Data processing method, model training method, apparatus, device, and storage medium

Abstract

Embodiments of the application disclose a data processing method, a model training method, a device, equipment and a storage medium, wherein the data processing method comprises: obtaining user feature data of a target object identifier; calling a first mapping module to perform vector generation processing on the user feature data to generate an identifier embedding vector of the target object identifier; the first mapping module is obtained based on a training multimedia data set having an association relationship with each training object identifier in a training object identifier set and training feature data associated with each training object identifier; performing feature mapping processing on the user feature data to generate a feature embedding vector of the target object identifier; determining an object embedding vector of the target object identifier according to the identifier embedding vector and the feature embedding vector; the identifier embedding vector represents multimedia preference features of the target object identifier, and the feature embedding vector represents user features associated with the target object identifier. Targeted recommendations can be made to improve the recommendation effect.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method, a model training method, an apparatus, a device, and a storage medium. Background Technology

[0002] With the rapid development of computer and multimedia technologies, various multimedia data platforms (such as short video platforms, news and information platforms, etc.) have sprung up. Every day, tens of thousands of new users use these multimedia data platforms. Therefore, data recommendation for new users has become a very important part of the recommendation system.

[0003] Currently, recommendation systems typically use user preference data, such as historical viewing data like frequently watched videos and news, to recommend various multimedia content. However, this approach is unsuitable for new users. When a new user joins a multimedia platform, the recommendation system can only make recommendations based on basic user information such as age, gender, and location, resulting in weak targeting and poor recommendation performance. Summary of the Invention

[0004] This application provides a data processing method, a model training method, an apparatus, a device, and a storage medium, which can make targeted recommendations to improve recommendation effectiveness.

[0005] On one hand, embodiments of this application disclose a data processing method, the method comprising:

[0006] Obtain user characteristic data associated with the target object identifier;

[0007] The first mapping module is invoked to perform vector generation processing on the user feature data to generate the identifier embedding vector of the target object identifier; the first mapping module is trained based on the training multimedia dataset that has a relationship with each training object identifier in the training object identifier set and the training feature data associated with each training object identifier;

[0008] The user feature data is subjected to feature mapping processing to generate the feature embedding vector of the target object identifier;

[0009] Based on the identifier embedding vector and the feature embedding vector, determine the object embedding vector of the target object identifier;

[0010] The identifier embedding vector is used to characterize the multimedia preference features of the target object identifier, and the feature embedding vector is used to characterize the user features associated with the target object identifier.

[0011] On the other hand, embodiments of this application disclose a model training method for user feature data, the method comprising:

[0012] Obtain a set of training object identifiers from the object identifiers registered on the multimedia platform, and obtain the training dataset corresponding to the set of training object identifiers. The training dataset includes: training feature data of the target training object identifier, training multimedia dataset of the target training object identifier, and multimedia data in which the user has applied positive feedback operations as instructed by the target training object identifier.

[0013] Based on the training multimedia dataset of the target training object identifier, obtain the actual preference feature vector of the target training object identifier;

[0014] The first initial mapping module performs vector generation processing on the training feature data of the target training object identifier to generate the training identifier embedding vector of the target object identifier.

[0015] The training feature data of the target training object identifier is processed by the second initial mapping module to generate the training feature embedding vector of the target object identifier.

[0016] The first initial mapping module and the second initial mapping module of the initial model are optimized and trained based on the actual preference feature vector, the training feature embedding vector, and the training identifier embedding vector to obtain the vector generation model.

[0017] On the other hand, embodiments of this application disclose a data processing apparatus, the apparatus comprising:

[0018] The acquisition unit is used to acquire user feature data associated with the target object identifier;

[0019] The first processing unit is used to call the first mapping module to perform vector generation processing on the user feature data to generate the identifier embedding vector of the target object identifier; the first mapping module is trained based on the training multimedia dataset that has a relationship with each training object identifier in the training object identifier set and the training feature data associated with each training object identifier.

[0020] The second processing unit is used to perform feature mapping processing on the user feature data to generate a feature embedding vector of the target object identifier.

[0021] The determining unit is configured to determine the object embedding vector of the target object identifier based on the identifier embedding vector and the feature embedding vector;

[0022] The identifier embedding vector is used to characterize the multimedia preference features of the target object identifier, and the feature embedding vector is used to characterize the user features associated with the target object identifier.

[0023] On the other hand, embodiments of this application disclose a model training apparatus for user feature data, the apparatus comprising:

[0024] The acquisition unit is used to acquire a set of training object identifiers from the object identifiers registered on the multimedia platform, and to acquire a training dataset corresponding to the set of training object identifiers. The training dataset includes: training feature data of the target training object identifier, and training multimedia dataset of the target training object identifier. The training multimedia dataset includes multimedia data in which the user has applied a positive feedback operation as instructed by the target training object identifier.

[0025] The acquisition unit is further configured to acquire the actual preference feature vector of the target training object identifier based on the training multimedia dataset of the target training object identifier;

[0026] The first generation unit is used to perform vector generation processing on the training feature data of the target training object identifier through the first initial mapping module to generate the training identifier embedding vector of the target object identifier.

[0027] The second generation unit is used to perform feature mapping processing on the training feature data of the target training object identifier through the second initial mapping module to generate the training feature embedding vector of the target object identifier.

[0028] The training unit is used to optimize and train the first initial mapping module and the second initial mapping module of the initial model based on the actual preference feature vector, the training feature embedding vector, and the training identifier embedding vector to obtain the vector generation model.

[0029] Accordingly, this application also discloses a smart device, including a processor, a memory, and a network interface, wherein the processor, memory, and network interface are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the above-described method.

[0030] Accordingly, embodiments of this application also disclose a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the above-described method.

[0031] Accordingly, this application also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the smart device reads the computer instructions from the computer-readable storage medium and executes the computer instructions to implement the above-described method.

[0032] In this embodiment, user feature data associated with the target object identifier can be obtained, and a first mapping module can be invoked to perform vector generation processing on the user feature data to generate an identifier embedding vector for the target object identifier. This identifier embedding vector can be used to characterize the multimedia preference features of the target object identifier. Feature mapping processing can also be performed on the user feature data to generate a feature embedding vector for the target object identifier. This feature embedding vector can be used to characterize the user features associated with the target object identifier. Furthermore, the object embedding vector of the target object identifier can be determined based on the identifier embedding vector and the feature embedding vector. By implementing the above method, an object embedding vector for the user can be obtained using the identifier embedding vector related to user preferences and the feature embedding vector related to user basic attribute data. When using this object embedding vector to make data recommendations to users, the user preference representation and feature representation can be fully utilized to achieve targeted recommendations for users, thereby improving the recommendation effect and accuracy. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic diagram of the architecture of a data processing system provided in an embodiment of this application;

[0035] Figure 2 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0036] Figure 3 This is a schematic diagram of the structure of a vector generation model provided in an embodiment of this application;

[0037] Figure 4 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0038] Figure 5a This is a schematic diagram of the training architecture of an initial model provided in an embodiment of this application;

[0039] Figure 5b This is a schematic diagram of a structure for obtaining actual preference feature vectors provided in an embodiment of this application;

[0040] Figure 6 This is a flowchart illustrating a model training method for user feature data provided in an embodiment of this application;

[0041] Figure 7 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0042] Figure 8 This is a schematic diagram of the structure of a model training device for user feature data provided in an embodiment of this application;

[0043] Figure 9 This is a schematic diagram of the structure of a smart device provided in an embodiment of this application;

[0044] Figure 10 This is a schematic diagram of the structure of a smart device provided in an embodiment of this application. Detailed Implementation

[0045] 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, 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.

[0046] 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.

[0047] 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.

[0048] Computer vision (CV) is a science that studies how to enable machines to "see." More specifically, it refers to machine vision, which uses cameras and computers to replace human eyes in recognizing, tracking, and measuring targets, and then performs image processing to create images more suitable for human observation or transmission to instruments. As a scientific discipline, computer vision studies related theories and technologies, attempting to build artificial intelligence systems capable of extracting information from images or multidimensional data. Computer vision technologies typically include image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, 3D object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping (SLAM), and common biometric recognition technologies such as facial recognition and fingerprint recognition.

[0049] 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 / deep learning typically includes techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0050] Research and analysis show that a user's representation can be divided into two parts: a user-identifier-specific embedding vector and a user-feature-specific feature embedding vector. When a new user enters the recommendation system, the system can assign them a new user identifier and a corresponding initialized embedding vector as the first part of the representation. For the second part, feature representation, since the features of the new user have almost all appeared in previous users, the feature embedding vectors of previous users can be directly referenced to generate the new user's feature representation. This application's embodiments focus on these two parts of features to process the feature representation of new users.

[0051] Based on the computer vision and machine learning technologies mentioned above in the artificial intelligence technologies, this application proposes a vector generation model (i.e., a model for generating object embedding vectors corresponding to object identifiers, where the object embedding vector can be a user representation of the object corresponding to the object identifier). This vector generation model can include a first mapping module and a second mapping module. The first mapping module can be used to generate the identifier embedding vector corresponding to the object identifier, and the second mapping module can be used to generate the feature embedding vector of the object identifier. The identifier embedding vector can be used to represent the multimedia preference features of the object identifier, and the feature embedding vector can be used to represent the user features associated with the object identifier. By performing user representation of the object identifier based on the above two types of embedding vectors, the final object embedding vector can more accurately represent an object, thereby improving the recommendation accuracy of multimedia data recommendations.

