vector are reduced in dimensionality by a preset fully connected layer to obtain the target video feature vector and the target generalized feature vector of a preset size; (2) Based on the target video feature vector and the target promotion feature vector, attention fusion processing is performed on the user feature vector to obtain the user video fusion vector and the user promotion fusion vector.

[0043] The preset fully connected layer can be a single fully connected layer, through which the video feature vector and the promotion feature vector are reduced in dimensionality to the target video feature vector and the target promotion feature vector of a preset size, which can be 1 by 8 dimensions.

[0044] Furthermore, the target video feature vector and the target promotion feature vector are used as context vectors in the attention operation. Attention fusion is performed on the vectors of each feature domain in the user feature vector to obtain the target video feature vector based on the video task type and the target promotion feature vector based on the promotion task type of the user feature information, respectively.

[0045] In some implementations, the step of performing attention fusion processing on the user feature vector based on the target video feature vector and the target promotion feature vector respectively to obtain the user video fusion vector and the user promotion fusion vector includes: (1.1) Multiply each user feature domain vector in the user feature vector by the first preset matrix vector to obtain the corresponding number of transition vectors; (1.2) Multiply each transition vector by the transpose of the target video feature vector to obtain the first weight value corresponding to each user feature domain vector; (1.3) Weight each user feature domain vector and its corresponding first weight value, and then perform vector averaging on the weighted user feature domain vectors to obtain the user video fusion vector; (1.4) Multiply each transition vector by the transpose of the target promotion feature vector to obtain the second weight value corresponding to each user feature domain vector; (1.5) Weight each user feature domain vector and its corresponding second weight value, and then perform vector averaging on the weighted user feature domain vectors to obtain the user promotion fusion vector.

[0046] Understandably, the Attention operation aims to better identify the unique focus of user features for different tasks. For example, in a promotion prediction task, a user's gender plays a significant role, while their video preference has a smaller impact. Therefore, the weights obtained from the attention operation would be higher for the 1*8 dimensional vector representing the user's gender feature field, and much lower for the vector representing the user's video preference field. Conversely, in a video prediction task, a user's video preference might have a large impact, resulting in a high attention weight, while the user's gender feature might have a small impact, resulting in a low attention weight.

[0047] This process involves multiplying each user feature vector in the user feature vector by a first preset matrix vector, which can be an 8x8 matrix, to obtain a 1x8 transition vector (also called an intermediate vector) for each user feature domain. Each transition vector is then multiplied by the transpose of the target video feature vector from the previous step to obtain a score for each feature domain vector. All scores are then subjected to softmax (normalization) to obtain a first weight value for each feature domain vector. A higher weight value indicates a higher level of attention from the video task type to that feature domain vector, while a lower weight value indicates a lower level of attention from the video task type. Each user feature domain vector is weighted according to its corresponding first weight value, and the weighted user feature domain vectors are then averaged to obtain a user video fusion vector, which serves as the attention vector representation for the video task.

[0048] Furthermore, each transition vector is transposed and multiplied by the target promotion feature vector from the previous step to obtain a score for each feature domain vector. Softmax is applied to all scores to obtain a second weight value for each feature domain vector. A higher weight value indicates a higher level of attention the promotion task type pays to that feature domain vector, while a lower weight value indicates a lower level of attention. Each user feature domain vector is weighted according to its corresponding second weight value, and the weighted user feature domain vectors are then averaged (i.e., summed and averaged) to obtain a user promotion fusion vector, which serves as the attention vector representation for the promotion type task.

[0049] In step 103, the user feature vector, video feature vector, promotion feature vector, user video fusion vector, and user promotion fusion vector are concatenated to obtain a joint vector.

[0050] The user feature vector, video feature vector, promotion feature vector, user video fusion vector, and user promotion fusion vector can be concatenated using the concat layer to obtain a joint vector, which serves as the input to the subsequent multi-task learning model.

[0051] In step 104, the preset multi-task learning model is trained by feature weighting according to different task types based on the joint vector and label information to obtain the trained preset multi-task learning model.

[0052] In one embodiment, the preset multi-task learning model can be an MMOE (Multi-gate Mixture-of-Experts) model. This preset multi-task learning model can process tasks that are not closely related. For example, the preset multi-task learning model can recommend videos that users like, while also taking into account whether users are interested in the products associated with those videos.

[0053] The tag information refers to the tag information corresponding to each task type. Assuming there are two task types: video task type and promotion task type, the tag information for the video task type contains 0 or 1, where 0 indicates the user did not click on the video, and 1 indicates the user clicked on the video. The tag information for the promotion task type can also contain 0 or 1, where 0 indicates the user did not click on the promotional information, and 1 indicates the user clicked on the promotional information. The preset multi-task learning model is trained based on the joint vector and tag information. According to different task types, feature weighting training is performed using the user video fusion vector and user promotion fusion vector in the joint vector. Compared to existing multi-task learning models, this embodiment, through the user video fusion vector and user promotion fusion vector, can make the preset multi-task learning model focus more on user feature domains with positive benefits during training for different task types. The network parameters of the preset multi-task learning model are guided based on the difference between the tag information and the output results until the difference converges, resulting in the trained preset multi-task learning model.

