Model training method, device, electronic device and computer program

Abstract

Embodiments of the present application disclose a model training method, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product, wherein the method includes the steps of: acquiring a data set, the data set including first domain data and second domain data, the first domain data including at least one object and first resource data of interest within the first domain for each object, and the second domain data including at least one object and second resource data of interest within the second domain for each object; invoking an interest alignment model to perform feature extraction on the first domain data to obtain a first domain feature representation; and The method includes: calling the interest alignment model to perform feature extraction on the second neighbor data to obtain a second domain feature representation; calling the interest alignment model to perform the following process, i.e., based on the first domain feature representation and the second domain feature representation, perform interest alignment processing between the first domain and the second domain; and training the interest alignment model according to the direction of reducing the first loss and the second loss to obtain a trained interest alignment model, where the first loss is a loss corresponding to the feature extraction and the second loss is a loss corresponding to the interest alignment processing, and the trained interest alignment model is used to recommend resource data for the target object in the second domain.

Description

[Technical Field]

[0001] This application is based on a Chinese patent application bearing application number 202211507481.6 and filed on November 29, 2022, and claims priority from the Chinese patent application, the entire contents of which are hereby incorporated by reference into this application.

[0002] The present application relates to the field of computers, and more particularly to the field of machine learning, and more particularly to model training methods, apparatus, electronic devices, computer-readable storage media, and computer program products. [Background technology]

[0003] A personalized recommender system is a platform used to make personalized resource recommendations for users (i.e., objects). For example, a personalized recommender system in the video scene supports recommending interesting video resources for users.

[0004] Currently, personalized recommendation systems support realizing personalized recommendation targets by collecting a large amount of behavioral data (e.g., interaction data between the user and resource data) of users (i.e., objects), constructing a resource recommendation model using a pre-designed recommendation algorithm, and generating a specific recommendation list for the object through the constructed resource recommendation model. However, through practice, it has been found that personalized recommendation systems do not possess abundant data for all objects. Therefore, a resource recommendation model trained based on a large amount of interaction data of an object cannot be applied to making resource recommendations for objects with relatively little interaction data, resulting in a relatively poor recommendation effect of the resource recommendation model. Summary of the Invention [Problem to be solved by the invention]

[0005] Embodiments of the present application provide a model training method, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product, which can utilize interest relationships between objects to train an interest alignment model and ensure the performance of the interest alignment model. [Means for solving the problem]

[0006] An embodiment of the present application provides a model training method, the method being performed by an electronic device, the method comprising: obtaining a data set, the data set including first domain data and second domain data, the first domain data including at least one object and first resource data of interest within the first domain for each object, and the second domain data including at least one object and second resource data of interest within the second domain for each object; Invoking an interest alignment model to perform feature extraction on the first domain data to obtain a first domain feature representation, and invoking an interest alignment model to perform feature extraction on the second neighbor data to obtain a second domain feature representation; calling the interest alignment model to perform the following process: performing interest alignment between the first domain and the second domain based on the first domain feature representation and the second domain feature representation; training the interest alignment model according to a direction of reducing the first loss and the second loss to obtain a trained interest alignment model; Here, the first loss is a loss corresponding to the feature extraction, and the second loss is a loss corresponding to the interest alignment process, and the trained interest alignment model is used to recommend resource data for the target object in the second domain.

[0007] An embodiment of the present application provides a model training device, the device comprising: an acquisition unit and a processing unit; The acquiring unit is configured to acquire a data set, the data set including first field data and second field data, the first field data including at least one object and first resource data of interest within the first field for each object, and the second field data including at least one object and second resource data of interest within the second field for each object; The processing unit is configured to invoke an interest alignment model to perform feature extraction on first domain data to obtain a first domain feature representation, and to invoke an interest alignment model to perform feature extraction on the second neighborhood data to obtain a second domain feature representation; The processing unit is further configured to call the interest alignment model to perform the following process: interest alignment between the first domain and the second domain based on the first domain feature representation and the second domain feature representation; The processing unit is further configured to train the interest alignment model according to a direction that reduces the first loss and the second loss, and obtain a trained interest alignment model; Here, the first loss is a loss corresponding to the feature extraction, and the second loss is a loss corresponding to the interest alignment process, and the trained interest alignment model is used to recommend resource data for the target object in the second domain.

[0008] An embodiment of the present application provides an electronic device, the electronic device comprising: a processor adapted to load and execute computer-executable instructions; a computer-readable storage medium having computer-executable instructions stored thereon, the computer-executable instructions, when executed by a processor, implementing the model training method.

[0009] An embodiment of the present application provides a computer-readable storage medium having computer-executable instructions stored thereon, the computer-executable instructions being suitable for being loaded by a processor and for executing the model training method.

[0010] An embodiment of the present application provides a computer program product, the computer program product including computer-executable instructions stored in a computer-readable storage medium, a processor of an electronic device reading the computer-executable instructions from the computer-readable storage medium, and the computer-executable instructions, when executed by the processor, realize the above-mentioned model training method. [Effects of the Invention]

[0011] In an embodiment of the present application, a data set is constructed by obtaining interest resource data in a corresponding domain of at least one object from multiple domains. For example, it supports obtaining first domain data from a first domain and second domain data from a second domain. Here, the first domain may be a domain with relatively abundant interaction data, while the second domain may be neighboring data with relatively scarce interaction data. Then, an interest alignment model is invoked to perform feature extraction on the first domain data and the second domain data, respectively, to extract first domain feature representations and second domain feature representations. An interest alignment model is invoked to perform interest alignment processing between the two domains based on the first domain feature representations and the second domain feature representations. Finally, it supports training the interest alignment model in a way that reduces the first loss in the feature extraction process and the loss in the interest alignment process. The embodiments of the present application take into full consideration the interest similarity between objects and adopt an object interest similarity alignment method to support training of an interest alignment model, so that the interest alignment model can generate a relatively accurate feature representation for an object even when applied to a domain with relatively little interaction data. For example, considering that the interaction data in a first domain is relatively abundant, the feature representation of the first domain extracted based on abundant interaction data is more accurate. Therefore, the interest alignment model is trained by adopting a method of aligning the interest similarity between two objects in a second domain with relatively little interaction data to the interest similarity between the two objects in the first domain with relatively much interaction data. Even when the interaction data in the second domain is relatively sparse, the trained interest alignment model can ensure that it can generate a relatively accurate feature representation for the object, so that resource data recommended for the object based on the accurate feature representation can meet the personalized needs of the object and achieve high-precision resource recommendation.The embodiment of the present application adopts a method of jointly training the interest alignment model using the first loss in the feature extraction process and the loss in the interest alignment processing process, which can enrich the learning targets of the entire interest alignment model and ensure the performance of the interest alignment model through multi-target learning.

[0012] In order to more clearly explain the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without making any creative efforts. [Brief explanation of the drawings]

[0013] [Figure 1a] 1 is a schematic diagram of the architecture of a resource recommendation system provided by one exemplary embodiment of the present application; [Figure 1b] FIG. 2 is a schematic diagram of the architecture of another resource recommendation system provided by an exemplary embodiment of the present application; [Figure 2] FIG. 1 is a schematic diagram of a model structure of one type of interest alignment model provided by one exemplary embodiment of the present application. [Figure 3] FIG. 1 is a flow diagram of a model training method provided by an exemplary embodiment of the present application. [Figure 4] FIG. 2 is a schematic diagram of constructing a second domain object resource graph provided by an exemplary embodiment of the present application; [Figure 5] FIG. 1 is a schematic diagram of one type of graph coding learning architecture provided by one exemplary embodiment of the present application. [Figure 6] FIG. 2 is a schematic diagram of one type of interest alignment process between a first field and a second field provided by one exemplary embodiment of the present application; [Figure 7]1 is a schematic diagram of one type of interest alignment process performed in a first field and a second field, respectively, provided by one exemplary embodiment of the present application. [Figure 8] FIG. 2 is a schematic diagram of one type of graph reconstruction provided by one exemplary embodiment of the present application. [Figure 9] FIG. 1 is a flow diagram of another model training method provided by an exemplary embodiment of the present application. [Figure 10] FIG. 1 is a schematic diagram of matching the fused feature representation of one target object with the resource feature representation of each candidate resource data provided by an exemplary embodiment of the present application; [Figure 11] 1 is a structural schematic diagram of a model training device provided by an exemplary embodiment of the present application; [Figure 12] 1 is a structural schematic diagram of an electronic device provided by an exemplary embodiment of the present application; DETAILED DESCRIPTION OF THE INVENTION

[0014] The following will clearly and completely describe the technical solutions in the embodiments of the present application in combination with the drawings in the embodiments of the present application, but of course, the described embodiments are only some of the embodiments of the present application, not all of the embodiments, and all other embodiments that can be obtained by those skilled in the art based on the embodiments of the present application without making any creative efforts are all within the scope of protection of the present application.

[0015] In the embodiments of the present application, a model training solution is provided, specifically, a solution for training and applying an interest alignment model under resource recommendation scenarios is provided. Below, technical terms and related concepts related to the model training solution provided in the embodiments of the present application are briefly introduced.

[0016] 1. Artificial Intelligence (AI).

[0017] Artificial intelligence (AI) refers to theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and enhance human intelligence to sense the environment, acquire knowledge, and use that knowledge to achieve optimal results. AI technology is a comprehensive academic field that encompasses a wide range of related fields, including both hardware and software. AI foundational technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technology primarily encompasses several major areas, such as computer vision, speech processing, natural language processing, and machine learning / deep learning. Machine learning (ML) is a multidisciplinary field that incorporates multiple disciplines, including probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory, and focuses on the study of how computers can simulate or realize human learning behavior, acquire new knowledge or skills, reorganize existing knowledge structures, and continuously improve their own performance. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, supervised learning, and meta-learning.

[0018] The model training solution provided in the embodiments of this application mainly relates to artificial neural networks and transfer learning in machine learning. Below, we will briefly introduce artificial neural networks and transfer learning, respectively.

[0019] (1) An artificial neural network is a method for achieving machine learning tasks. When discussing neural networks in the field of machine learning, it generally refers to "neural network learning." This is a type of network structure composed of many simple units. Such network structures resemble biological nervous systems and are used to simulate interactions between living organisms and their natural environments. The more network structures there are, the more functionally rich the neural network often becomes. Neural networks are a relatively broad concept, and neural network models tailored to specific learning tasks, such as speech, text, and images, have been derived. For example, graph neural networks (GNNs) are examples of graph neural networks. A graph neural network is a type of algorithm that uses deep learning models to perform feature mining and extraction on graph-structured data. Specifically, it is a general term for algorithms that can extract and explore features and patterns in graph-structured data through learning from graph-structured data, thereby meeting the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. For example, graph neural networks that meet the needs of the above graph learning tasks may include, but are not limited to, graph convolutional networks (GCNs), graph recurrent networks (GRNs), and graph attention networks (GATs).

[0020] The embodiments of the present application specifically relate to a graph attention network included in a graph neural network, which includes an attention mechanism that allows the neural network to focus only on information required for task learning and select specific inputs. That is, the introduction of the attention mechanism in the graph attention network allows the neural network to focus on graph structure data (e.g., nodes and edges included in the graph) that are more relevant to the task, thereby improving training effectiveness and test accuracy. Here, the attention mechanism simulates and produces human visual attention. For example, the human eye has a relatively wide field of view, but the field of view often only focuses on a small range. That is, the human eye is more likely to focus on important areas (i.e., areas of interest) in the field of view. Therefore, the attention mechanism aims to focus on important information in a task and omit unimportant information, similar to human visual attention. Attention mechanisms can be further subdivided into self-attention mechanisms, multi-head self-attention mechanisms, and global attention mechanisms.

[0021] (2) Transfer learning is a common technique used to address the lack of sufficient training data and / or annotations. Specifically, it involves improving a new task by transferring knowledge from related tasks that have already been learned, for example, transferring (copying) pre-trained model parameters (using data and annotations from a non-current task) to a new network model to aid training. While most machine learning algorithms are designed to solve a single task, developing algorithms that facilitate transfer learning is a topic of ongoing interest in machine learning research.