[0052] Furthermore, this application embodiment proposes a data processing scheme based on the vector generation model. Specifically, the general principle of this scheme is as follows: First, user feature data associated with the target object identifier is obtained. Then, the vector generation model is called to perform vector generation processing on the user feature data to generate an identifier embedding vector for the target object identifier. Next, the vector generation model is called to perform feature mapping processing on the user feature data to generate a feature embedding vector for the target object identifier. Then, the object embedding vector of the target object identifier can be determined based on the identifier embedding vector and the feature embedding vector. In other words, in this application embodiment, a vector generation model trained based on the relevant data of existing users can be used to generate identifier embedding vectors related to user preferences and feature embedding vectors related to user basic attribute data for new users or users with relatively little historical viewing data such as videos and news (referred to as new users in all embodiments of this application for ease of description). Thus, based on the identifier embedding vectors and feature embedding vectors, subsequent recommendation processing can be better performed for new users.

[0053] In specific implementations, the execution entity of the aforementioned data processing scheme can be a smart device, which can be a smart terminal or a server. The smart terminal mentioned here can be a smartphone, tablet, laptop, desktop computer, or other devices, or peripheral devices such as a gamepad or touchscreen; the server can be an independent 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, content delivery networks (CDNs), and big data and artificial intelligence platforms, etc. It should be noted that when the smart device is a server, this application embodiment provides a data processing system, such as... Figure 1 As shown, the data processing system includes at least one smart terminal and at least one server. The smart terminal can acquire user feature data associated with a target object identifier and upload the acquired user feature data associated with the target object identifier to the server (i.e., the smart device), so that the smart device can acquire the user feature data and determine the object embedding vector of the target object identifier based on the user feature data for subsequent data recommendation. At this time, the smart terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in this embodiment.

[0054] Practice has shown that the data processing scheme proposed in this application has the following beneficial effects: it allows the object embedding vector to represent two parts of an object identifier. The first part can be an identifier embedding vector specific to the object identifier, which contains the multimedia preference information of the object identifier; the second part can be a feature embedding vector specific to user features. By representing the object identifier based on the above two types of embedding vectors, the final object embedding vector can accurately represent an object. Therefore, when using the object embedding vector to make data recommendations for object identifiers, the preference representation part and the feature representation part of the object can also be fully applied. For example, in the user cold start recommendation scenario, that is, when a new user enters the recommendation system, an effective identifier embedding vector can be constructed for the new user, so that the identifier part of the user representation can work together with the feature part in the recommendation of new users, optimize the recommendation effect of new users, and effectively improve the recommendation accuracy.

[0055] The implementation details of the technical solutions in the embodiments of this application are described in detail below:

[0056] Please see Figure 2 , Figure 2This is a flowchart illustrating a data processing method provided in an embodiment of this application. The method described in this embodiment can be applied to smart devices, such as smart terminals, including smartphones, tablets, smart wearable devices, etc., and can also be servers. The data processing method described in this embodiment includes the following steps.

[0057] S201: Obtain user feature data associated with the target object identifier.

[0058] The target object identifier can be an identifier corresponding to a target object, which can refer to any user. In specific application scenarios, such as recommendation scenarios, the target object identifier can refer to a user who needs to make data recommendations on the recommendation platform. This user can refer to a new user or a returning user of the recommendation platform. A new user can refer to a user using the recommendation platform for the first time to make data recommendations, while a returning user can refer to a user who has used the recommendation platform for data recommendations before. User feature data can include the basic attributes of the target object, such as the target object's age, gender, place of residence, place of work, and industry.

[0059] S202: Call the first mapping module to perform vector generation processing on the user feature data and generate the identifier embedding vector of the target object identifier.

[0060] In one implementation, a vector generation model can be invoked to perform vector generation processing on user feature data to generate an identifier embedding vector for the target object. This vector generation model may include a first mapping module, for example... Figure 3 The illustration shows a vector generation model provided in an embodiment of this application. Specifically, the identifier embedding vector can be generated by calling the first mapping module in the vector generation model to perform vector generation processing on user feature data. In other embodiments, the first mapping module can also exist independently. For example, it can be constructed using a large amount of multimedia data such as video data and news data already viewed by long-time users, and supervised data can be determined based on the identified preference features of these long-time users. A network model can then be optimized and trained using a supervised training method, and the trained model can be used as the first mapping module.

[0061] The embedded vector can be used to characterize the multimedia preference features of the target object. These multimedia preference features refer to preferences for multimedia data, such as videos, music, news, novels, comics, etc. Optionally, for recommendation platforms that collect various types of multimedia data (e.g., a platform that can recommend videos, music, news, etc.), the multimedia data preference features can characterize the target object's degree of preference for different types of multimedia data; the target object's degree of preference for different types of multimedia data usually varies. Optionally, for recommendation platforms that collect only one type of multimedia data (e.g., a platform that can recommend videos, or a platform that can recommend news, etc.), the multimedia data preference features can also characterize the target object's degree of preference for various subcategories within that specific type of multimedia data. For example, taking videos as the multimedia data, the multimedia data preference features can characterize the target object's degree of preference for various video types within the video, where video types can include sports videos, entertainment videos, educational videos, legal videos, etc. For example, taking multimedia data as news, the preference features of multimedia data can characterize the target audience's preference for various news types in the news. These news types can include sports news, entertainment news, legal news, and so on.

[0062] For example, assuming that the degree of preference is represented by a value from 0 to 10, the degree of preference of a target object for any type of multimedia data can be any of these values, and different multimedia data can have different values ​​or the same value. Therefore, as can be seen from the above, the target object identifier can have different degrees of preference for various types of multimedia data or various subcategories included in a type of multimedia data. The identifier embedding vector mentioned in this application can then be used to represent the multimedia preference characteristics of the target object identifier.

[0063] In one implementation, the vector generation model, or the first mapping module, can be trained based on a training multimedia dataset associated with each training object identifier in the training object identifier set, and training feature data associated with each training object identifier. The training multimedia dataset can include one or more multimedia datasets. In this application, the corresponding training multimedia dataset can vary depending on the application scenario. For example, in a video recommendation scenario, the training multimedia dataset can be a dataset specifically for videos, meaning it can be a training video set with one or more training videos. Similarly, in a news recommendation scenario, the training multimedia dataset can be a dataset specifically for news, meaning it can be a training news set with one or more training news items. The association can refer to a positive feedback operation applied to the multimedia data by the training object identifier, such as watching, liking, or commenting. For example, taking videos as the multimedia data, for any training object identifier, the training multimedia dataset corresponding to that training object identifier can include: the set of videos watched by the training object identifier. For example, taking multimedia data such as news as an example, for any training object identifier, the corresponding training multimedia dataset can include: the news collections viewed by the training object corresponding to that training object identifier. Understanding training feature data can be referenced in the above explanation of user feature data.

[0064] S203: Perform feature mapping processing on user feature data to generate feature embedding vectors for the target object identifier.

[0065] The feature embedding vector is used to represent the user characteristics managed by the target object identifier. It can be represented by the basic user attributes of the target object identifier, such as age (25), place of residence (Beijing), place of work (Beijing), occupation (lawyer), etc. These user characteristics can be represented by a single feature embedding vector. Generally, if any two object identifiers manage different user characteristics, then the feature embedding vectors corresponding to the two object identifiers will also be different.

[0066] In one implementation, the feature embedding vector of the target object identifier can be generated through a vector generation model. That is, the user feature data can be processed using a vector generation model to perform feature mapping to generate the feature embedding vector of the target object identifier. Alternatively, such as... Figure 3 As shown, the vector generation model may include a second mapping module, which can be used to perform feature mapping processing on user feature data to generate feature embedding vectors for the target object identifier.

[0067] S204: Determine the object embedding vector of the target object identifier based on the identifier embedding vector and the feature embedding vector.

[0068] In one implementation, the identifier embedding vector and the feature embedding vector can be concatenated, and the result of the concatenation can be used as the object embedding vector of the target object identifier. Concatenation can be understood as joining the two vectors together; for example, the identifier embedding vector can be concatenated after the feature embedding vector to obtain the object embedding vector, or vice versa.

[0069] In one implementation, such as Figure 3 As shown, the vector generation model can also include a fusion module. This fusion module can then be used to fuse the identifier embedding vector and the feature embedding vector to obtain the object embedding vector of the target object identifier. The fusion process corresponding to the fusion module can be characterized using the following formula 1:

[0070]

[0071] Among them, u i′ The object embedding vector represents the identifier of the target object; ReLU is a non-linear transformation function; W u and b u These are the model parameters corresponding to the fusion module; Represents a concatenated vector sum vector X represents the feature embedding vector that identifies the target object. i′ User characteristic data representing the target object identifier, f c (·) indicates the mapping function of the second mapping module; f represents the identifier embedding vector of the target training object identifier. p (·) indicates the mapping function of the first mapping module.

[0072] Understandably, after the initial model training is completed to obtain the vector generation model, the model parameters in this vector generation model, namely the model parameters in the first initial mapping module, the model parameters in the second initial mapping module, and the model parameters in the fusion module (W) u and b u This can be solidified to process user feature data of other object identifiers using the vector generation model, and generate corresponding object embedding vectors. These object embedding vectors can be user representations of object identifiers, and are typically applied in multimedia data recommendation scenarios to perform multimedia data recommendations based on object identifiers.