[0054] In some implementations, the step of training a pre-defined multi-task learning model based on joint vectors and label information according to different task types with feature weighting to obtain the trained pre-defined multi-task learning model may include: (1) Input the joint vector into the preset multi-task learning model, train the multiple expert networks in the preset multi-task learning model, and obtain the trained multiple expert networks; (2) Determine the third weight value for each expert network under different task types; (3) The outputs of each expert network are weighted and connected according to the third weight value corresponding to each task type; (4) Load the corresponding weighted connection output into the corresponding task training network in the preset multi-task learning model according to the task type, and output the target output result corresponding to each task type; (5) Compare the target output of each task type with the corresponding label information to obtain the difference value; (6) Adjust the network parameters of the training network for the task according to the difference value until the difference value converges, and obtain the pre-set multi-task learning model after training.

[0055] The joint vector can be input into a preset multi-task learning model to train multiple expert networks in the preset multi-task learning model. The expert network is a standard deep neural network (DNN) structure. The number of expert networks can be set arbitrarily, and each expert network can learn different knowledge. For example, some expert networks learn more about video information preferences, while other expert networks learn more about promotional information preferences.

[0056] Furthermore, since different tasks focus on different aspects of each expert network's output—for example, in video tasks, expert networks that learn more about video information preferences are more focused on, while in promotional tasks, expert networks that learn more about promotional information preferences are more focused on—the third weight value for each expert network under different task types can be determined. This allows the pre-defined multi-task learning model to weight and connect the outputs of each expert network according to the corresponding third weight value for each task type. In this pre-defined multi-task learning model, a corresponding task training network is independently set for each task type. This task training network consists of several fully connected layers. The weighted outputs are loaded into the corresponding task training network according to the task type, which can output the target output result for each task type. The target output result can be between [0, 1], representing the predicted score. The closer the target output result is to 0, the closer the task is to a negative sample; the closer the target output result is to 1, the closer the task is to a positive sample.

[0057] The target output of each task type is compared with the corresponding label information to obtain the difference value. This difference value represents the degree of difference between the model's predicted value and the true value. Based on this difference value, the network parameters of the corresponding task training network are adjusted so that the task training network can make the prediction of the task more and more accurate. This continues until the difference value of the output of each task training network begins to converge, which means that the training is complete and the pre-trained multi-task learning model is obtained.

[0058] In step 105, target video information and target promotional information whose relevance to the target video information is greater than a preset threshold are displayed.

[0059] The trained, pre-defined multi-task learning model can infer not only whether a user will click on video information based on their user characteristics, but also whether they will click on the promotional information carried by that video. This allows the model to push target video information and promotional information highly relevant to the user's characteristics. In other words, it can push target video information that the user is interested in, and simultaneously push promotional information strongly related to that target video information. The preset threshold is the critical value for determining whether the target video information and promotional information are strongly correlated. When the correlation between the target video information and promotional information is greater than the preset threshold, they are considered strongly correlated, and the model can push video information that the user is interested in, while the user is also simultaneously interested in the promotional information carried by that video information. In one embodiment, the target video information and promotional information can be pushed to the terminal for real-time display, improving the accuracy of information push.

[0060] As described above, this embodiment of the application obtains user feature vectors transformed from user feature information, video feature vectors transformed from video feature information, and promotion feature vectors transformed from promotion feature information; it performs attention fusion processing on user feature vectors based on video feature vectors and promotion feature vectors respectively to obtain user video fusion vectors and user promotion fusion vectors; it concatenates user feature vectors, video feature vectors, promotion feature vectors, user video fusion vectors, and user promotion fusion vectors to obtain a joint vector; it trains a preset multi-task learning model according to different task types based on the joint vector and tag information to obtain a trained preset multi-task learning model; and it displays target video information and target promotion information whose correlation with the target video information is greater than a preset threshold. This target video information and target promotion information are obtained by pushing information to users through the trained preset multi-task learning model. Therefore, by utilizing an attention mechanism, user feature vectors are fused with video feature vectors and promotion feature vectors respectively, resulting in user video fusion vectors and user promotion fusion vectors. These vectors represent the features that video and promotion tasks are more interested in from the user feature vectors, preserving the commonalities of multiple tasks while capturing the information relevant to each task. Furthermore, the user feature vectors, video feature vectors, promotion feature vectors, user video fusion vectors, and user promotion fusion vectors are concatenated to obtain a joint vector. This joint vector is then used to train a pre-defined multi-task learning model with feature weighting according to different task types, resulting in a trained pre-defined multi-task learning model. The model outputs target video information and target promotion information with a correlation greater than a preset threshold, which are then displayed. Compared to traditional CTR prediction methods, the pre-defined multi-task learning model trained in this application can learn by focusing on the features required by the task through a self-attention mechanism and can simultaneously predict multiple tasks, resulting in higher accuracy of the model output and significantly improving the accuracy of information processing.

[0061] Based on the methods described in the above embodiments, the following examples will provide further detailed explanations.

[0062] In this embodiment, the information processing device will be specifically integrated into a server as an example for explanation. Please refer to the following description for details.

[0063] Please see Figure 3 , Figure 3 Another schematic flowchart of the information processing method provided in this application embodiment. The method flow may include: In step 201, the server obtains user feature information, video feature information, and promotion feature information.

[0064] For a better illustration of the embodiments of this application, please refer to Figure 4a As shown, Figure 4a This is a product illustration of the information processing method provided in the embodiments of this application. The product interface 10 includes a video playback area 11 and a push information display area 12. The push information can be advertising information, which is used to link to the product's sales interface. That is, when this product interface pushes information to users, it should recommend videos that users like in the video playback area 11, while also taking into account whether users are interested in the push information carried by the video.