[0022] Transfer learning is based on the concept of a domain. A domain is the main body of transfer learning and is composed of data features and feature distributions. In resource recommendation scenarios, a domain is generally used to represent different application domains, such as shopping, movies, reading, videos, social media accounts, documents, and live streaming. As can be seen from the above description, the task of transfer learning is to apply a model trained in an old domain to a new domain based on problem similarity. Therefore, domains related to transfer learning often include at least two domains, a source domain and a target domain. Here, the source domain is a domain different from the domain to which the sample data (i.e., resource data) of the model to be trained belongs. The source domain often contains abundant supervision information (i.e., contains abundant resource data used to train the model). The target domain is the same domain as the domain to which the sample data of the model to be trained belongs. The target domain usually contains only a small amount of resource data used to train the model.

[0023] Second, resource recommendations

[0024] Resource recommendation, also referred to as resource distribution, may refer to a process in a resource recommendation platform (or a resource recommendation system, or the personalized recommendation system mentioned above) of distributing resource data included in the platform to platform objects (e.g., one or more resource recipients who have registered a platform account or temporarily logged in to the platform). Here, the resource data included in the resource recommendation platform may be referred to as Internet resources (or simply referred to as resources), including, but not limited to, video (which may be divided into long videos and short videos according to the duration of the video), audio (e.g., music, voice audio, etc.), animation, or literature (e.g., journals, magazines, papers, etc.). The embodiments of the present application do not limit the resource type of the resource data. The resource type of the resource data is related to the field (i.e., the domain mentioned above) to which the resource recommendation platform that distributes the resource data belongs. The resource types of resource data distributed by different domains may be different. For example, the resource type of resource data in the live streaming field is video, while the resource type of resource data in the document field is document. The embodiments of the present application do not limit the resource type of the resource data.

[0025] A resource recommendation platform may refer to an application program that supports the distribution or recommendation of resource data. An application program may refer to a computer program for completing a specific task or tasks. Categorizing application programs according to different dimensions (e.g., the application program's operation method and functions) may result in different types of the same application program. For example, when classified according to the application program's operation method, application programs may include, but are not limited to, client terminals installed on a terminal, mini-programs (subprograms of the client terminal) that can be used without downloading and installation, and web (World Wide Web) application programs that open through a browser. Furthermore, when classified according to the application program's function type, application programs may include, but are not limited to, instant messaging (IM) application programs and content interaction application programs. Here, instant messaging application programs refer to application programs for internet-based instant messaging and social interaction, and instant messaging application programs may include, but are not limited to, social application programs with communication functions, map application programs with social interaction functions, and game application programs. The content interaction application program refers to an application program that can realize content interaction, and may be, for example, an application program such as online banking, a sharing platform, a personal space, and news.

[0026] The resource recommendation platform may be a plug-in (or function) that supports resource recommendation and is included in the application program mentioned above. For example, if the application program is a client terminal-type instant messaging application program, in this case, the resource distribution platform may be a resource distribution plug-in included in the instant messaging application program. For example, when the resource data is a short video, the resource recommendation function provided by the instant messaging application program is a short video recommendation function. In this way, even when a target object (e.g., any object using an instant messaging application program) is using the instant messaging application program to communicate, functions such as resource browsing and publishing can be performed, and no application jump (e.g., jumping from the instant messaging application program to a separate resource recommendation application program) is required.

[0027] It should be noted that the type of resource data distributed by a resource recommendation platform is not limited to one type. For example, the resource types of resource data supported by the resource recommendation platform may include both videos and documents. Furthermore, the embodiments of the present application do not limit the resource types of resource data distributed by the resource recommendation platform, nor the specific types of application programs that the resource recommendation platform is, or the resource recommendation functions provided by the application programs. For ease of discussion, the following embodiments will be described using the example of a resource distributed by a resource recommendation platform (or a resource recommendation system) being a short video.

[0028] Based on the above related descriptions of transfer learning and resource recommendation, the embodiments of the present application propose a model training solution, which relates to cross-domain recommendation. Cross-domain recommendation aims to combine data from multiple domains and introduce information from other domains (e.g., at least one source domain) to assist in capturing certain preferences or interests of an object through analyzing interaction data of the object in other domains, so as to provide better recommendations in the target domain and even in multiple domains.

[0029] In a specific implementation, the model training solution provided in the embodiments of the present application is a training solution for an interest alignment model, which may be referred to as a cross-domain recommendation model, etc. Specifically, the interest relationships between objects, i.e., the interest relationships between different objects and resource data (for example, the resource data that object A and object B are interested in are similar), are used to perform interest alignment learning for the interest alignment model, including cross-domain and intra-domain object interest relationships, thereby achieving the purpose of guiding target domain interest learning. A schematic training flow of the model training solution may include: obtaining a data set used for model training, where the data set includes first-domain data in a first domain and second-domain data in a second domain; then calling the interest alignment model to perform feature extraction on the first-domain data to obtain a first-domain feature representation (i.e., an embedding vector); and calling the interest alignment model to perform feature extraction on the second domain to obtain a second-domain representation. Based on the first and second domain feature representations, interest alignment (also called interest similarity alignment) is performed in the first and second domains, including cross-domain alignment and inter-domain alignment. Finally, the interest alignment model is trained according to the direction of reducing the first loss in the feature extraction process and the loss in the interest alignment process, to obtain a trained interest alignment model.

[0030] Based on the model training process, a trained interest alignment model can be obtained, which is used to recommend resource data for a target object in a second domain (e.g., any one of the objects in the resource data to be distributed). Specifically, after obtaining the trained interest alignment model, the embodiment of the present application further invokes the trained interest alignment model to support recommending resource data for a target object logging in to the resource recommendation platform. Specifically, the interest alignment model is employed to generate an accurate object feature representation (i.e., an embedding vector capable of characterizing the attributes or features of the target object) for the target object in the resource data to be distributed, and then, according to the object feature representation, a portion of resource data that meets the interests or hobbies of the target object is screened from the database of the resource recommendation platform, and resource data is subsequently selected from the portion of resource data to be recommended for the target object.

[0031] As can be seen from the above, the embodiments of the present application combine transfer learning to utilize the abundant knowledge and information of the first domain to enrich the data of the second domain, increase the available information, alleviate the problem of the sparse number of samples in the second domain, and ensure the model performance of the trained interest alignment model. When the trained interest alignment model is applied to the second domain with relatively little interaction data, it can also generate a relatively accurate feature representation for the object, so that the resource data recommended for the object based on the accurate feature representation can meet the object personalization needs, and achieve high-precision resource recommendation.

[0032] To facilitate understanding of the model training solution provided by the embodiments of the present application, the resource recommendation scenario related to the embodiments of the present application will be briefly introduced below by combining the resource recommendation system shown in Figure 1a. As shown in Figure 1a, the resource recommendation system includes a terminal 101, a server 102, and a server 103, and the embodiments of the present application do not limit the number and names of the terminal 101, the server 102, and the server 103.

[0033] The terminal 101 may refer to a terminal device used by a resource recipient of the resource data to be distributed, which is registered with the resource recommendation platform. The terminal device may include, but is not limited to, a smartphone (e.g., a smartphone equipped with the Android system or an Internetworking Operating System (IOS)), a tablet computer, a portable personal computer, a mobile Internet device (MID), an in-vehicle device, and a head-mounted device. The embodiments of the present application do not limit the type of terminal device and are described herein. The terminal device is equipped with a resource recommendation platform, specifically, an application program that implements the resource recommendation platform. The resource recommendation platform may be a recommendation platform for a second category. In this way, the resource recipient may perform operations such as receiving resource data for a second category through the resource recommendation platform equipped in the terminal device.

[0034] The server 102 is a server supported by the terminal 101, specifically a back-end server of a resource recommendation platform installed in the terminal 101, and is used to realize the provision of computational and application service support for the resource recommendation platform in the terminal 101 by interacting with the terminal 101. The server 103 may be a back-end server supported by a resource recommendation platform for a first domain, and the server 103 may communicate with the server 102 to provide data for model training in the first domain. The servers (e.g., the servers 102 and 103) may be independent physical servers, server clusters or distributed systems consisting of multiple physical servers, or cloud servers providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal 101 and the servers may be directly or indirectly connected via wired or wireless communication, but this application does not limit the same.

[0035] The server 102 further includes a database 1021, which may be used to store all resource data, objects, and object interaction data (user-item) between the objects and the resource data included in the resource recommendation platform of the second field (i.e., the target domain). As an example, under a video recommendation scenario, if the video platform includes Object 1 and Object 2, and Object 1 has clicked Video 1 and Video 2, and Object 2 has clicked Video 1 and Video 3, the database may store interaction data between the objects and the videos, such as Object 1-Video 1 (representing that Object 1 has triggered Video 1), Object 1-Video 2 (representing that Object 1 has triggered Video 2), Object 2-Video 1 (representing that Object 2 has triggered Video 1), and Object 2-Video 2 (representing that Object 2 has triggered Video 2). Similarly, the server 103 also includes a database 1031, which may be used to store all resource data, objects, and object interaction data between objects and resource data included in the resource recommendation platform of the first field (i.e., source domain). The specific content of the object interaction data may refer to the related description of the second field above, but will not be described in detail here.

[0036] There are two long-standing problems in resource recommendation systems: data sparseness and cold start. The so-called data sparseness problem refers to the fact that there is very little object interaction data for objects and resource data, making it very difficult to accurately capture the interests of objects or the characteristics of resource data based on a small amount of object interaction data. The so-called cold start problem refers to the fact that an object newly registered or logged in to a resource recommendation platform, or newly uploaded or published resource data to a resource recommendation platform, has no object interaction data. The degree of data sparseness and the scarcity of object interaction data under cold start make it difficult for traditional recommendation algorithms that make recommendations based on object interaction data of objects and resource data to make appropriate recommendations under these two types of cases. Differently, the present embodiment combines transfer learning to train the interest alignment model, i.e., it uses the abundant knowledge and information of the first domain to help improve the performance of the second domain and reduce the number of training samples required by the second domain, thereby more effectively resolving the problem of a relatively small number of training samples leading to data sparseness and cold start in recommendation scenes. Thus, when the interest alignment model obtained through cross-domain training is applied to the second domain, accurate feature representations can be generated for objects that still lack object interaction data or have sparse object interaction data, thereby facilitating the recommendation of resource data (i.e., resources included in the target domain) for users based on the accurate feature representations. As can be seen from the above, the model training solution provided by the present embodiment can be applied to resource recommendation scenes, and is particularly effective when applied to model training in data-sparse or cold-start scenes.

[0037] The model training solution provided by the embodiment of the present application may be executed by either the terminal 101 or the server 102 in the architecture shown in FIG. 1a, or may be executed jointly. That is, the electronic device serving as the execution body of the embodiment of the present application may be at least one of the terminal 101 and the server 102. The training process of the interest alignment model mentioned above is executed by the server 102, and the trained interest alignment model may be directly installed in the server 102 as shown in FIG. 1a. In this way, the trained interest alignment model is called by the server 102 each time a resource is distributed to realize resource recommendation. In this implementation manner, the electronic device serving as the execution body used to execute the solution provided by the embodiment of the present application is the server 102. The trained interest alignment model may also be installed in the terminal 101. As shown in FIG. 1b, after the interest alignment model is trained in the server 102, the trained interest alignment model is sent to the terminal 101, and the trained interest alignment model is then installed in the terminal 101. At this time, the electronic device that is the execution body of the model training solution means provided by the embodiment of the present application includes the terminal 101 and the server 102. If the training of the interest alignment model is performed by the terminal 101, the trained interest alignment model may be directly installed in the terminal 101, and at this time, the electronic device that is the execution body of the model training solution means provided by the embodiment of the present application includes the terminal 101.

[0038] 1a and 1b are merely schematic diagrams of an exemplary architecture of a resource recommendation system provided by an embodiment of the present application. In practical applications, the architecture may undergo adaptive changes. For example, the number of first domains provided by the embodiment of the present application may be multiple, i.e., the embodiment of the present application may be adapted to the case of multiple source domains. In this way, information from more domains may be used to support learning of the target domain, and the number of servers included in the architecture will increase accordingly. Furthermore, when the embodiment of the present application is implemented in a specific product or technology, for example, when making a resource recommendation based on resource data, it is necessary to obtain attribute information of the target object (e.g., information such as age, gender, or name), and therefore it is necessary to obtain the target object's permission or consent. Furthermore, the collection, use, and processing of related data must comply with relevant national and local laws, regulations, and standards.