[0073] For example, in a cold-start user recommendation scenario (which can be understood as recommending users or content lacking interactive behavior, i.e., newly added users or content), for a new user (such as the target object corresponding to the target object identifier), the user feature data of the new user can be mapped to the feature part representation of the new user through the second mapping module, i.e., the feature embedding vector; similarly, the user feature data of the new user can be mapped to the identifier (ID) part representation of the new user through the first mapping module, i.e., the identifier embedding vector. With these two representations, the final representation of the new user, i.e., the object embedding vector, can be obtained through the fusion transformation corresponding to the above formula (1).

[0074] In one implementation, the object embedding vector of the target object identifier obtained by the above method can also be specifically used in multimedia data recommendation scenarios. Specifically, firstly, the embedding vectors to be recommended for each multimedia data in the multimedia dataset to be recommended can be obtained. These embedding vectors can be pre-determined (the determination of these embedding vectors can refer to the method for determining multimedia vectors for multimedia data described below; for example, an embedding network can be used to extract features from the multimedia data to obtain the corresponding embedding vectors), and stored in a designated location for later retrieval. For example, taking video as the multimedia data, videos typically have corresponding video identifiers. The video identifier can be used to index the corresponding embedding vector at a designated location, and the obtained embedding vector can then be used as the embedding vector to be recommended for that video.

[0075] After obtaining the embedding vectors to be recommended, the recommendation score for each multimedia data point can be determined based on the object embedding vector and the embedding vectors of each multimedia data point to be recommended. Optionally, the embedding vectors of each multimedia data point to be recommended can be used to calculate the inner product with the object embedding vector, and the result of the inner product calculation can be used as the recommendation score for a multimedia data point to be recommended. For example, the recommendation score for any multimedia data point to be recommended can be calculated using the following formula 2:

[0076] r = (u i′ ) T e j Formula 2

[0077] Where r represents the recommendation score, u i′ Represents the object embedding vector, e j This represents the embedding vector of the multimedia data to be recommended, (·). T This indicates the transpose operation.

[0078] Therefore, after obtaining the recommendation scores for each multimedia data to be recommended using Formula 2 above, multimedia data can be recommended based on these scores. For example, target multimedia data can be determined from the multimedia data to be recommended based on their recommendation scores. Multiple multimedia data to be recommended can be sorted sequentially from highest to lowest recommendation score to obtain the corresponding sorting results. Then, the multimedia data in the top L positions of the sorting results is taken as the target multimedia data. Here, L can be a positive integer, such as 1, 5, 7, etc., and its specific value is not limited in this application. After determining the target multimedia data, multimedia data is recommended to the target object identifier based on the target multimedia data. Optionally, when there are multiple target multimedia data, if the target multimedia data can be displayed on the terminal screen used by the target object identifier, then when recommending multimedia data to the target object identifier based on the target multimedia data, the data can be displayed sequentially on the terminal screen from highest to lowest recommendation score to improve the user experience.

[0079] For example, taking videos as the multimedia data to be recommended, and L = 5, in a video recommendation scenario, firstly, the set of videos to be recommended in this scenario can be obtained. This set of videos can include multiple videos to be recommended. Then, the embedding vector corresponding to each video to be recommended is determined. Next, the inner product between the embedding vector to be recommended and the object embedding vector corresponding to each video to be recommended is calculated using Formula 2 above to obtain the recommendation score of each video to be recommended. Furthermore, the multiple videos to be recommended are sorted in descending order of recommendation score to obtain the corresponding video ranking result. Then, the top 5 videos to be recommended in the video ranking result can be used as target videos so that subsequent video recommendations can be made based on the target object identifier.

[0080] In this embodiment, user feature data associated with the target object identifier can be obtained, and a first mapping module can be invoked to perform vector generation processing on the user feature data to generate an identifier embedding vector for the target object identifier. This identifier embedding vector can be used to characterize the multimedia preference features of the target object identifier. Feature mapping processing can also be performed on the user feature data to generate a feature embedding vector for the target object identifier. This feature embedding vector can be used to characterize the user features associated with the target object identifier. Further, the object embedding vector of the target object identifier can be determined based on the identifier embedding vector and the feature embedding vector. By implementing the above method, the object embedding vector can characterize two parts of the target object identifier: the first part can be an identifier embedding vector specific to the object identifier, which contains the target object's preference information, or a representation of the target object's preferences; the second part can be a feature embedding vector specific to user features. Therefore, when using the object embedding vector to make data recommendations for the target object identifier, the preference representation part and feature representation part of the target object can be fully applied. For example, in the user cold start recommendation scenario, that is, when a new user enters the recommendation system, an effective identifier embedding vector can be constructed for the new user, so that the identifier part of the user representation can work together with the feature part in the recommendation of the new user, optimize the recommendation effect of the new user, and also effectively improve the recommendation accuracy.

[0081] Please see Figure 4 , Figure 4 This is a flowchart illustrating a data processing method provided in an embodiment of this application. The method described in this embodiment can be applied to smart devices, such as smart terminals, including smartphones, tablets, smart wearable devices, etc., and can also be servers. The data processing method described in this embodiment includes the following steps.

[0082] S401: Obtain the set of training object identifiers from the object identifiers registered on the multimedia platform, and obtain the training dataset corresponding to the set of training object identifiers.

[0083] The multimedia platform can be a platform used by users to share and browse multimedia data. For example, it could be a video platform, a news and information platform, or a multimedia platform that integrates multiple different multimedia data types. Typically, users can register on these multimedia platforms to become users of those platforms. Therefore, when obtaining the training dataset, the set of training object identifiers can be obtained from the object identifiers registered on the multimedia platform, and the corresponding training dataset can be retrieved.

[0084] Understandably, for a multimedia platform, the number of registered object identifiers may be very large. Therefore, it is advisable to extract a portion of the object identifiers from all the object identifiers to use as a training object identifier set. This training object identifier set may include one or more training object identifiers. The number of training object identifiers in the training object identifier set can be preset, for example, it can be set to 300, 500, etc., and the specific value is not limited in this application.

[0085] Optionally, when determining the training object identifier set, a preset number of object identifiers can be randomly selected from all object identifiers as the training object identifier set. Alternatively, a preset number of object identifiers can be selected from all object identifiers according to a preset selection rule. This preset selection rule could be a selection rule based on registration time, historical pageviews, etc., where registration time refers to the time corresponding to the object identifier's registration on the multimedia platform, and historical pageviews refer to the number of times the object identifier viewed multimedia data on the multimedia platform within a certain period before the current time (which could be the time corresponding to the acquisition of the training object identifier set). The following examples illustrate this further.

[0086] For example, the preset selection rule could be based on registration time. Specifically, the registration time of each object identifier can be obtained first, and then the object identifiers can be sorted sequentially from earliest to latest registration time to obtain the sorted object identifier result. After obtaining this sorted result, the first H object identifiers in the sorted result can be used as the training object identifier set, where H is the preset number described above. The early and late registration times are defined based on the current time. A larger time interval between a registration time and the current time indicates an earlier registration time, while a smaller time interval indicates a later registration time. For example, if the current time is 2020.10.10, the registration time of object identifier 1 is 2020.7.11, and the registration time of object identifier 2 is 2020.9.11, then when sorting these two object identifiers, the sorted result could be (object identifier 1, object identifier 2), meaning the object identifier with the earlier registration time should be ranked before the one with the later registration time. By using the methods described above, we can select object identifiers with earlier registration times to improve model efficiency.

[0087] For example, the preset selection rule could be based on historical pageviews. In practice, this could involve first obtaining the historical pageviews of each object identifier, then sorting the object identifiers sequentially from highest to lowest historical pageview count. After obtaining this sorting result, the top H object identifiers in the sorted result can be used as the training set. This method allows for the selection of object identifiers with the highest possible pageviews, thereby improving the model's efficiency.

[0088] In one implementation, the training dataset may include the training dataset corresponding to each training object identifier in the training object identifier set. Taking a target training object identifier (which can refer to any training object identifier in the training object identifier set) as an example, the training dataset may include: training feature data of the target training object identifier and training multimedia dataset of the target training object identifier. The training multimedia dataset may include multimedia data for which a positive feedback operation has been performed by the user at the instruction of the target training object identifier. This positive feedback operation may refer to actions such as watching, liking, or commenting. For example, taking a video as an example, the positive feedback operation may refer to liking, commenting, or watching a video.

[0089] For example, taking a short video recommendation platform as an example, we define the existing user set (which can also be understood as the training object identifier set mentioned above) U = {u1, u2, ..., u...} M} and the short video set V = {v1, v2, ..., v N} where M is the total number of existing users (i.e., the number of training object identifiers), and N is the total number of short videos. The set of short videos here can be understood as the short videos available on the short video recommendation platform. Define a user-short video interaction matrix. When R ij When R = 1, it means that user i (i∈U) has applied a positive feedback operation to short video j (j∈V), and when R ij When = 0, it indicates that user i has performed a negative feedback operation on short video j (a negative feedback operation can be understood as not performing operations such as watching, liking, or commenting). Here, we can let U(j) = {i∈U|R ij =1} represents the set of users who applied a positive feedback operation to short video j. Similarly, V(i) = {j∈V|R ij=1} can represent the set of short videos to which user i has applied positive feedback. Considering that this application mainly uses the target training object identifier in the training object identifier set and the training multimedia dataset corresponding to the target training object identifier as examples to illustrate the initial model training, then i in the following description can be used to represent the target training object identifier, and the corresponding V(i) can be used to represent the training multimedia dataset corresponding to the target training object identifier.

