Model training methods, apparatus, electronic equipment, and computer programs

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 with application number 202211507481.6 and filing date November 29, 2022, and claims priority from the Chinese patent application, the entire contents of which are incorporated herein by reference.

[0002] This application relates to the field of computer technology, and more specifically to the field of machine learning, and in particular to model training methods, apparatus, electronic equipment, computer-readable storage media, and computer program products. [Background technology]

[0003] Personalized Recommender Systems are platforms used to provide personalized resource recommendations for users (i.e., objects). For example, a Personalized Recommender System under a video scene would support recommending video resources of interest to the user.

[0004] Currently, personalized recommendation systems support achieving the goal of personalized recommendations by collecting large amounts of user (i.e., object) behavioral data (e.g., interaction data between users and resource data), building resource recommendation models using pre-designed recommendation algorithms, and generating specific recommendation lists for objects through these built resource recommendation models. However, practical experience has revealed that personalized recommendation systems do not possess rich data for all objects. Therefore, resource recommendation models trained on large amounts of interaction data for objects are not applicable to making resource recommendations for objects with relatively little interaction data, leading to relatively poor recommendation performance for these models. [Overview of the Initiative] [Problems that the invention aims to solve]

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

[0006] Embodiments of the present application provide a model training method, the method being performed by electronic equipment, and the method is A step of obtaining a data set, wherein the data set includes first-field data and second-field data, the first-field data includes at least one object and first-source data of interest within the first-field of each object, and the second-field data includes at least one object and second-source data of interest within the second-field of each object. The process involves calling an interest alignment model to extract features from the first domain data and obtain a first domain feature representation, and then calling the interest alignment model again to extract features from the second neighboring data and obtain a second domain feature representation. The steps include calling the interest alignment model and performing the following process, namely, performing an interest alignment process between the first and second domains based on the first and second domain feature representations, The process includes the steps of training the interest alignment model in a direction that reduces the first loss and the second loss, and obtaining a trained interest alignment model, Here, the first loss is the loss corresponding to the feature extraction, and the second loss is the loss corresponding to the interest alignment process, and the trained interest alignment model is used to recommend resource data for target objects in the second domain.

[0007] Embodiments of the present application provide a model training device, which is It includes an acquisition unit and a processing unit, The acquisition unit is configured to acquire a data set, the data set comprising first-field data and second-field data, wherein the first-field data comprises at least one object and first resource data of interest within the first-field of each object, and the second-field data comprises at least one object and second resource data of interest within the second-field of each object. The processing unit is configured to call an interest alignment model to extract features from the first domain data and obtain a first domain feature representation, and then call the interest alignment model again to extract features from the second neighboring data and obtain a second domain feature representation. The processing unit is further configured to call the interest alignment model and perform the following processing, namely, interest alignment processing between the first and second domains based on the first and second domain feature representations: The processing unit is further configured to train the interest alignment model in a manner that reduces the first loss and the second loss, and to obtain a trained interest alignment model. Here, the first loss is the loss corresponding to the feature extraction, and the second loss is the loss corresponding to the interest alignment process, and the trained interest alignment model is used to recommend resource data for target objects in the second domain.

[0008] Embodiments of the present application provide an electronic device, which is, A processor used to load and execute computer executable instructions, The present invention includes a computer-readable storage medium in which computer-executable instructions are stored, and which, when executed by a processor, implement the model training method described above.

[0009] Embodiments of the present invention provide a computer-readable storage medium in which computer-executable instructions are stored, which are loaded by a processor and are suitable for executing the model training method described above.

[0010] Embodiments of the present application provide a computer program product which includes computer executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer executable instructions from the computer-readable storage medium, and the computer executable instructions realize the model training method described above when executed by the processor. [Effects of the Invention]