[0039] Based on the model training solution described above, it can be seen that the model training solution provided by the embodiment of the present application mainly relates to two aspects: one is to train a model to obtain a trained interest alignment model; and the other is to employ the trained interest alignment model to perform resource recommendation (i.e., model application). To facilitate a better understanding of the method steps of the model training method provided subsequently in the embodiment of the present application, the following first combines the model structure schematic diagram of the interest alignment model provided by the embodiment of the present application, shown in Figure 2. As shown in Figure 2, the interest alignment model provided by the embodiment of the present application (also referred to as an object interest alignment model) may include a single-domain feature extraction module, a cross-domain alignment module, and an intra-domain alignment module.

[0040] The single-domain feature extraction module is used to perform feature extraction on domain data under a single domain (e.g., a certain domain) to obtain a domain feature representation for the single domain. Depending on the number of domains in a practical application, the number of single-domain feature extraction modules may vary. For example, in the case of a source domain (i.e., a first domain) and a target domain (i.e., a second domain), the single-domain feature extraction module may include a first feature extraction module and a second feature extraction module. Here, the first feature extraction module is used to perform feature extraction on the first domain data under the first domain to obtain a first domain feature representation, and the second feature extraction module is used to perform feature extraction on the second domain data under the second domain to obtain a second domain feature representation. It is worth noting that the first feature extraction module corresponding to the first domain is obtained through pre-training, i.e., in the pre-training stage, the first feature extraction model may be pre-trained to learn the object's interests and hobbies in the first domain. In this way, with the help of object interaction data from the first domain, the first feature extraction module and the feature representation of the object and resource data in the first domain can be optimized. When performing task learning for the second domain, the pre-trained first feature extraction module is directly transferred to the learning task of the second domain. That is, in the embodiment of the present application, based on the idea of ​​transfer learning, the pre-trained first feature extraction module is used as the initial point in the process of developing a model for task learning in the second domain. In this way, the feature representation of the object generated based on the pre-trained first feature extraction module can be used as the actual tag for supervised learning to optimize the interest alignment model.

[0041] The cross-domain alignment module, also referred to as the cross-domain interest alignment module or the inter-domain alignment module, may be used to align the interest relationship between two objects in a second domain (which may be referred to as the second interest similarity degree in the present embodiment) with the interest relationship between two corresponding objects in a first domain (which may be referred to as the first interest similarity degree in the present embodiment). That is, the present embodiment designs the inter-domain alignment module to align the interest similarity of the objects in the second domain with the interest similarity in the first domain. In this way, the interest relationship of the objects in the first domain can be transferred to the second domain, thereby guiding the interest learning process in the second domain.

[0042] The intra-domain alignment module, also referred to as the intra-domain interest alignment module, may be used to interest align an object graph (images containing only objects) in a first domain with an object resource graph (including objects and resource data of interest to the objects) in a first domain, and interest align an object graph in a second domain with an object resource graph in a second domain. That is, the present embodiment designs the intra-domain alignment module to align interest relationships between objects in a single domain, making the feature representations between objects with similar interests or hobbies in a single domain closer, so that when using the interest alignment model subsequently, even if an object has no object interaction data or very little object interaction data, an accurate feature representation for the object can be generated based on the accurate feature representations of the other objects through analyzing the interest relationships between the object and other objects that have object interaction data.

[0043] In the model training solution provided by the embodiments of the present application, the object interest alignment may include inter-domain interest alignment and intra-domain interest alignment, where inter-domain interest alignment is the alignment of the interest similarity between two objects in a second domain with the similarity between the same two objects in a first domain, and intra-domain interest alignment is the alignment of the predicted interest similarity between two objects for a single domain with the actual interest similarity between the two objects. In practical applications, the model training solution may include at least one of the two types of interest alignment given above. For ease of description, the following embodiments will be described, and will be particularly described herein, using the example in which the model training solution simultaneously includes inter-domain interest alignment and intra-domain interest alignment.

[0044] Based on the brief introduction to the model structure of the interest alignment model in combination with Figure 2 above, the following will introduce in more detail the method steps of the model training method proposed in the embodiments of the present application in combination with other drawings, specifically the model training process and model application process included in the model training method.

[0045] A specific training flow for training an interest alignment model according to the transfer learning learning method may be seen in Figure 3, which shows a flow diagram of a model training method provided by an exemplary embodiment of the present application. The model training method may be performed by the electronic device mentioned above. The method may include, but is not limited to, steps S301 to S304.

[0046] S301: A data set is obtained.

[0047] The data set is a sample set used to train the interest alignment model, and may include first domain data of a first domain and second domain data of a second domain. The first domain may refer to a source domain containing abundant object interaction data, and the second domain is a target domain to be learned containing relatively little object interaction data. The first domain data of the first domain includes at least one object and first resource data of interest for each object within the first domain, and at least one object included in the first domain data is an object registered or temporarily logged in in a resource recommendation platform corresponding to the first domain. The first resource data of interest in the first field of each of the at least one object refers to a portion of the resource data that the object triggers in the first field, and the triggering of the first resource data by an object described herein may include, but is not limited to, the object clicking the first resource data, the object commenting (or forwarding, liking, etc.) on the first resource data, and the object viewing the first resource data for a duration exceeding a time threshold (e.g., 10 seconds), etc. Similarly, the second field data of the second field includes at least one object and the second resource data of interest in the second field of each object, and at least one object included in the second field data is the same as at least one object included in the first field data mentioned above, i.e., the overlapping data included in the first field data and the second field data is an object. In this way, even if there is relatively little object interaction data in the second field of the object, the abundant object interaction data in the first field of the object can be used to guide supervised learning in the second field. Regarding the related content of the second resource data in which the object is interested, reference may be made to the related description above regarding the first field, which will not be described in detail here.

[0048] The first and second domain data in the data set may exist in the form of a graph, where the term "graph" refers to network structure data consisting of nodes and connecting edges (or simply "edges"). The nodes in this graph may include objects, first resource data, and second resource data, and the connecting edges refer to the connecting lines between the nodes, specifically, the connecting lines connecting the object nodes and the resource data nodes of which the objects are interested. Considering that the related data in the data set includes multiple types, i.e., objects, first resource data, and second resource data, i.e., the sum of the node types and the connecting edge types is greater than two, the graph constructed in the embodiment of the present application is a heterogeneous graph (also called a heterogeneous network). In the heterogeneous graph, the types of the nodes and the connecting edges are not uniform but diversified.

[0049] A schematic diagram showing that the first domain data and the second neighborhood data in the data set exist in the form of a heterogeneous graph may refer to FIG. 2 given above. The data set shown in FIG. 2 includes three types of nodes: triangle, circle, and square. The triangle represents the first resource data node in the first domain data, the circle represents the object node (e.g., u1 node corresponds to object 1, u2 node corresponds to object 2, and u3 node corresponds to object 3), and the square represents the second resource data node in the second domain data. A connection edge between a triangle node and a circle node indicates that the object corresponding to the circle node is interested in the first resource data corresponding to the triangle node connected by the connection edge. Similarly, a connection edge between a circle node and a square node indicates that the object corresponding to the circle node is interested in the second resource data corresponding to the square node connected by the connection edge.

[0050] In order to test the online effect of a trained interest alignment model (i.e., a cross-domain recommendation algorithm based on object interest alignment) (if resource recommendations are made for objects under a cold start scenario, the recommendation effect of the cold start object will be mainly observed during online testing), an embodiment of the present application supports implementing the interest alignment model in the recall module of a resource recommendation platform. In this way, the streaming updated object interaction data can be used as sample data to train the interest alignment model. In other words, an embodiment of the present application obtains the streaming updated object interaction data in the resource recommendation platform as a data set for model training, i.e., the first domain data and the second domain data included in the data set are dynamically updated, so that when the interest alignment model is trained according to the dynamically updated data set, the interest alignment model also changes dynamically. This enables the interest alignment model to adapt to changes in the interests and hobbies of objects in the resource recommendation platform, thereby ensuring that the interest alignment model can always generate accurate feature representations for objects in the resource recommendation platform and meet the dynamically changing personalized resource needs of the objects.

[0051] S302: Invoke an interest alignment model to perform feature extraction on the first domain data to obtain a first domain feature representation, and then invoke an interest alignment model to perform feature extraction on the second domain data to obtain a second domain feature representation.

[0052] After obtaining a dataset used to train an interest alignment model, an embodiment of the present application supports using the dataset to perform single-domain feature representation, that is, performing feature extraction (also called feature representation) on the first domain data and the second domain data based on the dataset, thereby obtaining a first domain feature representation for the first domain and a second domain feature representation for the second domain.

[0053] The domain feature representation of any one domain may be realized through an embedding vector, and therefore the feature representation is also called an embedding representation. The embedding vector represents the features or attributes of an object using a low-dimensional vector. The distance between the embedding vectors of any two objects may characterize the similarity between the two objects. For example, if the distance between the embedding vectors of the two objects is smaller than a distance threshold, the two objects are relatively similar (e.g., their features or attributes are similar); if the distance between the embedding vectors of the two objects is greater than or equal to the distance threshold, the similarity between the two objects is relatively low. In an embodiment of the present application, the domain feature of any one domain may include feature representations of objects in the domain and feature representations of resource data. For example, the first domain features of a first domain may include a feature representation of an object and a feature representation of first resource data, where the feature representation of the object may be used to characterize a feature or attribute of the object (e.g., an interest or hobby within the first domain of the object), and the feature representation of the first resource data may be used to characterize a feature or attribute of the resource data (e.g., a resource type of the resource data). The distance between the feature representations of any two objects may be used to characterize interest similarity between the two objects.

[0054] The process of performing feature extraction on single-domain domain data in the embodiments of the present application is similar. That is, the process of performing feature extraction on first domain data to obtain a first domain feature representation is similar to the process of performing feature extraction on second domain data to obtain a second domain feature representation. Below, we will provide an outline of the feature extraction process for the first domain feature representation and the second domain feature representation, respectively, and then take the second domain feature representation as an example to provide a detailed description of the specific implementation process of information feature extraction on the second domain data. The process of performing feature representation on the first field data to obtain the first field feature representation may include: constructing an object resource graph for the first field based on the first field data, where the object resource graph for the first field is a graph with each object and each first resource data as nodes and a first interest relationship between the object and the first resource data as connecting edges, where having a first interest relationship between an object and first resource data may refer to the object generating interaction data with the first resource data; and invoking an interest alignment model to perform a graph encoding process on the object resource graph for the first field to obtain a first object feature representation for each object and a first resource feature representation for each first resource data, and constructing a first field feature representation using the first object feature representation for each object and the first resource feature representation for each first resource data.

[0055] The process of performing feature expression on the second field data and obtaining the second field feature expression may include, but is not limited to, steps (1) and (2).

[0056] (1) Constructing an object resource graph for the second domain based on data from the second domain. The object resource graph for the second domain is a graph in which each object and each second resource data is a node and a first interest relationship between the object and the second resource data is a connection edge. The object resource graph for the second domain here may refer to a bigraph (also called a bipartite graph, which is a special type of heterogeneous graph) for the second domain. As described above, the dataset may be represented in the form of a heterogeneous graph, and nodes having a connection edge with any one node in the heterogeneous graph may be called neighbor nodes of the node. The neighbor nodes of the u1 node shown in FIG. 2 include two first resource data and two second resource data having a connection edge with the u1 node. Thus, the process of constructing the object resource graph for the second domain based on data from the second domain here may include obtaining an object resource graph for the first domain by performing neighbor sampling on the heterogeneous graph of the dataset. Here, neighbor sampling may include selecting a portion of neighbor nodes from all neighbor nodes (in this case, the first resource data node and the second resource data node) of the object node in the heterogeneous graph of the data set. Thus, the embodiment of the present application performs neighbor sampling for each object node in the heterogeneous graph of the data set, thereby supporting constructing an object resource graph obtained in the second field based on each object node and each second resource data obtained by sampling.

[0057] As shown in FIG. 4, the data set includes object node u1, object node u2, and object node u3, and further includes neighboring nodes of object node u1 including first resource data node j1, first resource data node j2, second resource data node i1, and second resource data node i2, neighboring nodes of object node u2 including first resource data node j2, first resource data node j3, and second resource data node i1, and neighboring nodes of object node u3 including first resource data node j3, first resource data node j4, second resource data node i1, and second resource data node i2. Thus, after neighbor sampling is performed on the heterogeneous graph of the dataset based on the second domain data, the neighbor nodes of object node u1 that can be obtained by sampling are second resource data node i1 and second resource data node i2, the neighbor node of object node u2 is second resource data node i1, and the neighbor nodes of object node u3 are second resource data node i1 and second resource data node i2. In this case, an object resource graph of the second domain can be constructed based on each object node and the second resource data that each object node is interested in. Here, a connection edge between an object node and a second resource data node in the object resource graph indicates an interaction behavior between the object corresponding to the object node and the second resource data corresponding to the second resource data node.