[0090] As described above, in the scenario of cold start user recommendation, for new users (such as the target object identifier mentioned above)... In other words, the new user i′ has not applied any positive feedback operation to any short videos on the short video recommendation platform; that is, the set of short videos to which the new user i′ has applied positive feedback operations is empty. The goal of cold-start user recommendation is to provide personalized recommendations to the new user i′ based on their user feature data. This involves processing the user feature data of the target object identifier using a vector generation model, as described in steps S201-S204 above, to obtain the corresponding object embedding vector, and then using this object embedding vector for multimedia data recommendation. The following mainly describes the process of training the initial model to obtain the vector generation model.

[0091] It is understood that in the specific implementation of this application, user information (such as the user characteristic data, registration time, historical browsing volume, etc. mentioned above) and other related data are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0092] S402: Train the initial model based on the training dataset to obtain a vector generation model, which includes a first mapping module and a second mapping module.

[0093] In one implementation, the initial model may include: a first initial mapping module and a second initial mapping module, where the first initial mapping module corresponds to the first mapping module, and the second initial mapping module corresponds to the second mapping module. That is, it can be understood that optimizing and training the first initial mapping module yields the first mapping module, and similarly, optimizing and training the second initial mapping module yields the second mapping module.

[0094] In one implementation, the training process of the initial model can be described as follows: Based on the training multimedia dataset of the target training object identifier, the actual preference feature vector of the target training object identifier is obtained. Then, the first initial mapping module of the initial model performs vector generation processing on the training feature data of the target training object identifier to generate the training identifier embedding vector of the target object identifier; and the second initial mapping module of the initial model performs feature mapping processing on the training feature data of the target training object identifier to generate the training feature embedding vector of the target object identifier. After obtaining the actual preference feature vector, the training feature embedding vector, and the training identifier embedding vector, the first and second initial mapping modules of the initial model can be optimized and trained based on these vectors to obtain the vector generation model.

[0095] In one implementation, the initial model may further include a fusion initial module, in which the aforementioned training fusion feature may be obtained by fusing the training identifier embedding vector and the actual preference feature vector through the fusion initial module.

[0096] In one implementation, the training architecture for the initial model is as follows: Figure 5a As shown, rectangle 51 refers to the second initial mapping module, rectangle 52 refers to the first initial mapping module, rectangle 53 refers to the fusion initial module, rectangle 54 refers to the capsule network, and rectangle 55 refers to the attention structure network; square 501 refers to the training feature data of the target training object identifier, solid triangle 502 refers to the training feature embedding vector, hollow pentagram 503 refers to the training identifier embedding vector, solid pentagram 504 refers to the actual preference feature vector, human figure 505 refers to the training fusion feature, hollow circle 506 refers to the multimedia feature vector corresponding to the multimedia reference data, solid circle 507 refers to the multimedia vector set obtained from the training multimedia dataset of the target training object identifier, and solid ellipse 508 refers to the initial preference feature vector.

[0097] The initial model may include, for example: Figure 5a The first initial mapping module, the second initial mapping module, and the fusion initial module are shown. Figure 5a As shown, the training architecture for the initial model can include not only the initial model but also a branch module, which includes a capsule network and an attention structure network to assist in the training of the initial model.

[0098] The capsule network and attention structure network in the branch module can be used to obtain the actual preference feature vector of the target training object identifier based on the training multimedia dataset of the target training object identifier. Optionally, this branch module can exist independently. In this case, when using this branch module to obtain the actual preference feature vector of the target training object identifier, it can be done by processing the training multimedia dataset of the target training object identifier using the pre-trained branch module to obtain the actual preference feature vector of the target training object identifier; or it can be jointly trained with other modules in the training architecture of the initial model to obtain the actual preference feature vector of the target training object identifier. The following mainly uses the example of jointly training this branch module with other modules in the training architecture of the initial model to illustrate the above.

[0099] In one implementation, when optimizing the initial model based on the actual preference feature vector, the training feature embedding vector, and the training label embedding vector, a joint training approach can be used to improve the model's efficiency. Specifically, two loss functions can be used to calculate the loss value, which can then be used for optimized model training. These two loss functions can include a first loss function and a second loss function.

[0100] (a) For the first loss function:

[0101] A first loss function can be used to simulate the loss between the training identifier embedding vector and the actual preference feature vector. This first loss function can be the distance loss between the training identifier embedding vector and the actual preference feature vector, which can be described as the distance between them, using the mean squared error as a metric. For example, the functional expression of this first loss function can be shown in Equation 3 below:

[0102]

[0103] in, The actual preference feature vector representing the target training object identifier (the acquisition of this data can be found in the relevant description below); f p (X i ) represents the training identifier embedding vector, f p (·) indicates the mapping effect of the first initial mapping module, X i Training feature data representing the identifier of the target training object.

[0104] The simulated loss part can guide the learning of parameters in the mapping from the training feature data of the target training object identifier to the training identifier embedding vector, so that the initial model can extract the preference information of the target training object identifier from the training feature data of the target training object identifier. That is, it can train the model parameters in the first initial mapping module to obtain the optimized first mapping module.

[0105] (ii) For the second loss function:

[0106] When calculating the second loss function, the training feature embedding vector and the actual preference feature vector can be fused first to obtain the training fused feature. As mentioned above, the training fused feature can be obtained by fusing the training feature embedding vector and the actual preference feature vector through the fusion initialization module in the initialization module. For example, this training fused feature can be represented by the following formula 4:

[0107]

[0108] Where i represents the i-th training object identifier in the training object identifier set, such as i representing the target training object identifier mentioned above. Assuming that the embodiments of this application are mainly illustrated with i as the target training object identifier; and These are the model parameters corresponding to the initial fusion module in the fusion transformation; Represents a concatenated vector sum vector This represents the training feature embedding vector representing the identifier of the target training object. f c (·) indicates the mapping function of the second initial mapping module, i.e. This can also be referred to as the mapping result from training feature data to training feature embedding vectors; and These are the two main components of the user representation of the target training object identifier.

[0109] After fusion processing, the final representation of the target training object identifier is obtained, namely the aforementioned training fusion feature u. i Then, the second loss function can be used to calculate the prediction loss for the multimedia feature vectors corresponding to the training fusion features and multimedia reference data. Specifically, the preference score of the target training object identifier i for the multimedia reference data can be calculated first. For example, the calculation of this preference score can be shown in Formula 5 below:

[0110]

[0111] Among them, e tV(i) represents the multimedia feature vector corresponding to the multimedia reference data; V(i) represents the training multimedia dataset corresponding to the target training object identifier; e j This represents the multimedia vector of the j-th multimedia data in the training multimedia dataset corresponding to the target training object identifier (the acquisition of this data can be found in the relevant description below).

[0112] Optionally, multimedia reference data can refer to the most recently pushed multimedia data to the target training object identifier, and the target training object identifier has provided positive feedback to the multimedia reference data. Based on this, a specific implementation method for determining multimedia reference data can include: First, a set of multimedia reference data for the target training object identifier recorded within a preset time range can be determined. The multimedia reference data set can include one or more multimedia data recommended to the target training object identifier and marked as having received positive feedback from the target training object identifier. The preset time range can be pre-set; for example, it can be a period of time before the time corresponding to the determination of the multimedia reference data set. For instance, if the time corresponding to the multimedia reference data set is determined to be January 5, 2020, then the preset time range can be January 3, 2020 – January 4, 2020. Then, after determining the multimedia reference data set, multimedia reference data can be determined from it. For example, a multimedia reference data set can be randomly selected from the multimedia reference data set as the multimedia reference data required for this application. Alternatively, when determining the multimedia reference data set, the recommendation time corresponding to each multimedia reference data set can also be determined (the time when the multimedia reference data is recommended to the target training object identifier). In this case, the multimedia reference data set with the latest recommendation time can be selected as the multimedia reference data required for this application. Selecting the multimedia reference data with the latest recommendation time also allows learning the recommendation effect of the most recent multimedia data, ensuring that the vector generation model trained subsequently has better model performance, and that the object embedding vectors generated by the vector generation model are more closely aligned with the user representation of the object.

[0113] After obtaining the preference score of the target training object identifier i for the multimedia reference data using Formula 5 above, the second loss function can be used to calculate the prediction loss based on this preference score. The second loss function can be the cross-entropy loss for recommending multimedia data, that is, the cross-entropy (Softmax) loss is calculated during training. The functional expression of this second loss function can be shown in Formula 6 below:

[0114] L s =-logP(e t |u i )Formula 6

[0115] The prediction loss can be used to train the base mapping, which can include the model parameters of the user multi-interest extraction part. These model parameters can include the model parameters for feature extraction of each multimedia data in the training multimedia dataset that identifies the target training object, as well as the model parameters in the capsule network. They can also include the parameters in the mapping of user feature data to feature embedding vectors (i.e., the parameters in the first initial mapping module), and the model parameters in the vector fusion of the training feature embedding vector and the actual preference feature vector.

[0116] In summary, after obtaining the simulated loss and predicted loss calculations, the first and second initial mapping modules of the initial model can be further optimized and trained based on these calculations. As described above, when training the initial model, a joint training approach can be used for optimization, combining the predicted loss and simulated loss as a joint loss, which can be L = L... s +L m The initial model is trained using L to obtain a vector generation model. This vector generation model can provide end-to-end recommendations for new user data without requiring an additional cold start module.