[0065] To address the aforementioned issues and achieve more accurate push notifications, this embodiment of the application requires the server to acquire user characteristic information, video characteristic information, and promotional characteristic information. The user characteristic information can consist of feature information from multiple user characteristic domains, such as user video preference information, user gender characteristic information, and recently clicked tag characteristic information. The video characteristic information can consist of feature information from multiple video characteristic domains, such as video category characteristic information and video duration characteristic information. The promotional characteristic information can consist of feature information from multiple promotional characteristic domains, such as promotional information category characteristic information and promotional information price range characteristic information.

[0066] In step 202, the server vectorizes the feature identifier corresponding to each feature domain in the user feature information to obtain the user feature domain vector corresponding to each feature domain, and concatenates the user feature domain vectors corresponding to each feature domain to obtain the user feature vector.

[0067] Please also refer to Figure 4b , Figure 4b This is a schematic diagram of the architecture of the multi-task learning model provided in this application embodiment. The user features can be composed of three feature domains: user preference for videos (feature domain 1), user gender (feature domain 2), and user recently clicked tags (feature domain 3). Feature domain 1 contains 3 feature identifiers (IDs), feature domain 2 contains 1 feature identifier, and feature domain 3 contains 8 feature identifiers. The feature embedding size is set to 8. After embedding, feature domain 1 obtains 3 vectors of 1x8 dimensions, feature domain 2 obtains 1 vector of 1x8 dimensions, and feature domain 3 obtains 8 vectors of 1x8 dimensions. Then, average pooling is performed on the multiple vectors in each feature domain so that each feature domain obtains a 1*8 dimension vector as the expression of its respective feature domain. The three feature domain vectors are concatenated to obtain a 1x24 dimension user feature vector.

[0068] In step 203, the server vectorizes the feature identifier corresponding to each feature domain in the video feature information to obtain the video feature domain vector corresponding to each feature domain. The video feature domain vectors corresponding to each feature domain are concatenated to obtain the video feature vector. The server also vectorizes the feature identifier corresponding to each feature domain in the promotion feature information to obtain the promotion feature domain vector corresponding to each feature domain. The promotion feature domain vectors corresponding to each feature domain are concatenated to obtain the promotion feature vector.

[0069] Please refer to the derivation process of the user feature vector. The generation process of the video feature vector and the promotion feature vector is similar, and a 1x16 dimension video feature vector and a 1x16 dimension promotion feature vector can be obtained. The details will not be elaborated here.

[0070] In step 204, the server reduces the dimensionality of the video feature vector and the generalized feature vector by using a preset fully connected layer to obtain a target video feature vector and a target generalized feature vector of a preset size.

[0071] Please refer to the following: Figure 4b As shown, the server reduces the dimensionality of the video feature vector and the promotion feature vector to a preset size of 1x8 through a fully connected layer.

[0072] In step 205, the server multiplies each user feature domain vector in the user feature vector by a first preset matrix vector to obtain a corresponding number of transition vectors. Each transition vector is then multiplied by the transpose of the target video feature vector to obtain a first weight value corresponding to each user feature domain vector. The user feature domain vectors are weighted according to their first weight values, and the weighted user feature domain vectors are then averaged to obtain the user video fusion vector.

[0073] Please refer to the following: Figure 4b As shown, the server multiplies the 1x8 dimension vector of each user feature domain in the user feature vector with an 8x8 dimension first preset matrix vector to obtain a 1x8 transition vector corresponding to each user feature domain. Each transition vector is then multiplied by the transpose of the target video feature vector from the previous step to obtain a score for each feature domain vector. All scores are then subjected to softmax (normalization) to obtain a first weight value for each feature domain vector. A higher weight value indicates a higher level of attention from the video task type to that feature domain vector. A lower weight value indicates a lower level of attention from the video task type to that feature domain vector. Each user feature domain vector is weighted according to its corresponding first weight value, and the weighted user feature domain vectors are then averaged to obtain a user video fusion vector, which serves as the self-attention vector representation for the video type task.

[0074] In step 206, the server multiplies each transition vector by the transpose of the target promotion feature vector to obtain the second weight value corresponding to each user feature domain vector. The server then weights each user feature domain vector and its corresponding second weight value, and performs vector averaging on the weighted user feature domain vectors to obtain the user promotion fusion vector.

[0075] Please refer to the following: Figure 4b As shown, each transition vector is multiplied by the transpose of the target promotion feature vector from the previous step to obtain a score for each feature domain vector. Softmax is applied to all scores to obtain a second weight value for each feature domain vector. A higher weight value indicates a higher level of attention the promotion task type pays to that feature domain vector, while a lower weight value indicates a lower level of attention. Each user feature domain vector is weighted according to its corresponding second weight value, and the weighted user feature domain vectors are then averaged to obtain a user promotion fusion vector, which serves as the self-attention vector representation for the promotion type task.

[0076] In step 207, the server concatenates the user feature vector, video feature vector, promotion feature vector, user video fusion vector, and user promotion fusion vector to obtain a joint vector.

[0077] Please refer to the following: Figure 4b As shown, the server can concatenate the user feature vector, video feature vector, promotion feature vector, user-video fusion vector, and user-promotion fusion vector through the concat input layer to obtain a joint vector, which serves as the input to the subsequent multi-task learning model.

[0078] In step 208, the server inputs the joint vector into a preset multi-task learning model and trains multiple expert networks in the preset multi-task learning model to obtain multiple trained expert networks.

[0079] Please refer to the following: Figure 4b As shown, the server can input the joint vector into multiple expert networks in a pre-defined multi-task learning model for training. These expert networks are standard deep neural networks, and there can be three expert networks. Two experts learn more about promotional information, while the third expert learns more about video information. The trained expert networks can then output the corresponding learned results.