[0058] (2) Invoke an interest alignment model to perform graph encoding on the object resource graph of the second domain to obtain a second object feature representation for each object and a second resource feature representation for each second resource data. Specifically, the present embodiment uses a graph encoder to capture rich semantic information of the object resource graph and adds a node-level attention mechanism to support distinguishing the importance of each neighbor node. A schematic diagram of an exemplary graph encoding learning architecture (or the architecture of the second feature extraction module) may be seen in FIG. 5. As shown in FIG. 5, the second feature extraction module includes a graph attention network (GAT). Through the attention mechanism included in the graph attention network (GAT), feature representation can be generated for each attention node in the input object resource graph of the second domain. Here, the attention node may refer to any node in the object resource graph, such as any object node or any second resource data node.

[0059] After inputting the object resource graph of the second domain into the second feature extraction module, the graph attention network GAT is invoked to distinguish the importance of each attention node's neighbor node, with each object node and each second resource data node as an attention node, thereby obtaining a feature representation of each attention node. When the attention node is an object node, the neighbor node of the object node refers to the second resource data node having a connecting edge between the object node and the object node. In other words, distinguishing the importance of each neighbor node may refer to determining the degree of preference (or degree of interest or hobby) of the object corresponding to the object node for each second resource data. When the attention node is a second resource data node, the neighbor node refers to the object node having a connecting edge between the second resource data node and the object node. In other words, distinguishing the importance of each neighbor node may refer to determining the degree of preference of the second resource data corresponding to the second resource data node by the object corresponding to each object node.

[0060] The specific implementation process of invoking the graph attention network to determine the feature representation of each attention node is similar. In the embodiment of the present application, one of at least one object included in the second domain data is represented as a training object, and the specific implementation process of invoking the graph attention mechanism to distinguish the importance of each neighbor node of the training object, thereby obtaining the corresponding second object feature representation of the training object, which may include but is not limited to steps s11 to s13.

[0061] s11: Obtain an initial feature representation of the training object in the second field and an initial feature representation of each second resource data of interest within the second field of the training object, where the initial feature representation of the training object is determined based on attribute information of the training object (e.g., basic object information (e.g., information such as age, gender, or selected category tag)), and the initial feature representation of the second resource data is determined based on attribute information of the second resource data (e.g., information such as upload time of the second resource data, category tag selected at the time of upload).

[0062] s12: Call an interest alignment model to calculate a degree of association between the training object and each piece of second resource data of interest of the training object according to the initial feature representation of the training object and the initial feature representation of each piece of second resource data of interest of the training object. Here, the degree of association between the training object and any one piece of second resource data of interest of the training object may be used to characterize the training object's interest to the one piece of second resource data, and the degree of association may be expressed in a probability format, for example, a degree of association of 20% indicates that the training object has a relatively low degree of interest to the second training resource data. Here, the formula for calculating the degree of association between the training object and any one piece of second resource data of interest of the training object is as follows:

number

[0063] where a ui indicates the degree of association (i.e., relevance) between object node u and second resource data node i, and h u denotes the initial feature representation of the training object corresponding to object node u, and h i denotes the initial feature representation of the second resource data corresponding to the second resource data node i, and N udenotes the neighbor set of object node u, which includes all second resource data nodes that have connecting edges with object node u; k denotes any one second resource data node in the neighbor set of object node u, i.e., second resource data node i is any one in the neighbor set of object node u; and LeakReLU function is the activation function.

[0064] As can be seen from the above formula (1), the embodiment of the present application calculates the relevance between the training object and each neighbor second resource data of the training object, for example, the similarity a between the object node u to which the training object corresponds and the second resource data node i to which the second resource data corresponds. ui Specifically, the relationship between the object node u and the second resource data node i is used as the numerator, and the relationship between the object node u and all neighbor second resource data nodes N u In this way, using the relevance for the second resource data node i as the numerator and the sum of the relevances of all neighbor nodes as the denominator is advantageous for quickly determining the proportion of the second resource data corresponding to the second resource data node i among all the second resource data of interest to the training object, and thereby determining the importance of the second resource data to the training object.

[0065] s13: obtain a first object feature representation of the training object based on the association degree between the training object and each second resource data of the training object's interest, the initial feature representation of the training object in the second field, and the initial feature representation of each second resource data of the training object's interest. That is, after obtaining the association degree between the training object and each neighbor second resource data based on step s12, the training object's preference for each second resource data can be roughly determined. To obtain the first object feature representation of the training object, i.e., the training object's overall interest preference, it is necessary to further specify the first object feature representation of the training object based on the association degree between the training object and each neighbor second resource data and the initial feature representation of the second resource data and the training object itself, thereby obtaining the training object's overall interest preference (e.g., preference for a certain type of resource data). Here, the calculation formula used to determine the second object feature representation of the training object is as follows:

number

[0066] where:

number

[0067] As described above, through the specific implementation process shown in steps s11 to s13 described above, the graph neural network GAT is invoked to capture the rich semantic information in the object resource graph, so that each object node can determine the second object feature representation of the corresponding object, and the object node can use it to describe the interests and preferences of the corresponding object.

[0068] The specific implementation process of invoking the graph attention network GAT to obtain the second resource feature representation of the second resource data corresponding to each second resource data node in the object resource graph of the second domain is similar to the specific implementation process of obtaining the first object feature representation described above, except that when calculating the second resource feature representation of the second resource data, the neighbor set is each object node adjacent to the second resource data node in the object resource graph of the second resource data, and the specific implementation process of determining the second resource feature representation of the second resource data will not be described in detail here. Similarly, the specific implementation process of performing feature extraction on the first domain data based on the first feature extraction model in the interest alignment model (specifically, the graph attention network GAT included in the first feature extraction model) to obtain the first domain feature representation (including the first object feature representation and the first resource feature representation) is similar to the specific implementation process of obtaining the first object feature representation given above, and will not be described in detail here.

[0069] The first feature extraction module corresponding to the first domain is pre-trained, i.e., the first feature extraction module has already been optimized using abundant object interaction data in the first domain during the pre-training stage. Therefore, during the process of training the interest alignment model, there is no need to optimize the first feature extraction module according to the first domain feature representation of the predicted first domain data. However, because object interaction data in the second domain is relatively scarce, the trained first feature extraction module is transferred during the model training process of the interest alignment model to help better optimize the second feature extraction module.

[0070] Based on this, after obtaining the feature representations of each object node and second resource data node in the second domain based on the above steps, it is also necessary to learn the object's affinity for the second resource data. Specifically, the affinity of each object for the second resource data of interest is predicted, and an interest alignment model is trained based on the difference between the predicted object's affinity for the second resource data of interest and the corresponding object's actual affinity for the corresponding second resource data of interest. Specifically, the second feature extraction model is optimized to improve the feature extraction performance of the second feature extraction model, i.e., the optimized second feature extraction model can predict more accurate object feature representations for objects.

[0071] First, the embodiment of the present application supports fusing a first object feature representation in a first domain of each object with a second object feature representation in a second domain to obtain a fused feature representation of each object, thereby realizing the object's interest features in the first domain and the interest features in the second neighborhood to be combined together to form a more complementary interest representation of the object. For example, the first object feature representation in the first domain of object node u is:

number

number

number

number

number

[0072] Then, the fusion feature representation of each object and the second resource feature representation of the second resource data of interest of the corresponding object are connected and operated to obtain the predicted attention (i.e., predicted preference) of each object for the second resource data of interest of the corresponding object, where the calculation formula for obtaining the predicted attention (i.e., predicted preference) of each object for the second resource data of interest of the corresponding object is as follows:

number

[0073] where:

number

number

number

number

[0074] Finally, the data set includes the actual interest level of each object for the corresponding second resource data of interest, thus the actual interest level of the object for the second resource data of interest can be obtained from the data set, and a first loss of the interest alignment model is constructed based on the difference between the actual interest level and the predicted interest level for each object for the corresponding second resource data of interest, thereby facilitating subsequent training of the interest alignment model based on the first loss. The embodiment of the present application supports learning the interest preferences of objects for the second resource data of interest through minimizing squared error, and the formula for minimizing squared error is as follows:

number

[0075] where R denotes the object-second resource data interaction matrix,

number

[0076] As described above, the first feature extraction module corresponding to the first domain is pre-trained. In the pre-training stage, when the first feature extraction module is pre-trained, the process of determining the predicted attention and loss for the first resource data of the object is similar to the model training process in the second domain described above, and will not be described in detail here. For ease of distinction, the embodiment of the present application uses the first loss of the second domain as

number

[0077] S303: Call the interest alignment model and perform the following process: interest alignment processing between the first field and the second field based on the first field feature expression and the second field feature expression.

[0078] As can be seen from the above description, the embodiments of the present application support inter-domain object interest similarity alignment and / or intra-domain object interest similarity alignment. Specific implementation methods for these two types of interest similarity alignment are introduced below.

[0079] (1) Inter-domain interest similarity alignment: In the present embodiment, a cross-domain alignment module (also referred to as an inter-domain alignment module) is designed based on the phenomenon that interests or hobbies of the same objects across domains have similarities, i.e., two objects with similar interests in a first domain also exhibit similar interests in a second domain. Specifically, considering that the first domain has abundant object interaction data and the first feature extraction model in the first domain is pre-trained (i.e., already has relatively good feature representation performance), a cross-domain alignment module is adopted to realize inter-domain object interest similarity alignment. The so-called inter-domain object interest similarity alignment refers to using the interest similarity of two objects in the first domain as actual tags to align the interest similarity of the two objects in the second domain with the interest similarity in the first domain, thereby guiding the learning of interest similarity in the second domain. That is, the embodiment of the present application fully considers the interest relationships between objects to transfer the object interest similarity in the first domain to the second domain, i.e., aligns the object interest similarity in the second domain with the first domain source domain, thereby achieving the purpose of guiding interest learning in the second domain.

[0080] In a specific implementation, as can be seen from the above steps, the first domain feature representation includes a first object feature representation of each object in the first domain data, and the second domain feature representation includes a second object feature representation of each object in the second domain data. Hereinafter, in combination with FIG. 6, a specific implementation process of calling an interest alignment model to perform interest alignment processing between the first domain and the second domain based on the first domain feature representation and the second domain feature representation will be introduced, where the interest alignment processing includes a cross-domain alignment processing, and the cross-domain alignment processing includes, but is not limited to, steps s21 to s22.

[0081] s21: Call an interest alignment model (specifically, a cross-domain alignment module) to determine an attention similarity between two objects in a single domain based on the object feature representation of each object in the single domain. Specifically, call the interest alignment model to determine a first attention similarity degree between two objects in a first domain based on the first object feature representation of each object, and call the interest alignment model to determine a second attention similarity degree between two objects in a second domain based on the second object feature representation of each object.

[0082] Considering the difference in the degree of personalization of a single domain, the difference in object interest similarity between different domains is relatively large. For example, the degree of personalization of some single domains (e.g., the book domain or the movie domain) is relatively high (i.e., the difference in resource types of resource data within the single domain is relatively large), resulting in a low object interest similarity (i.e., the interests and hobbies of different objects in the same resource data). Furthermore, for example, the degree of personalization of some single domains (e.g., the news domain) is relatively low (i.e., the difference in resource types of resource data within the single domain is relatively small), resulting in a higher object interest similarity (i.e., the interests and hobbies of different objects in the same resource data). Therefore, directly aligning the object interest similarity values ​​is affected by the nature of the domains and thus cannot reflect the actual interest similarity between objects. Based on this, the embodiment of the present application supports eliminating the influence of domain characteristics by converting the distance between the object feature representations of objects into a probability format to represent the interest similarity between objects.

[0083] Here, the specific implementation process for determining the degree of first attention similarity between two objects in the first and second domains is similar. Below, we will take determining the degree of first attention similarity between two objects in the first domain as an example to introduce the specific implementation process for calling the interest alignment model and determining the degree of first attention similarity between two objects in the first domain based on the first object feature representation of each object.