[0117] In one implementation, the following mainly focuses on obtaining the actual preference feature vector of the target training object identifier. For example, Figure 5b The diagram shown illustrates the structure for obtaining the actual preference feature vector of the target training object identifier. For a detailed implementation, please refer to [link / reference needed]. Figure 5b First, features can be extracted from each multimedia data point in the training multimedia dataset that identifies the target training object, resulting in a multimedia vector set. This feature extraction is the first step in user interest extraction; it maps the multimedia data in the training multimedia dataset into low-dimensional embedding vectors. This embedding vector is the multimedia vector mentioned above. Here, d is the dimension of the embedding vector (multimedia vector); for example, in practice, d can typically be 64. Then, by mapping each multimedia data point in the training multimedia dataset, we can obtain a multimedia vector set, which can be represented as E. i ={e j ,j∈V(i)}, where V(i) represents the training multimedia dataset corresponding to the target training object identifier; e j This represents the multimedia vector of the j-th multimedia data in the training multimedia dataset corresponding to the target training object identifier.

[0118] After obtaining the multimedia vector set, see [link / reference]. Figure 5bThe multimedia vector set can then be further processed using a capsule network to obtain the actual preference feature vector of the target training object identifier. A capsule network is a novel basic model of neural networks that uses vectors as neurons. In this application, the capsule network is the second step in user interest extraction and a key component of it. Using the capsule network, preference representations of the target training object identifier can be extracted from the multimedia vector set. The dynamic routing mechanism within the capsule network helps capture various preferences of the target training object identifier. Specifically, the capsule network can use the multimedia vector set as behavioral capsule features, and based on these behavioral capsule features and the initialized preference capsule features, it performs feature processing using the dynamic routing mechanism of the capsule network to obtain the actual preference feature vector of the target training object identifier.

[0119] The following section uses relevant formulas to explain how to obtain the actual preference feature vector of the target training object using capsule networks. The process may include the following steps S1-S4.

[0120] S1: First, the multimedia vector set can be used as the behavioral capsule feature. Based on this behavioral capsule feature and the initialized preference capsule feature, the routing value between the behavioral capsule feature and the initialized preference capsule feature is calculated using the dynamic routing mechanism of the capsule network. For example, the multimedia vector set is E i Assume that the initial preference capsule features can be represented as Here, K represents the total number of preferences. K can be a pre-set parameter; for example, practice shows that K can be 9. K may also need to be adjusted based on the model's performance. For the k-th preference of the target training object identifier (i.e., the k-th initialized preference capsule feature) and the j-th multimedia data in the training multimedia dataset, the routing value between the two can be calculated using the following formula 7:

[0121]

[0122] in, Let S represent the bilinear mapping matrix. S can be model parameters in the capsule network, and this matrix can be shared throughout the calculation of all route values. This can be understood as the degree of fit between the j-th multimedia data and the k-th preference, with respect to the target training object identifier i.

[0123] S2: After obtaining all the routing values ​​using Formula 7 above, the pairing coefficient can be calculated based on these values. Pairing Coefficient It can be calculated right The cross-entropy (Softmax) is obtained by normalizing the output, that is, by using the Softmax function to normalize the routing values. The result of performing a normalization operation is the pairing coefficient. For example, pairing coefficient It can be calculated using the function expression shown in Formula 8 below:

[0124]

[0125] In this case, the sum of the pairing coefficients calculated from all the preference capsule features of the j-th multimedia data and the target training object identifier i is 1, i.e.

[0126] S3: After obtaining the pairing coefficients, these coefficients can be used as weights for the behavioral capsule features, and a weighted calculation can be performed to determine the candidate preference capsule feature set. This candidate preference capsule feature set can include K candidate preference capsule features. For example, for the kth candidate preference capsule feature... The weighted sum of all behavioral capsules can be obtained, and the corresponding calculation formula can be expressed as the function expression shown in Formula 9 below:

[0127]

[0128] S4: Further, the actual preference feature vector of the target training object identifier can be determined based on the candidate preference capsule feature set. Specifically, determining the actual preference feature vector based on the candidate preference capsule feature set can include the following steps S41-S42.

[0129] S41: The initial preference feature vector can be determined by first calculating the feature set of the candidate preference capsules using a nonlinear squeezing function. For example, the nonlinear squeezing function can be a squash function, as shown in Equation 10 below. Therefore, the initial preference feature vector can be determined using Equation 10.

[0130]

[0131] Here, squash() represents the action of the squash function, ||·|| 2 This represents the L2 norm.

[0132] S42: Considering that each capsule in a capsule network can represent a certain type of user preference, and that each type of preference has different importance in describing the overall user preference, an attention structure can be used to aggregate various user preferences to obtain a representation of the overall user preference. In this application, an attention structure based on a Multilayer Perceptron (MLP) is mainly used to aggregate various user preferences. That is, the candidate preference capsule feature set obtained above can be aggregated using an attention structure network based on a Multilayer Perceptron to obtain the actual preference feature vector of the target training object identifier. The MLP-based attention structure can be represented by the following formula 11:

[0133]

[0134] in, Indicates attention weight; d represents the model parameters in the attention structure network. a S represents the intermediate layer dimension of the attention structure network, and ReLU is a nonlinear transformation function. In this application, the model parameters S mentioned above and the model parameters in the attention structure network here can both be referred to as the model parameters for the capsule network.

[0135] As can be seen from Equation 11, the attention weight... It can be calculated right The cross-entropy (Softmax) is obtained by normalizing the output, that is, by using the Softmax function to... Performing a normalization operation yields the final attention weights required. After obtaining the attention weights corresponding to the K candidate preference capsule features using Formula 11, the weighted sum of all preference capsules for the target training object identifier i can be calculated to obtain the overall preference of the target training object identifier i. That is, the attention weights can be used to perform a weighted summation of the K candidate preference capsule features, and the result of this weighted summation is the actual preference feature vector of the target training object identifier i. For example, the actual preference feature vector can be calculated using the function expression shown in Formula 12:

[0136]

[0137] It should be noted that in formulas 7, 10, and 11... The corresponding representations are somewhat different, in Formula 11 and Formula 12. It's the same one. In Formula 7... It is obtained from random initialization and represents the initial preference capsule characteristics, as shown in Formula 10. This represents the initial preference feature vector, as shown in Formula 10. It is obtained by iterating the process of Formula 7-Formula 10 a certain number of times (such as 1-3 times).

[0138] As described above, the input to the initial model can include the training feature data X of the target training object identifier i. i The training multimedia dataset (e.g., a set of short videos with positive feedback applied) V(i) and multimedia reference data (e.g., reference short videos) t are given. The output may include the preference score of the target training object identifier i for applying positive feedback to the multimedia reference data t. The training feature data X... i It can be used to extract feature information of the target training object identifier i. The training multimedia dataset V(i) with positive feedback applied can be used to extract preference information of the target training object identifier i.

[0139] S403: Obtain user feature data associated with the target object identifier.

[0140] S404: Call the first mapping module to perform vector generation processing on the user feature data and generate the identifier embedding vector of the target object identifier.

[0141] S405: Perform feature mapping processing on user feature data to generate feature embedding vectors for the target object identifier.

[0142] S406: Determine the object embedding vector of the target object identifier based on the identifier embedding vector and the feature embedding vector.

[0143] The specific implementation of steps S403-S406 can be found in the detailed description of steps S201-S204 in the above embodiment, and will not be repeated here.

[0144] In this embodiment, a training object identifier set can be obtained from the object identifiers registered on the multimedia platform, and a training dataset corresponding to the training object identifier set can be obtained. The initial model can then be trained using the training dataset to obtain a vector generation model. Furthermore, this vector generation model can be used to generate object embedding vectors. For example, user feature data associated with the target object identifier can be obtained, and the first mapping module can be called to perform vector generation processing on the user feature data to generate an identifier embedding vector for the target object identifier. This identifier embedding vector can be used to represent the multimedia preference features of the target object identifier. Feature mapping processing can also be performed on the user feature data to generate a feature embedding vector for the target object identifier. This feature embedding vector can be used to represent the user features associated with the target object identifier. Further, the object embedding vector of the target object identifier can be determined based on the identifier embedding vector and the feature embedding vector. As described above, the initial model can be trained based on a capsule network, effectively extracting information about user preferences and providing mappings from the original user identifiers and features to identifier embedding vectors and feature embedding vectors. The initial model is trained in conjunction with a joint training module, co-training the basic mapping and cold-start mapping, integrating the cold-start mapping into the training stage of the basic mapping, thus improving model efficiency and performance.

[0145] Please see Figure 6 , Figure 6 This is a flowchart illustrating a model training method for user feature data provided in an embodiment of this application. The method described in this embodiment can be applied to smart devices, such as smart terminals like smartphones, tablets, smart wearable devices, etc., and can also be a server. The model training method for user feature data described in this embodiment includes the following steps.

[0146] S601: Obtain the set of training object identifiers from the object identifiers registered on the multimedia platform, and obtain the training dataset corresponding to the set of training object identifiers.

[0147] The training dataset may include: training feature data of the target training object identifier, and training multimedia dataset of the target training object identifier, wherein the training multimedia dataset includes multimedia data in which the user has applied positive feedback operations as instructed by the target training object identifier.

[0148] S602: Based on the training multimedia dataset of the target training object identifier, obtain the actual preference feature vector of the target training object identifier.

[0149] S603: The first initial mapping module performs vector generation processing on the training feature data of the target training object identifier to generate the training identifier embedding vector of the target object identifier.