[0080] In step 209, the server concatenates the joint vector with the target video feature vector to obtain a first connection vector, multiplies the first connection vector by a second preset matrix vector to obtain the third weight value of each expert network under the video task type, concatenates the joint vector with the target promotion feature vector to obtain a second connection vector, and multiplies the second connection vector by the second preset matrix vector to obtain the third weight value of each expert network under the promotion task type.

[0081] Please refer to the following: Figure 4b As shown, since different tasks focus on different aspects of each expert network's output, an attention-gated network can be used to connect the joint vector and the target video feature vector to obtain a first connection vector. This first connection vector is then multiplied by a second preset matrix vector. Assuming the first connection vector is a 1x50 dimension vector and the second preset matrix vector is a 50x3 dimension vector, three scores can be obtained. These three scores are then normalized to obtain a third weight value for each expert network under the video task type. This weight value represents the degree of attention paid by the three expert networks to the video task type. The higher the weight value, the higher the degree of attention and the better the benefit to the task processing.

[0082] Furthermore, an attention-gated network can be used to connect the joint vector and the target promotion feature vector to obtain a second connection vector. This second connection vector is then multiplied by a second preset matrix vector. Assuming the second connection vector is a 1x50 dimension vector and the second preset matrix vector is a 50x3 dimension vector, three scores can be obtained. These three scores are then normalized to obtain a third weight value for each expert network under the promotion task type. This weight value represents the attention paid by the three expert networks to the promotion task type. The higher the weight value, the higher the attention and the better the benefit to the task processing.

[0083] In step 210, the server performs weighted concatenation of the outputs of each expert network according to the third weight value corresponding to the video task type, and performs weighted concatenation of the outputs of each expert network according to the third weight value corresponding to the promotion task type.

[0084] Please refer to the following: Figure 4b As shown, the server performs weighted concatenation of the outputs of each expert network according to the third weight value corresponding to the video type through task A Merge Layer, and performs weighted concatenation of the outputs of each expert network according to the third weight value corresponding to the promotion task type through task B Merge Layer.

[0085] In step 211, the server loads the output after the weighted connection according to the video task type into the first task training network in the preset multi-task learning model, outputs the first output result corresponding to the video task type, obtains the first low-order cross feature corresponding to the video task type, inputs the first low-order cross feature into the factorization machine model, and outputs the second output result corresponding to the first low-order cross feature.

[0086] It is understandable that, since the preset multi-task learning model only has high-order cross features and no low-order cross features, in order to better assist the preset multi-task learning model in fitting and avoid losing effective information, this application embodiment introduces low-order cross features. Please refer to the following steps for details: Please refer to the following: Figure 4b As shown, the server can load the output after weighted connection according to the video task type into the first task training network (TowerA) in the preset multi-task learning model through the task merging layer A, and output the first output result corresponding to the video task type. In order to introduce low-order cross features for combination calculation, the first low-order cross feature corresponding to the video task type can be obtained. The first low-order cross feature can include the user video fusion vector and the video feature vector. The user video fusion vector and the video feature vector are input into the video factor machine (FM) model, and the second output result predicted by the first low-order cross feature is output.

[0087] In step 212, the server loads the output after the weighted connection according to the promotion task type into the second task training network in the preset multi-task learning model, outputs the third output result corresponding to the promotion task type, obtains the second low-order cross feature corresponding to the promotion task type, inputs the second low-order cross feature into the factorization machine model, and outputs the fourth output result corresponding to the second low-order cross feature.

[0088] Please refer to the following: Figure 4b As shown, the server can load the output after weighted connection according to the promotion task type into the second task training network B (TowerB) in the preset multi-task learning model through the task merging layer B, and output the third output result corresponding to the promotion task type. In order to introduce low-order cross features for combination calculation, the second low-order cross features corresponding to the promotion task type can be obtained. The second low-order cross features can include the user promotion fusion vector and the promotion feature vector. The user promotion fusion vector and the promotion feature vector are input into the factorization machine (adsfactor machine, FM) model, and the fourth output result predicted by the second low-order cross features is output.

[0089] In step 213, the server adds the first output result and the second output result to obtain the target output result corresponding to the video task type, adds the third output result and the fourth output result to obtain the target output result corresponding to the promotion task type, compares the target output result of each task type with the corresponding label information to obtain the difference value, and adjusts the network parameters of the task training network according to the difference value until the difference value converges to obtain the pre-trained multi-task learning model.

[0090] Please refer to the following: Figure 4b As shown, the server adds the first output result and the second output result to obtain the target output result corresponding to the video task type. Since the output information corresponding to the low-order cross features is introduced into the target output result corresponding to the video task type, the preset multi-task model can better fit the data and avoid losing the effective information corresponding to the low-order feature data, so that the prediction result corresponding to the video task type will be more accurate.

[0091] Furthermore, the server adds the third and fourth output results to obtain the target output result corresponding to the promotion task type, thereby introducing low-order cross-feature information into the target output result corresponding to the promotion task type, making the prediction result corresponding to the promotion task type more accurate.

[0092] The label information corresponds to each task type. There are two task types: video task type and promotion task type. The label information for the video task type contains 0 or 1, where 0 indicates the user did not click on the video, and 1 indicates the user clicked on the video. The label information for the promotion task type can also contain 0 or 1, where 0 indicates the user did not click on the promotional information, and 1 indicates the user clicked on the promotional information. The target output results of the video task type and the promotion task type are compared with the corresponding label information to obtain the difference value. The network parameters of the training networks for the two tasks are adjusted according to the difference value until the difference value converges, resulting in the trained preset multi-task learning model.