[0011] In the embodiments of this application, a data set is constructed by acquiring resource data of interest in the corresponding field of at least one object from among multiple fields, and it is supported to acquire, for example, first-field data from the first field and second-field data from the second field. Here, the first field may be a field with relatively rich interaction data, while the second field may be neighborhood data with relatively little interaction data. Subsequently, an interest alignment model is invoked to perform feature extraction on the first-field data and the second-field data, respectively, thereby extracting first-field feature representations and second-field feature representations. The interest alignment model is invoked again to perform interest alignment processing in the two fields based on the first-field feature representations and second-field feature representations. Finally, it is supported to train the interest alignment model in a direction that reduces the first loss in the feature extraction process and the loss in the interest alignment processing process. The embodiments of this application fully consider the interest similarity between objects and employ an object interest similarity alignment method to support training an interest alignment model, enabling the model to produce relatively accurate feature representations for objects even when applied to domains with relatively little interaction data. For example, considering that the feature representations for the first domain extracted based on the abundant interaction data are more accurate because the interaction data for the first domain is relatively large, the interest alignment model is trained by employing a method that aligns the interest similarity of two objects in a second domain with relatively little interaction data to the interest similarity of the two objects in the first domain with relatively large interaction data. Even when the interaction data for the second domain is relatively small, the trained interest alignment model can be ensured to produce relatively accurate feature representations for objects, thereby ensuring that the resource data recommended for objects based on accurate feature representations meets the personalization needs of the objects and achieving high-precision resource recommendations.The embodiments of this application adopt a method of jointly training the interest alignment model with the first loss in the feature extraction process and the loss in the interest alignment process, which can enrich the learning target of the entire interest alignment model and ensure the performance of the interest alignment model through multi-target learning.

[0012] To more clearly explain the technical solutions in the embodiments of this application or the prior art, the drawings that need to be used in the following descriptions of the embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following descriptions are only some embodiments of this application. For those skilled in the art, other drawings can also be obtained according to these drawings without creative efforts.

Brief Description of the Drawings

[0013] [Figure 1a] It is a schematic diagram of the architecture of a resource recommendation system provided by one exemplary embodiment of this application. [Figure 1b] It is a schematic diagram of the architecture of another resource recommendation system provided by one exemplary embodiment of this application. [Figure 2] It is a schematic diagram of the model structure of an interest alignment model provided by one exemplary embodiment of this application. [Figure 3] It is a flow schematic diagram of a model training method provided by one exemplary embodiment of this application. [Figure 4] It is a schematic diagram of the construction of an object resource graph in a second field provided by one exemplary embodiment of this application. [Figure 5] It is a schematic diagram of the architecture of graph symbolic learning provided by one exemplary embodiment of this application. [Figure 6] It is a schematic diagram of the interest alignment process carried out between a first field and a second field provided by one exemplary embodiment of this application. [Figure 7]A schematic diagram of an interest alignment process performed in each of a first field and a second field provided by one exemplary embodiment of the present application. [Figure 8] A schematic diagram of a graph reconstruction provided by one exemplary embodiment of the present application. [Figure 9] A flow schematic diagram of another model training method provided by one exemplary embodiment of the present application. [Figure 10] A schematic diagram for matching the fusion feature representation of a target object and the resource feature representation of each candidate resource data provided by one exemplary embodiment of the present application. [Figure 11] A structural schematic diagram of a model training device provided by one exemplary embodiment of the present application. [Figure 12] A structural schematic diagram of an electronic device provided by one exemplary embodiment of the present application.

Modes for Carrying Out the Invention

[0014] Hereinafter, in combination with the drawings in the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Of course, the described embodiments are only a part of the embodiments of the present application, not all of the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the scope protected by the present application in any case.

[0015] In the embodiments of the present application, a model training solution is provided. Specifically, training and application solutions for an interest alignment model under a resource recommendation scenario are provided. Hereinafter, technical terms and related concepts related to the model training solution provided by the embodiments of the present application will be briefly introduced.

[0016] I. Artificial Intelligence (AI).

[0017] Artificial intelligence (AI) is a theory, method, technology, and application system that uses digital computers, or machines controlled by digital computers, to simulate, extend, and augment human intelligence to sense the environment, acquire knowledge, and use that knowledge to achieve the best results. AI technology is a comprehensive discipline encompassing a wide range of related fields, including both hardware-level and software-level technologies. Fundamental AI technologies generally include, for example, sensors, dedicated AI chips, cloud computing, distributed memory, big data processing technologies, operating / interaction systems, and mechatronics. AI software technologies primarily encompass several major areas, such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning. Machine learning (ML) is a multidisciplinary, interdisciplinary field that specializes in studying how computers simulate or realize human learning behavior, such as probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory, to acquire new knowledge or skills, reorganize existing knowledge structures, and continuously improve their own performance. Machine learning, and deep learning in particular, typically include techniques such as artificial neural networks, trust networks, reinforcement learning, transfer learning, inductive learning, guided learning, and meta-learning.