[0150] S604: The second initial mapping module performs feature mapping processing on the training feature data of the target training object identifier to generate the training feature embedding vector of the target object identifier.

[0151] S605: Optimize and train the first initial mapping module and the second initial mapping module of the initial model based on the actual preference feature vector, the training feature embedding vector, and the training identifier embedding vector to obtain the vector generation model.

[0152] In one implementation, a first loss function can be used to simulate the loss calculation on the training identifier embedding vector and the actual preference feature vector. Then, the training feature embedding vector and the actual preference feature vector are fused to obtain the training fused feature. A second loss function is then used to calculate the prediction loss on the training fused feature and the multimedia feature vector corresponding to the multimedia reference data. Furthermore, based on the simulated loss calculation and the prediction loss calculation, the first initial mapping module and the second initial mapping module included in the initial model can be optimized and trained.

[0153] In one implementation, the initial model may further include a fusion initial module, in which the training identifier embedding vector and the actual preference feature vector can be obtained by fusing them through the fusion initial module. Based on this, the fusion initial module included in the initial model can be optimized and trained according to the simulated loss calculation and the predicted loss calculation.

[0154] It should be noted that the training structure of the initial model described in the embodiments of this application can be as follows: Figure 5a As shown, the specific training process of the initial model can be referred to the relevant description in the above embodiments, and will not be repeated here.

[0155] In this embodiment, the initial model can be trained based on capsule networks to effectively extract information about user preferences, providing a mapping from the original user identifiers and features to identifier embedding vectors and feature embedding vectors; and the training of the initial model can be achieved by using joint training, which can also improve model efficiency and model performance.

[0156] Please see Figure 7 This is a schematic diagram of the structure of a data processing apparatus provided in an embodiment of this application. The apparatus described in this application embodiment can be applied in smart devices, such as smart terminals, including smartphones, tablets, smart wearable devices, etc., and can also be servers.

[0157] The data processing apparatus described in this embodiment includes:

[0158] The acquisition unit 701 is used to acquire user feature data associated with the target object identifier;

[0159] The first processing unit 702 is used to call the first mapping module to perform vector generation processing on the user feature data to generate the identifier embedding vector of the target object identifier; the first mapping module is trained based on the training multimedia dataset that has a relationship with each training object identifier in the training object identifier set and the training feature data associated with each training object identifier.

[0160] The second processing unit 703 is used to perform feature mapping processing on the user feature data to generate the feature embedding vector of the target object identifier.

[0161] The determining unit 704 is configured to determine the object embedding vector of the target object identifier based on the identifier embedding vector and the feature embedding vector;

[0162] The identifier embedding vector is used to characterize the multimedia preference features of the target object identifier, and the feature embedding vector is used to characterize the user features associated with the target object identifier.

[0163] In one implementation, the identifier embedding vector and the feature embedding vector are generated by a vector generation model, and the device further includes a training unit 705, specifically used for:

[0164] Obtain a set of training object identifiers from the object identifiers registered on the multimedia platform, and obtain the training dataset corresponding to the set of training object identifiers. The training dataset includes: training feature data of the target training object identifier, training multimedia dataset of the target training object identifier, and multimedia data in which the user has applied positive feedback operations as instructed by the target training object identifier.

[0165] The initial model is trained based on the training dataset to obtain a vector generation model, which includes a first mapping module and a second mapping module.

[0166] In one implementation, the initial model includes: a first initial mapping module and a second initial mapping module, wherein the first initial mapping module corresponds to the first mapping module, and the second initial mapping module corresponds to the second mapping module; the training unit 705 is specifically used for:

[0167] Based on the training multimedia dataset of the target training object identifier, obtain the actual preference feature vector of the target training object identifier;

[0168] The first initial mapping module performs vector generation processing on the training feature data of the target training object identifier to generate the training identifier embedding vector of the target object identifier.

[0169] The training feature data of the target training object identifier is processed by the second initial mapping module to generate the training feature embedding vector of the target object identifier.

[0170] The first initial mapping module and the second initial mapping module of the initial model are optimized and trained based on the actual preference feature vector, the training feature embedding vector, and the training identifier embedding vector to obtain the vector generation model.

[0171] In one implementation, the training unit 705 is specifically used for:

[0172] Feature extraction is performed on each multimedia data in the training multimedia dataset of the target training object identifier to obtain a multimedia vector set;

[0173] The multimedia vector set is processed by a capsule network to obtain the actual preference feature vector of the target training object identifier;

[0174] The capsule network uses the multimedia vector set as behavioral capsule features, and performs feature processing based on the dynamic routing mechanism of the capsule network according to the behavioral capsule features and the initialized preference capsule features to obtain the actual preference feature vector of the target training object identifier.

[0175] In one implementation, the training unit 705 is specifically used for:

[0176] The multimedia vector set is used as the behavior capsule feature, and the routing value between the behavior capsule feature and the initial preference capsule feature is calculated based on the dynamic routing mechanism of the capsule network according to the behavior capsule feature and the initial preference capsule feature.

[0177] The pairing coefficient is calculated based on the routing values.

[0178] The pairing coefficients are used as weight values ​​for behavioral capsule features, and a weighted calculation is performed to determine the candidate preference capsule feature set.

[0179] The actual preference feature vector of the target training object identifier is determined based on the candidate preference capsule feature set.

[0180] In one implementation, the training unit 705 is specifically used for:

[0181] The initial preference feature vector is determined by calculating the feature set of the candidate preference capsules based on the nonlinear squeezing function;

[0182] The attention structure network based on multilayer perceptron aggregates the feature set of the candidate preference capsules to obtain the actual preference feature vector of the target training object identifier.

[0183] In one implementation, the training unit 705 is specifically used for:

[0184] The first loss function is used to perform simulated loss calculation on the trained identifier embedding vector and the actual preference feature vector;

[0185] The training feature embedding vector and the actual preference feature vector are fused to obtain the training fused feature;

[0186] The second loss function is used to calculate the prediction loss of the multimedia feature vectors corresponding to the training fusion features and multimedia reference data.

[0187] Based on the simulated loss calculation and the predicted loss calculation, the first initial mapping module and the second initial mapping module included in the initial model are optimized and trained.

[0188] In one implementation, the training unit 705 is further configured to:

[0189] Determine a set of multimedia reference data for the target training object identifier recorded within a preset time range. The set of multimedia reference data includes one or more multimedia data that are recommended to the target training object identifier and are marked as having received positive feedback from the target training object identifier.

[0190] The multimedia reference data is determined from the multimedia reference data set.

[0191] In one implementation, the initial model includes: a fusion initial module, wherein the training fusion feature is obtained by fusing the training identifier embedding vector and the actual preference feature vector through the fusion initial module; the training unit 705 is further configured to:

[0192] Based on the simulated loss calculation and the predicted loss calculation, the fusion initial module included in the initial model is optimized and trained;

[0193] The determining unit 704 is specifically used for:

[0194] The fusion module, obtained by optimizing and training the initial fusion module, fuses the identifier embedding vector and the feature embedding vector to obtain the object embedding vector of the target object identifier.

[0195] In one implementation, the device further includes a recommendation unit 706, specifically used for:

[0196] Obtain the embedding vector of each multimedia data to be recommended in the multimedia dataset to be recommended;

[0197] Based on the object embedding vector and the recommendation embedding vector of each multimedia data to be recommended, the recommendation score of each multimedia data to be recommended is determined.

[0198] Based on the recommendation scores of each multimedia data to be recommended, target multimedia data is determined from the multimedia data to be recommended;

[0199] Multimedia data recommendation is performed on the target object identifier based on the target multimedia data.

[0200] It is understood that the division of units in the embodiments of this application is illustrative and merely a logical functional division; in actual implementation, there may be other division methods. The functional units in the embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units. Based on the same inventive concept, the principle and beneficial effects of the data processing device provided in the embodiments of this application in solving the problem are similar to the principle and beneficial effects of the data processing method in the embodiments of this application in solving the problem. Please refer to the principle and beneficial effects of the method implementation; for the sake of brevity, they will not be repeated here.

[0201] Please see Figure 8 This is a schematic diagram of a model training device for user feature data provided in an embodiment of this application. The device described in this embodiment can be applied to a smart device, such as a smart terminal, like a smartphone, tablet, smart wearable device, etc., or it can be a server. The model training device for user feature data described in this embodiment includes:

[0202] The acquisition unit 801 is used to acquire a set of training object identifiers from the object identifiers registered on the multimedia platform, and to acquire a training dataset corresponding to the set of training object identifiers. The training dataset includes: training feature data of the target training object identifier, training multimedia dataset of the target training object identifier, and multimedia data in which the user has applied a positive feedback operation as instructed by the target training object identifier.

[0203] The acquisition unit 801 is further configured to acquire the actual preference feature vector of the target training object identifier based on the training multimedia dataset of the target training object identifier;

[0204] The first generation unit 802 is used to perform vector generation processing on the training feature data of the target training object identifier through the first initial mapping module to generate the training identifier embedding vector of the target object identifier.

[0205] The second generation unit 803 is used to perform feature mapping processing on the training feature data of the target training object identifier through the second initial mapping module to generate the training feature embedding vector of the target object identifier.

[0206] Training unit 804 is used to optimize and train the first initial mapping module and the second initial mapping module included in the initial model based on the actual preference feature vector, the training feature embedding vector, and the training identifier embedding vector to obtain a vector generation model.