[0093] In step 214, the server displays the target video information and the target promotion information whose relevance to the target video information is greater than a preset threshold.

[0094] The pre-trained multi-task learning model can infer not only whether a user will click on a video based on their user characteristics, but also whether they will click on the promotional information accompanying the video. This allows the model to push not only target videos that the user is interested in, but also highly relevant promotional information. In other words, it can push videos that the user is interested in, and the user is also interested in the promotional information accompanying those videos. The server can directly display the target video and the promotional information whose relevance to the target video is greater than a preset threshold, and stream this display to the corresponding user terminal for real-time display, thus improving the accuracy of information push.

[0095] In some implementations, because the click-through rate of ads is very low, only 0.35%, the sample ratio is unbalanced. Therefore, Focal Loss is used to solve this imbalance problem, so that the model pays more attention to positive samples.

[0096] As described above, this embodiment of the application obtains user feature vectors transformed from user feature information, video feature vectors transformed from video feature information, and promotion feature vectors transformed from promotion feature information; it performs attention fusion processing on user feature vectors based on video feature vectors and promotion feature vectors respectively to obtain user video fusion vectors and user promotion fusion vectors; it concatenates user feature vectors, video feature vectors, promotion feature vectors, user video fusion vectors, and user promotion fusion vectors to obtain a joint vector; it trains a preset multi-task learning model according to different task types based on the joint vector and tag information to obtain a trained preset multi-task learning model; and it displays target video information and target promotion information whose correlation with the target video information is greater than a preset threshold. This target video information and target promotion information are obtained by pushing information to users through the trained preset multi-task learning model. Therefore, by utilizing an attention mechanism, user feature vectors are fused with video feature vectors and promotion feature vectors respectively, resulting in user video fusion vectors and user promotion fusion vectors. These vectors represent the features that video and promotion tasks are more interested in from the user feature vectors, preserving the commonalities of multiple tasks while capturing the information relevant to each task. Furthermore, the user feature vectors, video feature vectors, promotion feature vectors, user video fusion vectors, and user promotion fusion vectors are concatenated to obtain a joint vector. This joint vector is then used to train a pre-defined multi-task learning model with feature weighting according to different task types, resulting in a trained pre-defined multi-task learning model. The model outputs target video information and target promotion information with a correlation greater than a preset threshold, which are then displayed. Compared to traditional CTR prediction methods, the pre-defined multi-task learning model trained in this application can learn by focusing on the features required by the task through a self-attention mechanism and can simultaneously predict multiple tasks, resulting in higher accuracy of the model output and significantly improving the accuracy of information processing.

[0097] Further experiments revealed that the model in this embodiment significantly outperforms the MMoE model. Specific experimental data is as follows:

[0098] To facilitate better implementation of the information processing method provided in the embodiments of this application, the embodiments of this application also provide an apparatus based on the above-described information processing method. The meanings of the terms used are the same as in the above-described information processing method, and specific implementation details can be found in the descriptions in the method embodiments.

[0099] Please see Figure 5 , Figure 5This is a schematic diagram of the structure of an information processing device provided in an embodiment of this application. The information processing device may include an acquisition unit 301, an attention processing unit 302, a splicing unit 303, a training unit 304, and a display unit 305, etc.

[0100] The acquisition unit 301 is used to acquire the user feature vector after the conversion of user feature information, the video feature vector after the conversion of video feature information, and the promotion feature vector after the conversion of promotion feature information.

[0101] In some embodiments, the acquisition unit 301 is used for: Acquire user characteristic information, video characteristic information, and promotion characteristic information; The feature identifier corresponding to each feature domain in the user feature information is vectorized to obtain the user feature domain vector corresponding to each feature domain. The user feature vector is obtained by concatenating the user feature domain vectors corresponding to each feature domain. The feature identifier corresponding to each feature domain in the video feature information is vectorized to obtain the video feature domain vector corresponding to each feature domain. The video feature vector is obtained by concatenating the video feature vectors corresponding to each feature domain. The feature identifier corresponding to each feature domain in the promotion feature information is vectorized to obtain the promotion feature domain vector corresponding to each feature domain. The generalized feature vector is obtained by concatenating the generalized feature vectors corresponding to each feature domain.

[0102] Attention processing unit 302 is used to perform attention fusion processing on the user feature vector based on the video feature vector and the promotion feature vector respectively, to obtain user video fusion vector and user promotion fusion vector.

[0103] In some embodiments, the attention processing unit 302 includes: The dimension reduction subunit is used to reduce the dimension of the video feature vector and the generalized feature vector through a preset fully connected layer to obtain the target video feature vector and the target generalized feature vector of a preset size. The processing subunit is used to perform attention fusion processing on the user feature vector based on the target video feature vector and the target promotion feature vector, respectively, to obtain the user video fusion vector and the user promotion fusion vector.

[0104] In some implementations, the processing subunit is used for: Multiply each user feature domain vector in the user feature vector by the first preset matrix vector to obtain the corresponding number of transition vectors; Each transition vector is multiplied by the transpose of the target video feature vector to obtain the first weight value corresponding to each user feature domain vector. The user video fusion vector is obtained by weighting each user feature domain vector and its corresponding first weight value, and then averaging the weighted user feature domain vectors. Each transition vector is multiplied by the transpose of the target promotion feature vector to obtain the second weight value corresponding to each user feature domain vector; The user feature domain vector is weighted according to the corresponding second weight value, and the weighted user feature domain vectors are then averaged to obtain the user promotion fusion vector.