[0084] First, a distance calculation is performed on the first object feature representations of two objects among the plurality of objects to obtain distance information between the first object feature representations of two objects. Specifically, the distance calculation here is a Euclidean distance calculation, that is, the first object feature representations are in vector form, and the vector distance between two vectors is calculated to obtain distance information between the two vectors. As shown in FIG. 6, at least one object includes object u1, object u2, and object u3, and the first object feature representation of object u1 is:

number

number

number

number

number

number

number

number

number

number

number

number

[0085] Then, the distance information between the first object feature representations of each pair of objects is converted into a probability to obtain a first interest similarity between each pair of objects. In other words, converting the Euclidean distance between the object feature representations of the objects into a probability format to represent the interest similarity can avoid the problem that the alignment results are affected by the degree of domain personalization and cannot actually reflect the actual interest similarity between the objects. Here, the formula for converting the Euclidean distance into a probability format to represent the interest similarity is as follows:

number

[0086] where:

number

number

number

[0087] s22: Based on the specific implementation process shown in step s21 described above, a first attention similarity degree between each pair of objects in a first domain can be obtained, and the first attention similarity degree between any two objects can be used to characterize the similarity of interests or hobbies in the first domain between the two objects. Based on the first attention similarity degree between each object in the first domain, an attention similarity distribution P in the first domain can be determined, and the attention similarity distribution P is a probability distribution composed of the first attention similarities between each pair of objects in the first domain. Similarly, based on the specific implementation process shown in step s21 described above, a second attention similarity degree between each pair of objects in a second domain can be obtained, and the second attention similarity degree between any two objects can be used to characterize the similarity of interests or hobbies in the second domain between the two objects. Based on the second attention similarity degree between each object in the second field, an attention similarity distribution Q of the second field can be determined, and the attention similarity distribution Q is a probability distribution composed of the second attention similarity degrees between every two objects in the second field.

[0088] In an embodiment of the present application, the learning target for the cross-domain alignment module is to minimize the difference (or disparity) between the attention similarity distribution P of the first domain and the attention similarity distribution Q of the second domain. In this embodiment, KL divergence is used as the loss function, and a cross-domain alignment loss of the interest alignment model is constructed based on the difference between the attention similarity distribution P of the first domain and the attention similarity distribution Q of the second domain. The cross-domain alignment loss is used as the second loss, thereby supporting optimization of the interest alignment model through minimization of the cross-domain alignment loss. Here, KL divergence, also known as relative entropy, is a measure of the asymmetry of the difference between two probability distributions. In this embodiment, KL divergence is used as the loss function, and the calculation formula for obtaining the cross-domain alignment loss of the interest alignment model is as follows:

number

[0089] where P denotes the attention similarity distribution of the first field, and p uiuj denotes the first attention similarity degree between object ui and object uj in the first domain, Q denotes the attention similarity distribution in the second domain, and q uiuj denotes the second degree of similarity of interest between object ui and object uj in the second domain.

[0090] As described above, the cross-domain alignment loss of the interest alignment model can be constructed and obtained based on the specific implementation process shown in the above steps s21 to s22.

[0091] (2) Intra-domain interest similarity alignment: The embodiment of the present application mainly uses graph decoding targets to align the interest similarity of objects within a domain, so as to make the feature representations of objects with similar interests and hobbies within a single domain closer. As a result, when the model is applied, even if the object of the resource data to be distributed is a cold-start object, the accuracy of the object feature representation of the cold-start object can be ensured by analyzing the object feature representations of other objects with similar interests and hobbies to the cold-start object.

[0092] Considering that the domain data (in the form of a bigraph) input to the single-domain feature extraction module includes the interest similarity relationship between each object within a single domain, as shown in FIG. 5, object node u1 and object node u2 are both connected to second resource data i1, thus determining that the object corresponding to object node u1 and the object corresponding to object node u2 have similar interests or hobbies. To make the feature representations of objects with similar interests or hobbies within a single domain more similar, the present embodiment supports aligning intra-domain object interest similarities. The so-called intra-domain object interest similarity alignment refers to using the input information (single-domain domain data) of the single-domain feature extraction module as actual tags. Specifically, it refers to aligning the interest similarity between two objects in the domain data station input by the single-domain feature extraction module with the actual tags, and aligning the predicted interest similarity between the same two objects within a single domain with the actual tags. That is, aligning the interests of intra-domain objects to make the object feature representations of objects with similar interests or hobbies within a domain more similar. Based on this, the embodiments of the present application hope to be able to reconstruct an interest relationship graph between objects based on the fused feature representation of the objects, which respectively includes an interest relationship graph between objects in a first domain and an interest relationship graph between objects in a second domain, and the interest relationship graph is an object-object graph containing only object nodes, so that intra-domain alignment learning can be performed based on the reconstructed object-object graph and the object resource graph input into the model.

[0093] Below, in combination with Figure 7, we will introduce the specific implementation process of calling the interest alignment model and performing interest alignment processing in the first field and the second field based on the first field feature representation and the second field feature representation, respectively. At this time, the interest alignment processing includes intra-domain alignment processing, and the intra-domain alignment processing includes, but is not limited to, steps s31 to s33.

[0094] s31: Obtain a fused feature representation of each object, an object resource graph of a first domain, and an object resource graph of a second domain. Here, the fused feature representation of any one object is obtained by fusing the first object feature representation of the object in the first domain with the second object feature representation of the object in the second domain. By fusing the first object feature representation and the second object feature representation of the object, a complementary feature representation of the object can be formed, thereby reducing the problem of relatively little object interaction data in the second domain. The object resource graph of the first domain and the object resource graph of the second domain are input information for the interest alignment model. The related content of the object resource graph may refer to the above related description and will not be described in detail here.

[0095] s32: Call the interest alignment model (specifically, the intra-domain alignment module included in the interest alignment model) to perform graph decoding processing for the first domain and graph decoding processing for the second domain on the fused feature representation of each object, respectively, to obtain an object graph for the first domain and an object graph for the second domain. The object graph here is the object-object graph described above, where each object is a node and the second interest relationship between objects is a connecting edge. Having a so-called second interest relationship between objects may refer to two objects generating interaction data for the same resource data.

[0096] For example, to reconstruct an object graph of a first domain and an object graph of a second domain, the graph decoding weight of the first domain is obtained, and an interest alignment model is invoked to reconstruct the object graph of the first domain based on the graph decoding weight of the first domain and the fusion feature representation of each object; and the graph decoding weight of the second domain is obtained, and an interest alignment model is invoked to reconstruct the object graph of a second neighbor based on the graph decoding weight of the second domain and the fusion feature representation of each object. That is, an embodiment of the present application employs a graph decoder to reconstruct an object graph of a single domain, thereby supporting, through the object graph, prediction of whether two objects will generate interaction data for the same resource data within a single domain. If two objects generate interaction data for the same resource data, there will be a connecting edge between the two objects in the object graph. In this case, the two objects are deemed to have similar interests or hobbies, and it is hoped that the object feature representations of the two objects will be closer (i.e., the distance between the embedding vectors will be closer).

[0097] The graph decoder corresponding to each domain is different, specifically, the graph decoding weights of the graph decoder are different.

number

number

[0098] where:

number

number

number

number

number

number

[0099] As shown in Figure 8, after the fused feature representation of the object is input to the graph decoder corresponding to the first domain, the object graph G Scan be obtained by reconstructing the object resource graph G in the first domain, and as can be seen by comparing the object resource graph G in the first domain with the object graph G in the first domain, the second-order relationships of the objects in the first domain are transformed into first-order relationships (i.e., the meta-path: object-first resource data-object in the object resource graph is transformed into the meta-path: object-object in the object graph. A meta-path may refer to a specific path connecting two or more nodes, and may be used to describe a composite relationship between nodes, and different meta-paths have different explicit definitions). This can achieve the description of interest similarity between two objects in the first domain. Similarly, after the feature fusion representation of the objects is input into the corresponding graph decoder in the second domain, the object graph G in the second domain is transformed into T can be reconstructed to obtain the object graph G T The related content is in the object graph G S Please refer to the related description for the same, and it will not be described in detail here.

[0100] s33: Align the object graph of the first domain with the object resource graph of the first domain, and align the object graph of the second domain with the object resource graph of the second domain, thereby realizing intra-domain alignment. Specifically, after reconstructing the object graph of the first domain and the object graph of the second domain based on the above steps, the object graph is used as the predicted object-object relationship, and the object resource graph of the corresponding domain is used as the actual object-object relationship, thereby supporting training of the interest alignment model according to the difference between the actual object-object relationship and the predicted object-object relationship.

[0101] For example, the intra-domain alignment loss for the first domain is obtained based on the difference between the object graph of the first domain and the corresponding object resource graph, and the intra-domain alignment loss for the second domain is obtained based on the difference between the object graph of the second domain and the corresponding object resource graph. Then, considering the existence of object graphs of multiple domains, the intra-domain alignment loss of the interest alignment model is obtained by merging the intra-domain alignment loss of the first domain with the intra-domain alignment loss of the second domain, and the intra-domain alignment loss is used as the second loss. As described above, graph reconstruction is a binary classification task, and thus it is supported to adopt the binary classification cross-entropy loss as the graph reconstruction loss. The intra-domain alignment loss of the interest alignment model constructed in the multi-domain (i.e., first and second domain) case is as follows:

number

[0102] where:

number

number

number

number

number

[0103] where:

number

number

number

number

number

number

number

number

number

number

number

number

[0104] As described above, the intra-domain alignment loss of the alignment model of interest can be constructed and obtained based on the specific implementation process shown in the above steps s31 to s33.

[0105] It is worth noting that, in consideration of the relatively large number of objects and resource data contained in the data set (or the first and second domain data), in order to reduce computational complexity, the embodiments of the present application support dividing the domain data of a single domain into multiple data subsets (batches), each of which contains some objects and resource data. In this way, an object graph of some object nodes can be constructed, rather than constructing an object graph based on all objects, thereby reducing computational complexity and improving the efficiency and speed of model training. For example, assuming that one data subset batch contains n objects, n≦N, where N is all objects, the computational complexity is O(N 2 )from

number

[0106] In some embodiments, the interest alignment process includes a cross-domain alignment process and an intra-domain alignment process, and in step 303, calling the interest alignment model to perform the following process, i.e., performing the interest alignment process between the first domain and the second domain based on the first domain feature representation and the second domain feature representation, can be realized through the following technical solutions: calling the interest alignment model to perform the following process, i.e., performing the cross-domain alignment process between the first domain and the second domain based on the first domain feature representation and the second domain feature representation; calling the interest alignment model to perform the following process, i.e., performing the intra-domain alignment process between the first domain and the second domain based on the first domain feature representation and the second domain feature representation.

[0107] Here, the implementation method of performing cross-domain alignment processing between the first field and the second field based on the first field feature representation and the second field feature representation may refer to the implementation method of steps s21 to s22, and here, the implementation method of performing intra-domain alignment processing between the first field and the second field based on the first field feature representation and the second field feature representation may refer to the implementation method of steps s31 to s33.

[0108] In some embodiments, a cross-domain alignment loss corresponding to the cross-domain alignment process is obtained, and an intra-domain alignment loss corresponding to the intra-domain alignment process is obtained, and the cross-domain alignment loss and the intra-domain alignment loss constitute the second loss.

[0109] S304: Training the interest alignment model according to the direction of reducing the first loss in the feature extraction process and the loss in the interest alignment process.

[0110] As can be seen from the above description, the overall learning goal of the model training method provided by the embodiment of the present application can be divided into three parts: a) learning a recommendation target in a second domain, i.e., learning the interests and preferences of an object in the second domain according to the object interaction data between the object and the second resource data. A first loss of the interest alignment model can be constructed and obtained under this target learning. b) learning a cross-domain interest alignment target, i.e., transferring the object interest similarity in the first domain, which has abundant object interaction data, to the second domain, which has scarce object interaction data, and aligning the object interest similarity in the second domain with the first domain, thereby achieving the goal of guiding interest learning in the second domain. A cross-domain alignment loss of the interest alignment model can be constructed and obtained under this learning goal. c) learning an intra-domain interest alignment target, i.e., using a graph decoding target to align the interests and preferences of objects in a domain, aiming to make the object feature representations of objects with similar interests in the domain closer. The intra-domain alignment loss of the interest alignment model can be constructed and obtained under this target learning.

[0111] Here, the loss in the feature extraction process may refer to the first loss of the interest alignment model mentioned above, while the loss in the interest alignment process may include at least one of cross-domain alignment loss and intra-domain alignment loss.It should be noted that, depending on the differences in the modules included in the interest alignment model, the process of training the interest alignment model will be different according to the direction of reducing the first loss in the feature extraction process and the loss in the interest alignment process.