[0207] In one implementation, the training unit 804 is specifically used for:

[0208] The first loss function is used to perform simulated loss calculation on the trained identifier embedding vector ☆ and the actual preference feature vector;

[0209] The training feature embedding vector and the actual preference feature vector are fused to obtain the training fused feature;

[0210] The second loss function is used to calculate the prediction loss of the multimedia feature vectors corresponding to the training fusion features and multimedia reference data.

[0211] Based on the simulated loss calculation and the predicted loss calculation, the first initial mapping module and the second initial mapping module included in the initial model are optimized and trained.

[0212] In one implementation, the initial model includes: a fusion initial module, wherein the training fusion feature is obtained by fusing the training identifier embedding vector and the actual preference feature vector through the fusion initial module; the training unit 804 is further configured to:

[0213] Based on the simulated loss calculation and the predicted loss calculation, the fusion initial module included in the initial model is optimized and trained.

[0214] It is understood that the division of units in this application embodiment is illustrative and merely a logical functional division; in actual implementation, there may be other division methods. The functional units in this application embodiment can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units. Based on the same inventive concept, the principle and beneficial effects of the user feature data model training device provided in this application embodiment are similar to the principle and beneficial effects of the user feature data model training method in this application embodiment. Please refer to the implementation principle and beneficial effects of the method; for brevity, they will not be repeated here.

[0215] Please see Figure 9 This is a schematic diagram of the structure of a smart device provided in an embodiment of this application. The smart device described in this embodiment includes: a processor 901, a memory 902, and a network interface 903. The processor 901, memory 902, and network interface 903 can exchange data. The smart device described in this embodiment can be, for example, a smart terminal, such as a smartphone, tablet computer, smart wearable device, etc., and the smart device can also be a server.

[0216] The processor 901 described above can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0217] The aforementioned memory 902 may include read-only memory and random access memory, and provides program instructions and data to the processor 901. A portion of the memory 902 may also include non-volatile random access memory. The processor 901 executes the program instructions by calling them:

[0218] Obtain user characteristic data associated with the target object identifier;

[0219] The first mapping module is invoked to perform vector generation processing on the user feature data to generate the identifier embedding vector of the target object identifier; the first mapping module is trained based on the training multimedia dataset that has a relationship with each training object identifier in the training object identifier set and the training feature data associated with each training object identifier;

[0220] The user feature data is subjected to feature mapping processing to generate the feature embedding vector of the target object identifier;

[0221] Based on the identifier embedding vector and the feature embedding vector, determine the object embedding vector of the target object identifier;

[0222] The identifier embedding vector is used to characterize the multimedia preference features of the target object identifier, and the feature embedding vector is used to characterize the user features associated with the target object identifier.

[0223] In one implementation, the identifier embedding vector and the feature embedding vector are generated through a vector generation model, and the processor 901 is further configured to:

[0224] Obtain a set of training object identifiers from the object identifiers registered on the multimedia platform, and obtain the training dataset corresponding to the set of training object identifiers. The training dataset includes: training feature data of the target training object identifier, training multimedia dataset of the target training object identifier, and multimedia data in which the user has applied positive feedback operations as instructed by the target training object identifier.

[0225] The initial model is trained based on the training dataset to obtain a vector generation model, which includes a first mapping module and a second mapping module.

[0226] In one implementation, the initial model includes: a first initial mapping module and a second initial mapping module, the first initial mapping module corresponding to the first mapping module, and the second initial mapping module corresponding to the second mapping module; the processor 901 is specifically used for:

[0227] Based on the training multimedia dataset of the target training object identifier, obtain the actual preference feature vector of the target training object identifier;

[0228] The first initial mapping module performs vector generation processing on the training feature data of the target training object identifier to generate the training identifier embedding vector of the target object identifier.

[0229] The training feature data of the target training object identifier is processed by the second initial mapping module to generate the training feature embedding vector of the target object identifier.

[0230] The first initial mapping module and the second initial mapping module of the initial model are optimized and trained based on the actual preference feature vector, the training feature embedding vector, and the training identifier embedding vector to obtain the vector generation model.

[0231] In one implementation, the processor 901 is specifically used for:

[0232] Feature extraction is performed on each multimedia data in the training multimedia dataset of the target training object identifier to obtain a multimedia vector set;

[0233] The multimedia vector set is processed by a capsule network to obtain the actual preference feature vector of the target training object identifier;

[0234] The capsule network uses the multimedia vector set as behavioral capsule features, and performs feature processing based on the dynamic routing mechanism of the capsule network according to the behavioral capsule features and the initialized preference capsule features to obtain the actual preference feature vector of the target training object identifier.

[0235] In one implementation, the processor 901 is specifically used for:

[0236] The multimedia vector set is used as the behavior capsule feature, and the routing value between the behavior capsule feature and the initial preference capsule feature is calculated based on the dynamic routing mechanism of the capsule network according to the behavior capsule feature and the initial preference capsule feature.

[0237] The pairing coefficient is calculated based on the routing values.

[0238] The pairing coefficients are used as weight values ​​for behavioral capsule features, and a weighted calculation is performed to determine the candidate preference capsule feature set.

[0239] The actual preference feature vector of the target training object identifier is determined based on the candidate preference capsule feature set.

[0240] In one implementation, the processor 901 is specifically used for:

[0241] The initial preference feature vector is determined by calculating the feature set of the candidate preference capsules based on the nonlinear squeezing function;

[0242] The attention structure network based on multilayer perceptron aggregates the feature set of the candidate preference capsules to obtain the actual preference feature vector of the target training object identifier.

[0243] In one implementation, the processor 901 is specifically used for:

[0244] The first loss function is used to perform simulated loss calculation on the trained identifier embedding vector and the actual preference feature vector;

[0245] The training feature embedding vector and the actual preference feature vector are fused to obtain the training fused feature;

[0246] The second loss function is used to calculate the prediction loss of the multimedia feature vectors corresponding to the training fusion features and multimedia reference data.

[0247] Based on the simulated loss calculation and the predicted loss calculation, the first initial mapping module and the second initial mapping module included in the initial model are optimized and trained.

[0248] In one implementation, the processor 901 is further configured to:

[0249] Determine a set of multimedia reference data for the target training object identifier recorded within a preset time range. The set of multimedia reference data includes one or more multimedia data that are recommended to the target training object identifier and are marked as having received positive feedback from the target training object identifier.

[0250] The multimedia reference data is determined from the multimedia reference data set.

[0251] In one implementation, the initial model includes: a fusion initial module, wherein the training fusion feature is obtained by fusing the training identifier embedding vector and the actual preference feature vector through the fusion initial module; the processor 901 is further configured to:

[0252] Based on the simulated loss calculation and the predicted loss calculation, the fusion initial module included in the initial model is optimized and trained;

[0253] The fusion module, obtained by optimizing and training the initial fusion module, fuses the identifier embedding vector and the feature embedding vector to obtain the object embedding vector of the target object identifier.

[0254] In one implementation, the processor 901 is further configured to:

[0255] Obtain the embedding vector of each multimedia data to be recommended in the multimedia dataset to be recommended;

[0256] Based on the object embedding vector and the recommendation embedding vector of each multimedia data to be recommended, the recommendation score of each multimedia data to be recommended is determined.

[0257] Based on the recommendation scores of each multimedia data to be recommended, target multimedia data is determined from the multimedia data to be recommended;

[0258] Multimedia data recommendation is performed on the target object identifier based on the target multimedia data.

[0259] Based on the same inventive concept, the principle and beneficial effects of the intelligent device provided in the embodiments of this application in solving the problem are similar to the principle and beneficial effects of the data processing method in the embodiments of this application in solving the problem. Please refer to the principle and beneficial effects of the method implementation. For the sake of brevity, they will not be repeated here.

[0260] Please see Figure 10This is a schematic diagram of the structure of a smart device provided in an embodiment of this application. The smart device described in this embodiment includes: a processor 1001, a memory 1002, and a network interface 1003. The processor 1001, memory 1002, and network interface 1003 can exchange data. The smart device described in this embodiment can be, for example, a smart terminal, such as a smartphone, tablet computer, smart wearable device, etc., and the smart device can also be a server.

[0261] The processor 1001 mentioned above can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0262] The aforementioned memory 1002 may include read-only memory and random access memory, and provides program instructions and data to the processor 1001. A portion of the memory 1002 may also include non-volatile random access memory. The processor 1001 executes the program instructions by calling them:

[0263] Obtain a set of training object identifiers from the object identifiers registered on the multimedia platform, and obtain the training dataset corresponding to the set of training object identifiers. The training dataset includes: training feature data of the target training object identifier, training multimedia dataset of the target training object identifier, and multimedia data in which the user has applied positive feedback operations as instructed by the target training object identifier.

[0264] Based on the training multimedia dataset of the target training object identifier, obtain the actual preference feature vector of the target training object identifier;

[0265] The first initial mapping module performs vector generation processing on the training feature data of the target training object identifier to generate the training identifier embedding vector of the target object identifier.

[0266] The training feature data of the target training object identifier is processed by the second initial mapping module to generate the training feature embedding vector of the target object identifier.

[0267] The first initial mapping module and the second initial mapping module of the initial model are optimized and trained based on the actual preference feature vector, the training feature embedding vector, and the training identifier embedding vector to obtain the vector generation model.

[0268] In one implementation, the processor 1001 is specifically used for:

[0269] The first loss function is used to perform simulated loss calculation on the trained identifier embedding vector ☆ and the actual preference feature vector;

[0270] The training feature embedding vector and the actual preference feature vector are fused to obtain the training fused feature;

[0271] The second loss function is used to calculate the prediction loss of the multimedia feature vectors corresponding to the training fusion features and multimedia reference data.