[0105] The splicing unit 303 is used to splice the user feature vector, video feature vector, promotion feature vector, user video fusion vector and user promotion fusion vector to obtain a joint vector.

[0106] Training unit 304 is used to perform feature-weighted training on the preset multi-task learning model according to different task types based on the joint vector and label information, so as to obtain the trained preset multi-task learning model.

[0107] In some implementations, the training unit 304 includes: The input sub-unit is used to input the joint vector into a preset multi-task learning model, train multiple expert networks in the preset multi-task learning model, and obtain multiple trained expert networks. Determine the sub-unit, which is used to determine the third weight value for each expert network under different task types; The weighted subunit is used to weight and connect the outputs of each expert network according to the third weight value corresponding to each task type. The output subunit is used to load the corresponding weighted connection output into the corresponding task training network in the preset multi-task learning model according to the task type, and output the target output result corresponding to each task type. The comparison subunit is used to compare the target output of each task type with the corresponding label information to obtain the difference value; The adjustment sub-unit is used to adjust the network parameters of the task training network according to the difference value until the difference value converges, thus obtaining the pre-set multi-task learning model after training.

[0108] In some implementations, the determining subunit is used for: The joint vector is concatenated with the target video feature vector to obtain the first concatenation vector; Multiply the first connection vector by the second preset matrix vector to obtain the third weight value of each expert network under the video task type; The joint vector is concatenated with the target generalization feature vector to obtain the second concatenation vector; Multiply the second connection vector by the second preset matrix vector to obtain the third weight value of each expert network under the promotion task type.

[0109] In some implementations, the weighting subunit is used for: The outputs of each expert network are weighted and connected according to the third weight value corresponding to the video task type. The outputs of each expert network are weighted and connected according to the third weight value corresponding to the promotion task type.

[0110] In some implementations, the output subunit is used for: According to the video task type, the output after the corresponding weighted connection is loaded into the first task training network in the preset multi-task learning model, and the first output result corresponding to the video task type is output. Obtain the first low-order cross feature corresponding to the video task type, input the first low-order cross feature into the factorization machine model, and output the second output result corresponding to the first low-order cross feature; According to the promotion task type, the output after the corresponding weighted connection is loaded into the second task training network in the preset multi-task learning model, and the third output result corresponding to the promotion task type is output. Obtain the second low-order cross feature corresponding to the promotion task type, input the second low-order cross feature into the factorization machine model, and output the fourth output result corresponding to the second low-order cross feature; Add the first output result and the second output result to obtain the target output result corresponding to the video task type; Add the third and fourth output results to obtain the target output result corresponding to the promotion task type.

[0111] Display unit 305 is used to display target video information and target promotion information whose correlation with the target video information is greater than a preset threshold. The target video information and target promotion information are obtained by pushing them to users through the preset multi-task learning model after training.

[0112] The specific implementation of each of the above units can be found in the previous embodiments, and will not be repeated here.

[0113] As described above, in this embodiment, the acquisition unit 301 acquires the user feature vector after the transformation of user feature information, the video feature vector after the transformation of video feature information, and the promotion feature vector after the transformation of promotion feature information; the attention processing unit 302 performs attention fusion processing on the user feature vector based on the video feature vector and the promotion feature vector respectively to obtain the user video fusion vector and the user promotion fusion vector; the splicing unit 303 splices the user feature vector, video feature vector, promotion feature vector, user video fusion vector, and user promotion fusion vector to obtain a joint vector; the training unit 304 performs feature weighted training on the preset multi-task learning model according to different task types based on the joint vector and label information to obtain the trained preset multi-task learning model; and the display unit 305 displays the target video information and the target promotion information whose correlation with the target video information is greater than a preset threshold. The target video information and the target promotion information are obtained by pushing the preset multi-task learning model to the user. Therefore, by utilizing an attention mechanism, user feature vectors are fused with video feature vectors and promotion feature vectors respectively, resulting in user video fusion vectors and user promotion fusion vectors. These vectors represent the features that video and promotion tasks are more interested in from the user feature vectors, preserving the commonalities of multiple tasks while capturing the information relevant to each task. Furthermore, the user feature vectors, video feature vectors, promotion feature vectors, user video fusion vectors, and user promotion fusion vectors are concatenated to obtain a joint vector. This joint vector is then used to train a pre-defined multi-task learning model with feature weighting according to different task types, resulting in a trained pre-defined multi-task learning model. The model outputs target video information and target promotion information with a correlation greater than a preset threshold, which are then displayed. Compared to traditional CTR prediction methods, the pre-defined multi-task learning model trained in this application can learn by focusing on the features required by the task through a self-attention mechanism and can simultaneously predict multiple tasks, resulting in higher accuracy of the model output and significantly improving the accuracy of information processing.

[0114] This application also provides a computer device, such as... Figure 6 As shown, it illustrates a schematic diagram of the server structure involved in an embodiment of this application. Specifically: The computer device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 6 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein: Processor 401 is the control center of the computer device. It connects various parts of the computer device via various interfaces and lines. By running or executing software programs and / or modules stored in memory 402, and by calling data stored in memory 402, it performs various functions of the computer device and processes data, thereby performing overall detection of the computer device. Optionally, processor 401 may include one or more processing cores; optionally, processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the aforementioned modem processor may also not be integrated into processor 401.