[0018] The model training solution provided in the embodiment 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) Artificial neural networks are a type of method for realizing machine learning tasks, and when neural networks are discussed 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, and such a network structure resembles the biological nervous system and is used to simulate the interaction between living organisms and the natural environment. Furthermore, the more network structures there are, the richer the functionality of the neural network tends to be. Neural networks are one relatively large concept, and for different learning tasks such as speech, text, and images, neural network models have been derived that are applied more specifically to the learning task, such as graph neural networks (GNNs). Graph neural networks are 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 search for features and patterns in graph structured data through learning of graph structured data, and can satisfy the needs of graph learning tasks such as clustering, classification, prediction, partitioning, and generation. For example, graph neural networks that satisfy the above graph learning task needs may include, but are not limited to, graph convolutional networks (GCN), graph recurrent networks (GRN), and graph attention networks (GAT).

[0020] The embodiments of the present invention specifically relate to a graph attention network included in a graph neural network, wherein the graph attention network includes an attention mechanism, through which the neural network can focus only on the information necessary for task learning and achieve the selection of specific inputs. In other words, the introduction of an attention mechanism in a graph attention network causes the neural network to focus on graph structure data that is more relevant to the task (e.g., nodes and edges included in the graph), thereby increasing the effectiveness of training and the accuracy of testing. Here, the attention mechanism is produced by simulating human vision attention, for example, the field of view of the human eye is relatively wide, but the field of view often focuses on only one small area, that is, the human eye focuses more on important areas in the field of view (i.e., areas of interest), and therefore, to resemble human vision attention, the attention mechanism is intended to focus on important information in the task and omit unimportant information. 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 situations where training data and / or annotations are insufficient. Specifically, it improves a new task to be learned by transferring knowledge from related tasks that have already been learned. For example, it helps train a new network model by transferring (copying) pre-trained model parameters (using data and annotations from non-current tasks) to the new network model. While most machine learning algorithms are designed to solve a single task, developing algorithms that facilitate transfer learning is a topic that machine learning cells continue to focus on.

[0022] Transfer learning relates to the concept of a domain, where the domain is the core of transfer learning and consists of data features and feature distributions. In the context of resource recommendation, a domain is generally used to represent different application areas, such as shopping, movies, reading, videos, social public accounts, documents, and live streaming. As can be seen from the above description, the task of transfer learning is to apply a model that has been trained in an old field to a new field, starting from the similarity of the problems. Therefore, the domains to which transfer learning relates often include at least two domains: the source domain and the target domain. Here, the source domain is a different domain from the field to which the sample data (i.e., resource data) of the model being trained belongs, and the source domain often contains rich supervision information (i.e., rich resource data used to train the model), while the target domain is the same domain as the field to which the sample data of the model being trained belongs, and the target domain usually contains only a small amount of resource data used to train the model.

[0023] 2. Resource Recommendations

[0024] Resource recommendation, also known as resource delivery, may refer to the process by which a resource recommendation platform (or resource recommendation system, or personalized recommendation system as described above) delivers resource data contained within the platform to platform objects (e.g., one or more resource recipients who have registered a platform account or have temporarily logged into the platform). Here, resource data contained within the resource recommendation platform may be called Internet resources (or abbreviated as resources) and may include, but are not limited to, videos (which can be divided into long and short videos depending on their length), audio (e.g., music, or audio), video, or literature (e.g., journals, or articles). Embodiments of the present application do not limit the resource types of resource data. The resource types of resource data are related to the field (i.e., the domain as described above) to which the resource recommendation platform delivering the resource data belongs, and the resource types of resource data delivered by different domains may differ. For example, the resource type of resource data in the live streaming domain is video, while the resource type of resource data in the document domain is documents. Embodiments of the present application do not limit the resource types of 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 that completes a specific task in one or more items. Classifying application programs according to different dimensions (e.g., the operation method and functionality of the application program) may result in different types of the same application program under different dimensions. For example, classifying according to the operation method of an application program may include, but are not limited to, client terminals installed on a terminal, mini-programs that can be used without requiring download and installation (as sub-programs of the client terminal), and web (World Wide Web) application programs opened through a browser. Furthermore, classifying according to, for example, the functional type of an application program may include, but are not limited to, IM (Instant Messaging) application programs and content interaction application programs. Here, an instant messaging application program refers to an application program for internet-based instant messaging and social interaction, and an instant messaging application program may include, but are not limited to, social application programs with communication functions, map application programs with social interaction functions, and game application programs. A content interaction application program refers to an application program that can implement content interaction, and may include, for example, online banking, sharing platforms, personal spaces, and news applications.