[0112] The interest alignment model may include only a single-domain feature extraction module and a cross-domain alignment module. In this implementation, the step of training the interest alignment model according to the direction of reducing the first loss of the feature extraction process and the loss of the interest alignment process may include jointly training the interest alignment model using the first loss of the interest alignment model and the cross-domain alignment loss of the interest alignment model obtained in the above steps. More specifically, the first loss of the interest alignment model and the cross-domain alignment loss of the interest alignment model are added to obtain a target loss of the interest alignment model, and the interest alignment model is trained according to the direction of reducing the target loss. Here, the target loss is calculated as follows:

number

[0113] L denotes the overall loss of the interest alignment model,

number

number

[0114] The interest alignment model may only include a single-domain feature extraction module and an intra-domain alignment module. In this implementation, the step of training the interest alignment model according to the direction of reducing the first loss of the feature extraction process and the loss of the interest alignment processing process may include jointly training the interest alignment model using the first loss of the interest alignment model obtained according to the above steps and the intra-domain alignment loss of the interest alignment model. More specifically, the first loss of the interest alignment model and the intra-domain alignment loss of the interest alignment model are added to obtain the target loss of the interest alignment model, and the interest alignment model is trained according to the direction of reducing the target loss. Here, the target loss is calculated as follows:

number

[0115]

number

[0116] The interest alignment model may include a single-domain feature extraction module, a cross-domain alignment module, and an intra-domain alignment module. In this implementation, the step of training the interest alignment model according to the direction of reducing the first loss of the feature extraction process and the loss of the interest alignment processing process may include jointly training the interest alignment model using the first loss of the interest alignment model, the cross-domain alignment loss of the interest alignment model, and the intra-domain alignment loss of the interest alignment model obtained according to the above steps. More specifically, the first loss of the interest alignment model, the cross-domain alignment loss, and the intra-domain alignment loss are added to obtain the target loss of the interest alignment model, and the interest alignment model is trained according to the direction of reducing the target loss. Here, the target loss is calculated as follows:

number

[0117] As can be seen from the above, the embodiment of the present application supports the use of a joint training method to train the interest alignment model, specifically, by adjusting the model parameters of each module included in the interest alignment model, so that the predictive performance of the interest alignment model after adjusting the model parameters is more accurate. It is worth noting that the first domain already has abundant object interaction data, while the second domain has relatively little object interaction data. Furthermore, the embodiment of the present application aims to use the object interaction data of the first domain to guide the learning of the second domain. Therefore, according to the first loss in the feature extraction process and the decreasing direction of the loss in the interest alignment processing process, in the process of training the interest alignment model, the attention similarity distribution P of the first domain is used as an anchor to support interrupting the gradient backpropagation for the first feature extraction module based on the cross-domain alignment loss, that is, there is no need to adjust the parameters of the first feature extraction module based on the cross-domain alignment loss. However, the embodiments of the present application still support fine-tuning the first domain feature extraction model based on the intra-domain alignment loss, allowing the first domain feature representation of the first domain to contribute to the first loss of the second domain.

[0118] In an embodiment of the present application, the interest alignment model is trained by fully considering the interest similarity between objects and adopting an object interest similarity alignment method, so that the interest alignment model can generate a relatively accurate feature representation for an object even when applied to a domain with relatively little interaction data. For example, considering that the interaction data in a first domain is relatively abundant, the feature representation of the first domain extracted based on the abundant interaction data is more accurate. Therefore, the interest similarity between two objects in a second domain with relatively little interaction data is aligned with the interest similarity of the two objects in the first domain with relatively much interaction data to train the interest alignment model. This ensures that the trained interest alignment model can generate a relatively accurate feature representation for an object even when the interaction data in the second domain is relatively sparse, so that the resource data recommended for the object based on the accurate feature representation can meet the needs of object personalization and achieve high-precision resource recommendation. In the embodiment of the present application, a method of jointly training the interest alignment model is adopted with the first loss in the feature extraction process and the loss in the interest alignment processing process, which can enrich the overall learning targets of the interest alignment model and ensure the performance of the interest alignment model through multi-target learning.

[0119] The embodiment shown in Figure 3 above mainly provides a detailed description of the model training portion of the interest alignment model. The following section will discuss the model application portion of the interest alignment model. In a specific implementation, after obtaining a trained interest alignment model based on the above steps, the embodiment of the present application supports implementing the trained interest alignment model in a resource recommendation platform, specifically, applying it to the resource recall stage of the resource recommendation platform. Therefore, the interest alignment model may also be referred to as a resource recall model. The resource recall stage mainly involves quickly screening a portion of resource data that is interesting to an object from a large resource database, thereby facilitating subsequent high-precision sorting of the portion of resource data for recommendation to the object. An exemplary flow diagram of applying the interest alignment model to model application (i.e., resource recommendation) may be seen in Figure 8. As shown in Figure 8, when a target object (e.g., any one object) has resource distribution needs, the object features (or attribute information, such as age, gender, etc.) of the target object are obtained, and an object resource graph can be constructed based on the interaction data of the target object and other objects in the resource recommendation platform. Then, in the resource recall stage, the trained interest alignment model in the embodiment of the present application is employed to generate a feature representation for the target object and a feature representation for resource data in the resource recommendation platform. The feature representation of the target object is matched with the feature representation of each resource data, specifically, the similarity between the feature representation of the target object and the resource data is calculated. Then, according to the similarity result, multiple resource data are selected and put into a precisely sorted candidate pool for use in the subsequent sorting stage.

[0120] Referring to FIG. 9, FIG. 9 shows a flow diagram of another model training method provided by one exemplary embodiment of the present application, which may be performed by the electronic device mentioned above, and the method may include, but is not limited to, steps S901 to S907.

[0121] S901: A data set is acquired.

[0122] S902: Invoke an interest alignment model to perform feature extraction on the first domain data to obtain a first domain feature representation, and then invoke an interest alignment model to perform feature extraction on the second domain data to obtain a second domain feature representation.

[0123] S903: An interest alignment model is called and interest alignment processing is performed in the first and second fields based on the first and second field feature expressions.

[0124] S904: Training the interest alignment model according to the direction of reducing the first loss in the feature extraction process and the loss in the interest alignment process.

[0125] It is necessary to explain that the specific implementation steps shown in steps S901 to S904 may refer to the relevant description of the specific implementation steps shown in steps S301 to S304 in the embodiment shown in Figure 3 above, and will not be described in detail here.

[0126] S905: Invoke the trained interest alignment model to perform feature extraction on the target object to be delivered, and obtain a fused feature representation of the target object.

[0127] After the resource recommendation platform is equipped with the trained interest alignment model, if there is a target object of resource data to be distributed in the resource recommendation platform, and the target object is an object that has just been registered in the resource recommendation platform, the object attributes of the target object of the resource data to be distributed can be obtained, where the object attributes may include related attribute information entered when the target object registers in the resource recommendation platform (e.g., set nickname, age, gender, selected resource type tag, etc.), and then the trained interest alignment model is invoked to perform feature extraction on the object attributes of the target object, thereby obtaining a fusion feature representation of the target object.

[0128] Specifically, a first feature extraction module included in the interest alignment model is called to predict a first object feature representation of the target object in a first region, and a second feature extraction module included in the interest alignment model is called to predict a second object feature representation of the target object in a second region, and then the first object feature representation and the second object feature representation of the target object are fused to obtain a fused feature representation of the target object. Here, the specific implementation process of the first feature extraction module predicting the first object feature representation of the target object (or the second feature extraction module predicting the second object feature representation of the target object) may refer to the relevant description in the embodiment shown in Figure 3 above, and will not be described in detail here.

[0129] S906: The similarity between the fused feature representation of the target object and the resource feature representation of each candidate resource data is compared.

[0130] S907: The candidate resource data whose similarity comparison result is greater than the comparison result threshold is determined as second resource data to be distributed that is associated with the target object.

[0131] In steps S906-S907, a fusion feature representation of the target object of the resource data to be distributed can be obtained based on the above steps, and the fusion feature representation can be used to characterize the attributes or features of the target object (for example, to characterize that the target object is interested in a certain type of resource data). In this way, a resource feature representation of candidate resource data to be distributed in the second field can be obtained, and the resource feature representation of any one of the candidate resource data can also be used to characterize the attributes or features of the candidate resource data (for example, to characterize trigger cases within the history time of the candidate resource data).

[0132] The candidate resource data whose resource feature representation is closer to the fused feature representation of the target object is considered to be more likely to be resource data of interest to the target object. Therefore, the similarity between the fused feature representation of the target object and the resource feature representation of each candidate resource data is matched to obtain a similarity comparison result between the target object and each candidate resource data. The purpose of similarity matching is to find candidate resource data that is close to the fused feature representation of the target object. Finally, candidate resource data whose similarity comparison result satisfies a resource recommendation rule is selected as resource data to be distributed associated with the target object. When the interest alignment model is implemented in the resource recall phase of the resource recommendation platform, candidate resource data whose similarity comparison result is greater than a comparison result threshold can be placed in a precisely sorted candidate pool as resource data to be distributed, thereby facilitating subsequent recommendation by precise sorting. Here, the resource recommendation rule may be customized by the operator according to business needs. For example, the resource recommendation rule may include selecting candidate resource data whose similarity comparison result is greater than a result threshold as resource data to be distributed associated with the target object. Further, for example, the resource recommendation rule may include sorting the result values ​​from the highest to the lowest, and determining the first k candidate resource data as resource data to be distributed associated with the target object, etc. The embodiments of the present application are not limited to the specific content of the resource recommendation rule.

[0133] FIG. 10 may be referred to for an exemplary schematic diagram of matching the fusion feature representation of the target object with the resource feature representation of each candidate resource data. As shown in FIG. 10, the candidate resource data to be distributed in the second category includes candidate resource data 1, candidate resource data 2, candidate resource data 3, and candidate resource data 4. The similarity comparison result between the fusion feature representation of the target object and the resource feature data of the candidate resource data 1 is 60% (or expressed in a numerical form such as a fraction or decimal), the similarity comparison result between the fusion feature representation of the target object and the resource feature data of the candidate resource data 1 is 50%, the similarity comparison result between the fusion feature representation of the target object and the resource feature data of the candidate resource data 1 is 20%, and the similarity comparison result between the fusion feature representation of the target object and the resource feature data of the candidate resource data 1 is 70%. If the resource recommendation rule includes sorting the result values ​​from largest to smallest and determining the first k (k=2) candidate resource data as resource data to be distributed and associated with the target object, a sorting among four candidate resource data can be obtained: candidate resource data 4 → candidate resource data 1 → candidate resource data 2 → candidate resource data 3. In this way, the first candidate resource data 4 and candidate resource data 1 in the sorting can be determined as resource data to be distributed and associated with the target object.

[0134] In practical applications, the trained interest alignment model of the present embodiment can be implemented in the recall module of a resource recommendation platform and used to recommend resources for millions of online objects in the resource recommendation platform. Here, the results of evaluating the model effectiveness of the interest alignment model using key indicators that determine whether the model can be accurately recommended in the industrial sector can be seen in Table 1.

[0135] [Table 1]

[0136] In Table 1, Uctr (User-click-through-rate) may be abbreviated as "user click-through rate," and this evaluation index is obtained by comparing the number of objects with click behaviors and the number of objects visited in the resource data. After resource data is recommended using the interest alignment model provided by the embodiment of the present application, if the number of objects visited in the resource data is 100 and 13 users have click behaviors on the resource data, the user click-through rate is 13 / 100*% = 13%. Here, UctrX indicates the increase in clicks on a video when the video playback time is X seconds under a video recommendation scene. As shown in Table 1, X may take values ​​of 180, 60, or 30 seconds. Pctr (Page-click-through-rate) may be referred to as "page click-through rate." dau (daily-active-users) may be abbreviated as "daily active users" and may be used to reflect the number of objects that are active daily in the resource recommendation platform. As described above, after being equipped with the trained interest alignment model of the embodiment of the present application, the resource recommendation platform can generate object feature representations for objects with higher accuracy, thereby recommending resource data for objects with higher accuracy and further promoting the popularity of the resource recommendation platform.

[0137] It is understood that the embodiments of the present application involve relevant data such as user information, and when the embodiments of the present application are applied to specific products or technologies, user permission or consent must be obtained, and further, the collection, use, and processing of relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0138] The above has discussed in detail the method of the present embodiment, and in order to better facilitate implementation of the above method of the present embodiment, the following provides the apparatus of the present embodiment accordingly.