[0115] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the server, etc. In addition, the memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0116] The computer equipment also includes a power supply 403 that supplies power to the various components. Optionally, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0117] The computer device may also include an input unit 404, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0118] Although not shown, the computer device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the computer device loads the executable files corresponding to the processes of one or more applications into the memory 402 according to the following instructions, and the processor 401 runs the applications stored in the memory 402, thereby implementing the various method steps provided in the foregoing embodiments, as follows: The system acquires user feature vectors (converted from user feature information), video feature vectors (converted from video feature information), and promotion feature vectors (converted from promotion feature information). Based on these video and promotion feature vectors, it performs attention fusion processing on the user feature vector to obtain a user-video fusion vector and a user-promotion fusion vector. It then concatenates these vectors to obtain a joint vector. Based on this joint vector and tag information, it trains a pre-defined multi-task learning model using feature weighting according to different task types, resulting in a trained pre-defined multi-task learning model. Finally, it displays target video information and target promotion information whose relevance to the target video information is greater than a preset threshold. This target video information and target promotion information are obtained by pushing content to users through the trained pre-defined multi-task learning model.

[0119] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the detailed description of the information processing method above, which will not be repeated here.

[0120] As described above, the computer device in this application embodiment can obtain user feature vectors transformed from user feature information, video feature vectors transformed from video feature information, and promotion feature vectors transformed from promotion feature information; perform attention fusion processing on the user feature vectors based on the video feature vectors and promotion feature vectors respectively to obtain user video fusion vectors and user promotion fusion vectors; concatenate the user feature vectors, video feature vectors, promotion feature vectors, user video fusion vectors, and user promotion fusion vectors to obtain a joint vector; perform feature weighted training on a preset multi-task learning model according to different task types based on the joint vectors and label information to obtain a trained preset multi-task learning model; and display target video information and target promotion information whose correlation with the target video information is greater than a preset threshold. The target video information and target promotion information are obtained by pushing information to users through the trained preset multi-task learning model. Therefore, by utilizing an attention mechanism, user feature vectors are fused with video feature vectors and promotion feature vectors respectively, resulting in user video fusion vectors and user promotion fusion vectors. These vectors represent the features that video and promotion tasks are more interested in from the user feature vectors, preserving the commonalities of multiple tasks while capturing the information relevant to each task. Furthermore, the user feature vectors, video feature vectors, promotion feature vectors, user video fusion vectors, and user promotion fusion vectors are concatenated to obtain a joint vector. This joint vector is then used to train a pre-defined multi-task learning model with feature weighting according to different task types, resulting in a trained pre-defined multi-task learning model. The model outputs target video information and target promotion information with a correlation greater than a preset threshold, which are then displayed. Compared to traditional CTR prediction methods, the pre-defined multi-task learning model trained in this application can learn by focusing on the features required by the task through a self-attention mechanism and can simultaneously predict multiple tasks, resulting in higher accuracy of the model output and significantly improving the accuracy of information processing.

[0121] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0122] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the information processing methods provided in embodiments of this application. For example, the instructions can execute the following steps: The system acquires user feature vectors (converted from user feature information), video feature vectors (converted from video feature information), and promotion feature vectors (converted from promotion feature information). Based on these video and promotion feature vectors, it performs attention fusion processing on the user feature vector to obtain a user-video fusion vector and a user-promotion fusion vector. It then concatenates these vectors to obtain a joint vector. Based on this joint vector and tag information, it trains a pre-defined multi-task learning model using feature weighting according to different task types, resulting in a trained pre-defined multi-task learning model. Finally, it displays target video information and target promotion information whose relevance to the target video information is greater than a preset threshold. This target video information and target promotion information are obtained by pushing content to users through the trained pre-defined multi-task learning model.

[0123] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations of the above embodiments.

[0124] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0125] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0126] Since the instructions stored in the computer-readable storage medium can execute the steps of any of the information processing methods provided in the embodiments of this application, the beneficial effects that any of the information processing methods provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0127] The above provides a detailed description of an information processing method, apparatus, and computer-readable storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An information processing method, characterized in that, include: Obtain user feature vectors after converting user feature information, video feature vectors after converting video feature information, and promotion feature vectors after converting promotion feature information; Based on the video feature vector and the promotion feature vector, attention fusion processing is performed on the user feature vector to obtain the user video fusion vector and the user promotion fusion vector. The user feature vector, video feature vector, promotion feature vector, user video fusion vector, and user promotion fusion vector are concatenated to obtain a joint vector. Based on the joint vector and label information, a preset multi-task learning model is trained using feature weighting according to different task types to obtain a trained preset multi-task learning model. The preset multi-task learning model includes a first task training network for inferring whether a user will click on video information and a second task training network for inferring whether a user will click on promotional information carried by the video information. The promotional information is advertising information. The preset multi-task learning model uses Focal Loss. The joint vector is input into the preset multi-task learning model, and multiple expert networks in the preset multi-task learning model are trained to obtain multiple trained expert networks. The joint vector is connected to the target video feature vector and multiplied by a second preset matrix vector through an attention mechanism gating network to obtain a third weight value corresponding to each expert network under the video task type. The joint vector is then connected to the target promotional feature vector and multiplied by the second preset matrix vector to obtain a third weight value corresponding to each expert network under the promotional task type. The outputs of each expert network are weighted and connected according to the third weight value corresponding to each task type. The weighted and connected outputs are loaded into the first task training network according to the video task type to output a first output result. The user video fusion vector and video feature vector are input into the factorization machine model, which outputs a second output result. The first and second output results are added together to obtain the target output result corresponding to the video task type. The output after weighted connection according to the promotion task type is loaded into the second task training network, which outputs a third output result. The user promotion fusion vector and promotion feature vector are input into the factorization machine model, which outputs a fourth output result. The third and fourth output results are added together to obtain the target output result corresponding to the promotion task type. The target output result of each task type is compared with the corresponding label information to obtain the difference value. The network parameters of the task training network are adjusted according to the difference value until the difference value converges, thus obtaining the trained preset multi-task learning model. The display includes target video information and target promotional information whose correlation with the target video information is greater than a preset threshold. The target video information and target promotional information are obtained by pushing them to users through the trained preset multi-task learning model.