[0272] Based on the simulated loss calculation and the predicted loss calculation, the first initial mapping module and the second initial mapping module included in the initial model are optimized and trained.

[0273] In one implementation, the initial model includes: a fusion initial module, wherein the training fusion feature is obtained by fusing the training identifier embedding vector and the actual preference feature vector through the fusion initial module; the processor 1001 is further configured to:

[0274] Based on the simulated loss calculation and the predicted loss calculation, the fusion initial module included in the initial model is optimized and trained.

[0275] Based on the same inventive concept, the principle and beneficial effects of the intelligent device provided in the embodiments of this application in solving the problem are similar to the principle and beneficial effects of the model training method for user feature data in the method embodiments of this application in solving the problem. Please refer to the principle and beneficial effects of the method implementation. For the sake of brevity, they will not be repeated here.

[0276] This application also provides a computer storage medium storing program instructions, which, when executed, may include, for example... Figure 2 or Figure 4 The data processing method in the corresponding embodiments may further include some or all of the steps, such as Figure 6 Some or all of the steps of the model training method for user feature data in the corresponding embodiments.

[0277] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0278] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0279] This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. The processor of a smart device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the smart device to perform the steps described in the embodiments of the above methods.

[0280] The data processing method, model training method, apparatus, device, and storage medium provided in the embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are 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. A data processing method, characterized in that, include: Obtain user characteristic data associated with the target object identifier; The first mapping module is invoked to perform vector generation processing on the user feature data to generate the identifier embedding vector of the target object identifier; The first mapping module is trained based on a training multimedia dataset that is associated with each training object identifier in the training object identifier set and training feature data associated with each training object identifier. The user feature data is subjected to feature mapping processing to generate a feature embedding vector for the target object identifier; Based on the identifier embedding vector and the feature embedding vector, an object embedding vector of the target object identifier is determined; wherein, the identifier embedding vector is used to characterize the multimedia preference features of the target object identifier, and the feature embedding vector is used to characterize the user features associated with the target object identifier; The identifier embedding vector and the feature embedding vector are generated through a vector generation model, and the method further includes: Obtain a set of training object identifiers from the object identifiers registered on the multimedia platform, and obtain the training dataset corresponding to the set of training object identifiers. The training dataset includes: training feature data of the target training object identifier, training multimedia dataset of the target training object identifier, and multimedia data in which the user has applied positive feedback operations as instructed by the target training object identifier. The initial model is trained based on the training dataset to obtain a vector generation model, which includes the first mapping module; wherein, training the initial model includes training based on the actual preference feature vector of each training object identifier; the acquisition of the actual preference feature vector includes: Feature extraction is performed on each multimedia data in the training multimedia dataset of the target training object identifier to obtain a multimedia vector set; the multimedia vector set is then processed by a capsule network to obtain the actual preference feature vector of the target training object identifier; wherein, the capsule network uses the multimedia vector set as behavioral capsule features, and performs feature processing based on the behavioral capsule features and the initialized preference capsule features, according to the dynamic routing mechanism of the capsule network, to obtain the actual preference feature vector of the target training object identifier.

2. The method as described in claim 1, characterized in that, The initial model includes: a first initial mapping module and a second initial mapping module, wherein the first initial mapping module corresponds to the first mapping module, and the second initial mapping module corresponds to the second mapping module included in the vector generation module; training the initial model based on the training dataset to obtain the vector generation model includes: Based on the training multimedia dataset of the target training object identifier, obtain the actual preference feature vector of the target training object identifier; The first initial mapping module performs vector generation processing on the training feature data of the target training object identifier to generate the training identifier embedding vector of the target object identifier. The training feature data of the target training object identifier is processed by the second initial mapping module to generate the training feature embedding vector of the target object identifier. The first initial mapping module and the second initial mapping module of the initial model are optimized and trained based on the actual preference feature vector, the training feature embedding vector, and the training identifier embedding vector to obtain the vector generation model.

3. The method as described in claim 2, characterized in that, The step of performing feature processing on the multimedia vector set through a capsule network to obtain the actual preference feature vector of the target training object identifier includes: The multimedia vector set is used as the behavior capsule feature, and the routing value between the behavior capsule feature and the initial preference capsule feature is calculated based on the dynamic routing mechanism of the capsule network according to the behavior capsule feature and the initial preference capsule feature. The pairing coefficient is calculated based on the routing values. The pairing coefficients are used as weight values ​​for behavioral capsule features, and a weighted calculation is performed to determine the candidate preference capsule feature set. The actual preference feature vector of the target training object identifier is determined based on the candidate preference capsule feature set.

4. The method as described in claim 3, characterized in that, The step of determining the actual preference feature vector of the target training object identifier based on the candidate preference capsule feature set includes: The initial preference feature vector is determined by calculating the feature set of the candidate preference capsules based on the nonlinear squeezing function; The attention structure network based on multilayer perceptron aggregates the feature set of the candidate preference capsules to obtain the actual preference feature vector of the target training object identifier.

5. The method as described in claim 2, characterized in that, The optimization training of the first initial mapping module and the second initial mapping module of the initial model based on the actual preference feature vector, the trained feature embedding vector, and the trained identifier embedding vector to obtain the vector generation model includes: The first loss function is used to perform simulated loss calculation on the trained identifier embedding vector and the actual preference feature vector; The training feature embedding vector and the actual preference feature vector are fused to obtain the training fused feature; The second loss function is used to calculate the prediction loss of the multimedia feature vectors corresponding to the training fusion features and multimedia reference data. Based on the simulated loss calculation and the predicted loss calculation, the first initial mapping module and the second initial mapping module included in the initial model are optimized and trained.

6. The method as described in claim 5, characterized in that, Before calculating the prediction loss of the multimedia feature vectors corresponding to the training fusion features and multimedia reference data using the second loss function, the method further includes: Determine a set of multimedia reference data for the target training object identifier recorded within a preset time range. The set of multimedia reference data includes one or more multimedia data that are recommended to the target training object identifier and are marked as having received positive feedback from the target training object identifier. The multimedia reference data is determined from the multimedia reference data set.

7. The method as described in claim 5, characterized in that, The initial model includes a fusion initial module, wherein the training fusion feature is obtained by fusing the training identifier embedding vector and the actual preference feature vector through the fusion initial module. The method further includes: Based on the simulated loss calculation and the predicted loss calculation, the fusion initial module included in the initial model is optimized and trained; The step of determining the object embedding vector of the target object identifier based on the identifier embedding vector and the feature embedding vector includes: The fusion module, obtained by optimizing and training the initial fusion module, fuses the identifier embedding vector and the feature embedding vector to obtain the object embedding vector of the target object identifier.

8. The method according to any one of claims 1-7, characterized in that, The method further includes: Obtain the embedding vector of each multimedia data to be recommended in the multimedia dataset to be recommended; Based on the object embedding vector and the recommendation embedding vector of each multimedia data to be recommended, the recommendation score of each multimedia data to be recommended is determined. Based on the recommendation scores of each multimedia data to be recommended, target multimedia data is determined from the multimedia data to be recommended; Multimedia data recommendation is performed on the target object identifier based on the target multimedia data.

9. A method for training a model using user feature data, characterized in that, The initial model to be trained includes a first initial mapping module and a second initial mapping module, and the method includes: Obtain a set of training object identifiers from the object identifiers registered on the multimedia platform, and obtain the training dataset corresponding to the set of training object identifiers. The training dataset includes: training feature data of the target training object identifier, training multimedia dataset of the target training object identifier, and multimedia data in which the user has applied positive feedback operations as instructed by the target training object identifier. Based on the training multimedia dataset of the target training object identifier, obtain the actual preference feature vector of the target training object identifier; The first initial mapping module performs vector generation processing on the training feature data of the target training object identifier to generate the training identifier embedding vector of the target object identifier. The training feature data of the target training object identifier is processed by the second initial mapping module to generate the training feature embedding vector of the target object identifier. Based on the actual preference feature vector, the trained feature embedding vector, and the trained identifier embedding vector, the first initial mapping module and the second initial mapping module of the initial model are optimized and trained to obtain the vector generation model. The step of optimizing and training the first initial mapping module and the second initial mapping module of the initial model based on the actual preference feature vector, the trained feature embedding vector, and the trained identifier embedding vector to obtain the vector generation model includes: The first loss function is used to perform simulated loss calculation on the trained identifier embedding vector and the actual preference feature vector; The training feature embedding vector and the actual preference feature vector are fused to obtain the training fused feature; The second loss function is used to calculate the prediction loss of the multimedia feature vectors corresponding to the training fusion features and multimedia reference data. Based on the simulated loss calculation and the predicted loss calculation, the first initial mapping module and the second initial mapping module included in the initial model are optimized and trained.

10. The method as described in claim 9, characterized in that, The initial model includes a fusion initial module, wherein the training fusion feature is obtained by fusing the training identifier embedding vector and the actual preference feature vector through the fusion initial module. The method further includes: Based on the simulated loss calculation and the predicted loss calculation, the fusion initial module included in the initial model is optimized and trained.

11. A smart device, characterized in that, The intelligent device includes a memory and a processor; The memory stores computer programs; The processor invokes a computer program stored in the memory to implement the method as described in any one of claims 1-10.

12. A computer storage medium, characterized in that, The computer storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-10.

13. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method as described in any one of claims 1-10.

Images