[0139] 11 shows a structural diagram of a model training device provided by an exemplary embodiment of the present application, which may be a set of computer-executable instructions (including program code) running on an electronic device. The model training device may be used to perform some or all of the steps of the method embodiments shown in FIGS. 3 and 9, and the model training device includes an acquisition unit 1101 and a processing unit 1102. The acquisition unit 1101 is configured to acquire a data set, where the data set includes first domain data and second domain data, where the first domain data includes at least one object and first resource data of interest within the first domain for each object, and the second domain data includes at least one object and second resource data of interest within the second domain for each object. The processing unit 1102 is configured to call an interest alignment model to perform feature extraction on the first domain data to obtain a first domain feature representation, and to call an interest alignment model to perform feature extraction on the second neighbor data to obtain a second domain feature representation, the processing unit 1102 is further configured to call the interest alignment model to perform the following process, i.e., interest alignment processing between the first domain and the second domain based on the first domain feature representation and the second domain feature representation, the processing unit 1102 is further configured to train the interest alignment model according to a direction of reducing a first loss and a second loss to obtain a trained interest alignment model, where the first loss is a loss corresponding to the feature extraction and the second loss is a loss corresponding to the interest alignment processing, and the trained interest alignment model is used to recommend resource data for a target object in the second domain.

[0140] In some embodiments, the processing unit 1102 is further configured to construct an object resource graph for the first domain based on the first domain data, where the object resource graph for the first domain is a graph in which each object and each first resource data is a node and a first interest relationship between the object and the first resource data is a connecting edge. The interest alignment model is invoked to perform a graph encoding process on the object resource graph for the first domain to obtain a first object feature representation for each object and a first resource feature representation for each first resource data. The first object feature representation for each object and the first resource feature representation for each first resource data constitute the first domain feature representation.

[0141] In some embodiments, the processing unit 1102 is further configured to construct an object resource graph for the second domain based on the second domain data, where the object resource graph for the second domain is a graph in which each object and each second resource data is a node and a first interest relationship between the object and the second resource data is a connecting edge. The interest alignment model is invoked to perform a graph encoding process on the object resource graph for the second domain to obtain a second object feature representation for each object and a second resource feature representation for each second resource data. The second domain feature representation is composed of the second object feature representation for each object and the second resource feature representation for each second resource data.

[0142] In some embodiments, any one object of the at least one object is represented as a training object, and the processing unit 1102 is further configured to obtain an initial feature representation in a second field of the training object and an initial feature representation of each second resource data of interest of the training object, invoke an interest alignment model to calculate a degree of association between the training object and each second resource data of interest of the training object according to the initial feature representation of the training object and the initial feature representation of each second resource data of interest of the training object, and obtain a second object feature representation of the training object based on the degree of association between the training object and each second resource data of interest of the training object, the initial feature representation in the second field of the training object, and the initial feature representation of each second resource data of interest of the training object.

[0143] In some embodiments, the dataset includes an actual attention level for each object's corresponding second resource data of interest, and the step of obtaining a first loss of the feature extraction process includes: fusing the first object feature representation and the second object feature representation for each object to obtain a fused feature representation for each object; concatenating and operating the fused feature representation for each object with the second resource feature representation of the corresponding object's second resource data of interest to obtain a predicted attention level for each object's corresponding second resource data of interest; and constructing a first loss of the interest alignment model based on a difference between the actual attention level and the predicted attention level for each object's corresponding second resource data of interest.

[0144] In some embodiments, the interest alignment process includes a cross-domain alignment process, where the first domain feature representation includes a first object feature representation of each object in the first domain data, and the second domain feature representation includes a second object feature representation of each object in the second domain data. The processing unit 1102 is further configured to: invoke an interest alignment model to determine a first focused similarity degree between each pair of objects in the first domain based on the first object feature representation of each object; and invoke an interest alignment model to determine a second focused similarity degree between each pair of objects in the second domain based on the second object feature representation of each object; and perform an alignment process on the second focused similarity degree between the two objects in the second domain to the first focused similarity degree between two corresponding objects in the first domain.

[0145] In some embodiments, the processing unit 1102 is further configured to perform a distance operation on the first object feature representations of each two objects of the plurality of objects, obtain distance information between the first object feature representations of each two objects, perform a probability transformation on the distance information between the first object feature representations of each two objects, and obtain a first noteworthy similarity degree between each two objects.

[0146] In some embodiments, the process of obtaining a loss in the interest alignment processing process includes determining an attention similarity distribution in a first field based on a first attention similarity degree between every two objects in the first field, and determining an attention similarity distribution in a second field based on a second attention similarity degree between every two objects in the second field; constructing a cross-domain alignment loss of the interest alignment model based on the difference between the attention similarity distribution in the first field and the attention similarity distribution in the second field, and taking the cross-domain alignment loss as the second loss.

[0147] In some embodiments, the processing unit 1102 is further configured to obtain the fused feature representation of each object, the object resource graph of the first domain, and the object resource graph of the second domain, and invoke an interest alignment model to perform a graph decoding process for the first domain and a graph decoding process for the second domain on the fused feature representation of each object, respectively, to obtain an object graph of the first domain and an object graph of the second domain, where the object graph is a graph with each object as a node and second interest relationships between objects as connecting edges, and to perform an alignment process on the object graph of the first domain with the object resource graph of the first domain, and an alignment process on the object graph of the second domain with the object resource graph of the second domain.

[0148] In some embodiments, the processing unit 1102 is further configured to obtain graph decoding weights of a first domain, and invoke an interest alignment model to reconstruct an object graph of the first domain based on the graph decoding weights of the first domain and the fused feature representation of each object, and to obtain graph decoding weights of a second domain, and invoke an interest alignment model to reconstruct an object graph of the second domain based on the graph decoding weights of the second domain and the fused feature representation of each object.

[0149] In some embodiments, the process of obtaining a second loss in the interest alignment processing process includes: obtaining an intra-domain alignment loss of the first domain based on the difference between the object graph of the first domain and the corresponding object resource graph; obtaining an alignment loss of the second domain based on the difference between the object graph of the second domain and the corresponding object resource graph; merging the alignment loss of the first domain and the alignment loss of the second domain to obtain an intra-domain alignment loss of the interest alignment model; and taking the cross-domain alignment loss as the second loss; and taking the cross-domain alignment loss as the second loss.

[0150] In some embodiments, the interest alignment processing includes a cross-domain alignment processing and an intra-domain alignment processing, and the processing unit 1102 is further configured to invoke the interest alignment model to perform the following processing: a cross-domain alignment processing between the first domain and the second domain based on the first domain feature representation and the second domain feature representation; and to invoke the interest alignment model to perform the following processing: an intra-domain alignment processing between the first domain and the second domain based on the first domain feature representation and the second domain feature representation.

[0151] In some embodiments, the step of obtaining a second loss of the alignment of interest process includes obtaining a cross-domain alignment loss corresponding to the cross-domain alignment process, and obtaining an intra-domain alignment loss corresponding to the intra-domain alignment process, and constructing the second loss with the cross-domain alignment loss and the intra-domain alignment loss.

[0152] In some embodiments, the interest alignment model includes a first feature extraction module, a second feature extraction module, a cross-domain alignment module, and an intra-domain alignment module, where the first feature extraction module is used to perform feature extraction on first domain data, which is obtained by pre-training, and the second feature extraction module is used to perform feature extraction on second domain data, the cross-domain alignment module is used to perform interest alignment processing on a second attention similarity degree between two objects in the second domain to a first attention similarity degree between two corresponding objects in the first domain, and the intra-domain alignment module is used to perform interest alignment processing on an object graph in the first domain with an object resource graph in the first domain, and to perform interest alignment processing on an object graph in the second domain with an object resource graph in the second domain.

[0153] In some embodiments, the loss in the interest alignment process includes a cross-domain alignment loss obtained in the cross-domain alignment process and an intra-domain alignment loss obtained in the intra-domain alignment process. The processing unit 1102 is further configured to: add the first loss of the interest alignment model, the cross-domain alignment loss, and the intra-domain alignment loss to obtain a target loss of the interest alignment model, and train the interest alignment model according to a direction that reduces the target loss, wherein, in the process of training the interest alignment model according to a direction that reduces the target loss, the training of the first feature extraction module based on the cross-domain alignment loss is stopped.

[0154] In some embodiments, the processing unit 1102 is further used for: obtaining object attributes of the target object to be distributed and resource feature representations of each candidate resource data to be distributed in the second field; invoking the trained interest alignment model to perform feature extraction based on the object attributes of the target object to obtain a fusion feature representation of the target object; performing a similarity comparison between the fusion feature representation of the target object and the resource feature representation of each candidate resource data to obtain a similarity comparison result between the target object and each candidate resource data; and setting the candidate resource data whose similarity comparison result satisfies the resource recommendation rule as the second resource data to be distributed associated with the target object.

[0155] In some embodiments, each or all of the units in the model training apparatus shown in FIG. 11 may be merged into one or several other units, or some units may be further divided into multiple smaller units, which can achieve similar operations and do not affect the technical effects of the embodiments of the present application. The units are divided based on logic functions, and in practical applications, the function of one unit may be realized by multiple units, or the functions of multiple units may be realized by one unit. In other embodiments of the present application, the model training apparatus may include other units, and in practical applications, these functions may be realized in cooperation with other units and may be realized cooperatively by multiple units. According to another embodiment of the present application, the model training apparatus shown in Figure 11 may be configured and the model training method of the embodiment of the present application may be realized by running a computer program (including program code) that can execute the steps associated with the corresponding methods shown in Figures 3 and 9 on a general-purpose electronic device, such as a computer, including processing and storage elements such as a central processing unit (CPU), a random access memory (RAM), and a read-only memory (ROM). The computer program may be stored on a computer-readable recording medium, for example, and installed in and operated by the electronic device through the computer-readable recording medium.

[0156] The embodiments of the present application take into full consideration the interest similarity between objects and employ an object interest similarity alignment method to support training of an interest alignment model, so that the interest alignment model can generate a relatively accurate feature representation for an object even when applied to a domain with relatively little interaction data. For example, considering that the interaction data in a first domain is relatively abundant, the feature representation of the first domain extracted based on abundant interaction data is more accurate. Therefore, the interest alignment model is trained by aligning the interest similarity between two objects in a second domain with relatively little interaction data to the interest similarity between the two objects in the first domain with relatively much interaction data. Even when the interaction data in the second domain is relatively sparse, the trained interest alignment model can generate a relatively accurate feature representation for the object, so that resource data recommended for the object based on the accurate feature representation can meet the personalized needs of the object and achieve high-precision resource recommendation. The embodiment of the present application adopts a method of jointly training the interest alignment model using the first loss in the feature extraction process and the loss in the interest alignment processing process, which can enrich the learning targets of the entire interest alignment model and ensure the performance of the interest alignment model through multi-target learning.

[0157] 12 shows a structural diagram of an electronic device provided by an exemplary embodiment of the present application. As shown in FIG. 12, the electronic device includes a processor 1201, a communication interface 1202, and a computer-readable storage medium 1203. The processor 1201, the communication interface 1202, and the computer-readable storage medium 1203 may be connected via a bus or other means. The communication interface 1202 is used for receiving and transmitting data. The computer-readable storage medium 1203 may be stored in a storage device of the electronic device. The computer-readable storage medium 1203 is used to store a computer program, the computer program including program instructions, and the processor 1201 is used to execute the program instructions stored in the computer-readable storage medium 1203. The processor 1201 (also called CPU (Central Processing Unit)) is the computational and control core of an electronic device, and is suitable for implementing one or more instructions, specifically, for loading and executing one or more instructions to implement a corresponding method flow or corresponding function.

[0158] The embodiment of the present application further provides a computer-readable storage medium (Memory), which is a storage device in an electronic device and is used to store programs and data. It is to be understood that the computer-readable storage medium here may include an embedded storage medium in the electronic device, but may also include an extended storage medium supported by the electronic device. The computer-readable storage medium provides a storage space in which a processing system of the electronic device is stored. The storage space also stores one or more instructions suitable for being loaded and executed by the processor 1201, which may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here may be a high-speed RAM storage device, a non-volatile memory, such as at least one magnetic disk storage device, or at least one computer-readable storage medium located remotely from the processor.