2. The information processing method according to claim 1, characterized in that, The step of performing attention fusion processing on the user feature vector based on the video feature vector and the promotion feature vector respectively to obtain the user video fusion vector and the user promotion fusion vector includes: The video feature vector and the promotion feature vector are reduced in dimensionality by a preset fully connected layer to obtain a target video feature vector and a target promotion feature vector of a preset size. Attention fusion processing is performed on the user feature vector based on the target video feature vector and the target promotion feature vector to obtain the user video fusion vector and the user promotion fusion vector.

3. The information processing method according to claim 2, characterized in that, The step of performing attention fusion processing on the user feature vector based on the target video feature vector and the target promotion feature vector respectively to obtain the user video fusion vector and the user promotion fusion vector includes: Multiply each user feature domain vector in the user feature vector by the first preset matrix vector to obtain the corresponding number of transition vectors; Each transition vector is multiplied by the transpose of the target video feature vector to obtain the first weight value corresponding to each user feature domain vector; The user video fusion vector is obtained by weighting each user feature domain vector and its corresponding first weight value, and then averaging the weighted user feature domain vectors. Each transition vector is multiplied by the transpose of the target promotion feature vector to obtain the second weight value corresponding to each user feature domain vector. The user feature domain vector is weighted according to the corresponding second weight value, and the weighted user feature domain vectors are then averaged to obtain the user promotion fusion vector.

4. The information processing method according to claim 1, characterized in that, The step of weighting and connecting the outputs of each expert network according to the third weight value corresponding to each task type includes: The outputs of each expert network are weighted and connected according to the third weight value corresponding to the video task type. The outputs of each expert network are weighted and connected according to the third weight value corresponding to the promotion task type.

5. The information processing method according to any one of claims 1 to 4, characterized in that, The steps of obtaining the user feature vector after conversion of user feature information, the video feature vector after conversion of video feature information, and the promotion feature vector after conversion of promotion feature information include: Acquire user characteristic information, video characteristic information, and promotion characteristic information; The feature identifier corresponding to each feature domain in the user feature information is vectorized to obtain the user feature domain vector corresponding to each feature domain. The user feature vector is obtained by concatenating the user feature domain vectors corresponding to each feature domain. The feature identifier corresponding to each feature domain in the video feature information is vectorized to obtain the video feature domain vector corresponding to each feature domain. The video feature vector is obtained by concatenating the video feature vectors corresponding to each feature domain. The feature identifier corresponding to each feature domain in the promotion feature information is vectorized to obtain the promotion feature domain vector corresponding to each feature domain. The generalized feature vector is obtained by concatenating the generalized feature vectors corresponding to each feature domain.

6. An information processing device, characterized in that, include: The acquisition unit is used to acquire the user feature vector after the conversion of user feature information, the video feature vector after the conversion of video feature information, and the promotion feature vector after the conversion of promotion feature information. An attention processing unit is used to perform attention fusion processing on the user feature vector based on the video feature vector and the promotion feature vector respectively, to obtain a user video fusion vector and a user promotion fusion vector. The splicing unit is used to splice the user feature vector, video feature vector, promotion feature vector, user video fusion vector, and user promotion fusion vector to obtain a joint vector; The training unit is used to train a preset multi-task learning model according to different task types based on the joint vector and label information, thereby obtaining a trained preset multi-task learning model. The preset multi-task learning model includes a first task training network for inferring whether a user will click on video information and a second task training network for inferring whether a user will click on promotional information carried by the video information. The promotional information is advertising information. The preset multi-task learning model uses Focal Loss. The joint vector is input into the preset multi-task learning model to train multiple expert networks, resulting in trained expert networks. An attention mechanism gating network is used to concatenate the joint vector with the target video feature vector and multiply it by a second preset matrix vector to obtain a third weight value for each expert network under the video task type. The joint vector is then concatenated with the target promotional feature vector and multiplied by the second preset matrix vector to obtain a third weight value for each expert network under the promotional task type. The outputs of each expert network are weighted and concatenated according to the third weight value corresponding to each task type. The weighted concatenated outputs are then loaded into the first task training network according to the video task type, and a first output result is output. The user video fusion vector and video feature vector are input into the factorization machine model, which outputs a second output result. The first and second output results are added together to obtain the target output result corresponding to the video task type. The output after weighted connection according to the promotion task type is loaded into the second task training network, which outputs a third output result. The user promotion fusion vector and promotion feature vector are input into the factorization machine model, which outputs a fourth output result. The third and fourth output results are added together to obtain the target output result corresponding to the promotion task type. The target output result of each task type is compared with the corresponding label information to obtain the difference value. The network parameters of the task training network are adjusted according to the difference value until the difference value converges, thus obtaining the trained preset multi-task learning model. The display unit is used to display target video information and target promotional information whose correlation with the target video information is greater than a preset threshold. The target video information and target promotional information are obtained by pushing them to users through the trained preset multi-task learning model.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the information processing method according to any one of claims 1 to 5.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the information processing method according to any one of claims 1 to 5.

Images