[0159] In some embodiments, the computer-readable storage medium stores one or more instructions, and the processor 1201 loads and executes the one or more instructions stored in the computer-readable storage medium to implement corresponding steps in the embodiments of the model training method.

[0160] An embodiment of the present application further provides a computer program product, the computer program product including computer-executable instructions stored in a computer-readable storage medium, a processor of an electronic device reading the computer-executable instructions from the computer-readable storage medium, and executing the computer-executable instructions to cause the electronic device to perform the model training method.

[0161] As can be appreciated by those skilled in the art, each exemplary unit and algorithm step described in the embodiments disclosed herein can be combined and realized as electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Experts may use different methods to implement the described functions for each specific application, but such implementation should not be deemed to exceed the scope of the present application.

[0162] The above embodiments may be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, they may be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer-executable instructions. When the computer program instructions are loaded and executed on a computer, the above-described flow or functions according to the embodiments of the present invention are generated entirely or partially. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer-executable instructions may be stored in or transmitted through a computer-readable storage medium. The computer-executable instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optics, digital user line (DSL)) or wireless (e.g., infrared, radio, microwave, etc.) methods. The computer-readable storage medium may be any available medium accessible by a computer, or may be a data processing device, such as a server or data center, that includes one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, and magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid state disks (SSDs)).

[0163] The above are only specific embodiments of the present application, but the scope of protection of the present application is not limited thereto, and any changes or substitutions that a person skilled in the art can easily make within the technical scope of the present invention are also included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be in accordance with the scope of protection of the claims. [Explanation of symbols]

[0164] 101 terminals 102 Server 103 Server 1021 Database 1031 Database 1101 Acquisition Units 1102 Processing Unit 1201 processor 1202 Communication Interface 1203 Computer-readable storage medium

Claims

1. 1. A method for training a model, the method being performed by an electronic device, the method comprising: obtaining a data set, the data set including first domain data and second domain data, the first domain data including at least one object and first resource data of interest within a first domain for each object, and the second domain data including the at least one object and second resource data of interest within a second domain for each object; Invoking an interest alignment model to perform feature extraction on the first domain data to obtain a first domain feature representation, and invoking the interest alignment model to perform feature extraction on second neighbor data to obtain a second domain feature representation; calling the interest alignment model to perform the following process: interest alignment between the first domain and the second domain based on the first domain feature representation and the second domain feature representation; training the interest alignment model according to a direction of reducing the first loss and the second loss to obtain a trained interest alignment model; A model training method, in which the first loss is a loss corresponding to the feature extraction, the second loss is a loss corresponding to the interest alignment process, and the trained interest alignment model is used to recommend resource data for a target object in the second field.

2. The step of invoking an interest alignment model to perform feature extraction on the first domain data to obtain a first domain feature representation includes: a step of constructing an object resource graph of the first field based on the first field data, the object resource graph of the first field being a graph in which each object and each first resource data is a node and a first interest relationship between the object and the first resource data is a connection edge; Invoking an interest alignment model to perform a graph encoding process on the object resource graph of the first domain to obtain a first object feature representation for each of the objects and a first resource feature representation for each of the first resource data; and constructing the first domain feature representation with a first object feature representation for each of the objects and a first resource feature representation for each of the first resource data.

3. The step of invoking the interest alignment model to perform feature extraction on the second neighborhood data to obtain a second domain feature representation includes: a step of constructing an object resource graph of the second field based on the second field data, the object resource graph of the second field being a graph in which each object and each second resource data is a node and a first interest relationship between the object and the second resource data is a connection edge; Invoking the interest alignment model to perform a graph encoding process on the object resource graph of the second domain to obtain a second object feature representation for each of the objects and a second resource feature representation for each of the second resource data; and constructing the second domain feature representation with a second object feature representation for each of the objects and a second resource feature representation for each of the second resource data.

4. Any one of the at least one object is represented as a training object, and the step of calling the interest alignment model to perform a graph encoding process on the object resource graph of the second domain to obtain a second object feature representation of each of the objects includes: obtaining an initial feature representation of the training objects in the second domain and an initial feature representation of each second resource data of interest of the training objects; calling the interest alignment model to perform the following process: calculating a degree of association between the training object and each second resource data of interest of the training object according to the initial feature representation of the training object and the initial feature representation of each second resource data of interest of the training object; and determining a second object feature representation for the training object based on a degree of association between the training object and each second resource data of interest of the training object, an initial feature representation for the training object in the second domain, and an initial feature representation for each second resource data of interest of the training object.

5. The data set includes an actual attention level for the corresponding second resource data of interest for each of the objects, and the method further comprises: fusing the first object feature representation and the second object feature representation for each of the objects to obtain a fused feature representation for each of the objects; connecting and manipulating the fused feature representation of each object with the second resource feature representation of the second resource data of interest of the corresponding object to obtain a predicted attention level for the second resource data of interest of the corresponding object; 4. The method of claim 3, further comprising: constructing a first loss of the interest alignment model based on a difference between actual attention and predicted attention to the corresponding second resource data of interest of each object.

6. The interest alignment process includes a cross-domain alignment process, the first domain feature representation includes a first object feature representation of each object in the first domain data, and the second domain feature representation includes a second object feature representation of each object in the second domain data, and the step of calling the interest alignment model to perform the following process, i.e., performing interest alignment process between the first domain and the second domain based on the first domain feature representation and the second domain feature representation, includes: Invoking the interest alignment model to determine a first interest similarity degree between each pair of objects in the first domain based on a first object feature representation of each object; Invoking the interest alignment model to determine a second interest similarity degree between each pair of objects in the second domain based on a second object feature representation of each object; and performing an alignment process on a second degree of similarity between two objects in the second field to a first degree of similarity between two corresponding objects in the first field.

7. Invoking the interest alignment model to determine a first interest similarity degree between each pair of objects in the first domain based on the first object feature representation of each object, performing a distance calculation on first object feature representations of two of the plurality of objects to obtain distance information between the first object feature representations of the two of the objects; and performing a probability transformation on the distance information between the first object feature representations of each pair of objects to obtain a first salient similarity measure between each pair of objects.

8. The method further comprises: determining a salient similarity distribution of the first domain based on a first salient similarity degree between every two objects in the first domain; determining a second attention similarity distribution of the second field according to a second attention similarity degree between every two objects in the second field; 7. The method of claim 6, further comprising: constructing a cross-domain alignment loss of the interest alignment model based on the difference between the attention similarity distribution of the first domain and the attention similarity distribution of the second domain; and setting the cross-domain alignment loss as a second loss.

9. The interest alignment process includes an intra-domain alignment process, and the step of calling the interest alignment model to perform the following process, i.e., performing interest alignment between the first domain and the second domain based on the first domain feature representation and the second domain feature representation, includes: obtaining a fused feature representation of each object, an object resource graph of the first domain, and an object resource graph of the second domain; a step of calling the interest alignment model and performing a graph decoding process for the first field and a graph decoding process for the second field on the fused feature representation of each of the objects, respectively, to obtain an object graph of the first field and an object graph of the second field, wherein the object graph is a graph in which each object is a node and second interest relationships between objects are connection edges; 2. The method of claim 1, further comprising the steps of: aligning the object graph of the first domain to the object resource graph of the first domain; and aligning the object graph of the second domain to the object resource graph of the second domain.

10. The step of calling the interest alignment model and performing a graph decoding process for a first domain and a graph decoding process for a second domain on the fused feature representation of each of the objects to obtain an object graph for the first domain and an object graph for the second domain, Obtaining graph decoding weights of the first domain, and invoking the interest alignment model to reconstruct an object graph of the first domain based on the graph decoding weights of the first domain and the fused feature representation of each object; 10. The method of claim 9, further comprising: obtaining graph decoding weights for the second domain; and invoking the interest alignment model to reconstruct an object graph for the second domain based on the graph decoding weights for the second domain and the fused feature representation of each object.

11. The method further comprises: obtaining an intra-domain alignment loss for the first domain based on the difference between the object graph of the first domain and a corresponding object resource graph; obtaining an intra-domain alignment loss for the second domain based on the difference between the object graph of the second domain and a corresponding object resource graph; 10. The method of claim 9, further comprising: merging the intra-domain alignment loss of the first domain with the intra-domain alignment loss of the second domain to obtain an intra-domain alignment loss of the alignment model of interest; and setting the intra-domain alignment loss as the second loss.

12. The interest alignment process includes a cross-domain alignment process and an intra-domain alignment process, and the step of calling the interest alignment model to perform the following process, i.e., performing interest alignment between the first domain and the second domain based on the first domain feature representation and the second domain feature representation, includes: Invoking the interest alignment model to perform the following process: performing a cross-domain alignment process between the first domain and the second domain based on the first domain feature representation and the second domain feature representation; 2. The method of claim 1, further comprising: invoking the interest alignment model to perform the following process: performing intra-domain alignment processing between the first domain and the second domain based on the first domain feature representation and the second domain feature representation.

13. The method further comprises: obtaining a cross-domain alignment loss corresponding to the cross-domain alignment process, and obtaining an intra-domain alignment loss corresponding to the intra-domain alignment process; and constituting the second loss with the cross-domain alignment loss and the intra-domain alignment loss.

14. The interest alignment model includes a first feature extraction module, a second feature extraction module, a cross-domain alignment module, and an intra-domain alignment module; The first feature extraction module is used to perform feature extraction on the first field data, and the first feature extraction module is obtained by pre-training; The second feature extraction module is used to perform feature extraction on the second field data; The cross-domain alignment module is used for performing an interest alignment process on a second interest similarity degree between two objects in the second domain with a first interest similarity degree between two corresponding objects in the first domain; The method of any one of claims 1 to 13, wherein the intra-domain alignment module is used to perform interest alignment processing of the object graph of the first domain to the object resource graph of the first domain, and to perform interest alignment processing of the object graph of the second domain to the object resource graph of the second domain.

15. The second loss includes a cross-domain alignment loss obtained by a cross-domain alignment process and an intra-domain alignment loss obtained by an intra-domain alignment process, and the step of training the alignment model of interest according to a direction of reducing the first loss and the second loss to obtain a trained alignment model of interest includes: adding the first loss of the alignment model of interest, the cross-domain alignment loss, and the intra-domain alignment loss to obtain a target loss of the alignment model of interest; training the interest alignment model according to a direction that reduces the target loss; The method of claim 14 , further comprising: suspending training of the first feature extraction module based on the cross-domain alignment loss during training of the interest alignment model according to a direction that reduces the target loss.

16. The method further comprises: obtaining object attributes of the target object to be distributed and resource feature representations of each candidate resource data to be distributed in the second field; Invoking the trained interest alignment model to perform feature extraction based on object attributes of the target object to obtain a fused feature representation of the target object; performing a similarity comparison between the fusion feature representation of the target object and the resource feature representation of each of the candidate resource data, and obtaining a similarity comparison result between the target object and each of the candidate resource data; The method according to any one of claims 1 to 15, further comprising a step of selecting candidate resource data whose similarity comparison result satisfies a resource recommendation rule as second resource data to be distributed and associated with the target object.

17. A model training device, an acquisition unit and a processing unit; The acquiring unit is configured to acquire a data set, the data set including first field data and second field data, the first field data including at least one object and first resource data of interest within a first field for each object, and the second field data including the at least one object and second resource data of interest within a second field for each object; the processing unit is configured to invoke an interest alignment model to perform feature extraction on the first domain data to obtain a first domain feature representation, and to invoke the interest alignment model to perform feature extraction on the second neighborhood data to obtain a second domain feature representation; The processing unit is further configured to call the interest alignment model to perform the following process: interest alignment between the first domain and the second domain based on the first domain feature representation and the second domain feature representation; The processing unit is further configured to train the interest alignment model according to a direction that reduces the first loss and the second loss, and obtain a trained interest alignment model; A model training device, wherein the first loss is a loss corresponding to the feature extraction, the second loss is a loss corresponding to the interest alignment process, and the trained interest alignment model is used to recommend resource data for a target object in the second field.

18. An electronic device, a processor adapted to execute computer-executable instructions; and a computer-readable storage medium having stored thereon computer-executable instructions that, when executed by the processor, implement the model training method of any one of claims 1 to 16.

19. 1. A computer-readable storage medium, comprising:

17. The computer-readable storage medium having stored thereon computer-executable instructions that, when executed, implement the model training method of any one of claims 1 to 16.

20. 1. A computer program product comprising: The computer program product comprises computer-executable instructions that, when executed by a processor, implement the model training method of any one of claims 1 to 16.

Images