[0026] The resource recommendation platform may be a plugin (or function) that supports resource recommendation and is included in the application program mentioned above. For example, if the application program is an instant messaging application program in client terminal format, then the resource distribution platform may be a resource distribution plugin 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 in the process of a target object (e.g., any one object using the instant messaging application program) interacting using the instant messaging application program, functions such as viewing and publishing resources can be performed without the need for application jumps (e.g., jumps from the instant messaging application program to a resource recommendation application program independent of the instant messaging application program).

[0027] It is important to explain that the types of resource data distributed by the same resource recommendation platform are not limited to just one type. For example, the resource types of resource data supported by the same resource recommendation platform may include both video and document. Furthermore, the embodiments of this application do not limit the resource types of resource data distributed by the resource recommendation platform, nor which specific type of application program the resource recommendation platform is, or which type of application program provides the resource recommendation function. For the sake of clarity, subsequent embodiments will use the example that the resources distributed by the resource recommendation platform (or resource recommendation system) are short videos.

[0028] Based on the above-mentioned descriptions of transfer learning and resource recommendation, the embodiment of the present application proposes a model training solution relating to cross-domain recommendation. Cross-domain recommendation aims to capture preferences or interests of an object in certain aspects by combining data from multiple domains and introducing information from other domains (e.g., at least one source domain) and analyzing interaction data of the object in other domains, thereby enabling better recommendations on the target domain and even on multiple domains.

[0029] In concrete implementation, the model training solution provided by the embodiment of the present application is a training solution related to an interest alignment model, which may also be called a cross-domain recommendation model, and specifically, it utilizes interest relationships between objects, i.e., interest relationships of different objects to resource data (for example, resource data of interest to object A and object B are similar), to perform interest alignment learning on the interest alignment model, which includes cross-domain and intra-domain object interest relationships, thereby achieving the objective of guiding target domain interest learning. The general training flow of the model training solution may include the following: Obtain a data set to be used for model training, which includes first-domain data of the first domain and second-domain data of the second domain. Then, call 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 call 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 in a direction that reduces 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 described above, a trained interest alignment model can be obtained, and this interest alignment model is used to recommend resource data for a target object in the second domain (for example, any one of the objects in the resource data to be delivered). Specifically, after obtaining the trained interest alignment model, the embodiment of the present invention further supports calling the trained interest alignment model to recommend resource data for a target object that logs into the resource recommendation platform. Specifically, the interest alignment model is used to generate an accurate object feature representation (i.e., an embedding vector that can characterize the attributes or features of the target object) for the target object of the resource data to be delivered, and then, according to the object feature representation, a portion of the resource data that satisfies the interests and hobbies of the target object is screened from the database of the resource recommendation platform, and resource data is subsequently selected from this portion of resource data and recommended for the target object.

[0031] As can be seen from the above, the embodiment of the present application combines transfer learning to enrich the data in the second domain by utilizing the rich knowledge and information of the first domain, increasing the available information, mitigating the problem of the sparse number of samples in the second domain, and ensuring the model performance of the trained interest alignment model. When the trained interest alignment model is applied to the second domain, which has relatively little interaction data, it can also produce relatively accurate feature representations for objects, thereby ensuring that the resource data recommended for objects based on accurate feature representations meets the object personalization needs and achieving high-precision resource recommendations.

[0032] To facilitate understanding of the model training solution provided by the embodiment of this application, the resource recommendation scenario to which the embodiment of this application relates will be briefly introduced below, using the resource recommendation system shown in Figure 1a. As shown in Figure 1a, the resource recommendation system includes terminal 101, server 102, and server 103, and the embodiment of this application does not limit the number or naming of terminal 101, server 102, and server 103.

[0033] Terminal 101 may refer to a terminal device used by a resource recipient of resource data to be distributed, which is registered in the resource recommendation platform. The terminal device may include, but is not limited to, devices such as smartphones (e.g., smartphones equipped with the Android system or smartphones equipped with the Internetworking Operating System (iOS)), tablet computers, portable personal computers, mobile internet devices (MIDs), in-vehicle devices, and head-mounted devices, and the embodiments of this application will be described without limiting the type of terminal device. The terminal device is equipped with a resource recommendation platform, specifically an application program that incorporates the resource recommendation platform, and the resource recommendation platform may be a recommendation platform for the second field. In this way, the resource recipient may perform operations such as receiving resource data under the second field through the resource recommendation platform equipped in the terminal device.

[0034] Server 102 is the server that terminal 101 corresponds to, specifically, the backend server of the resource recommendation platform equipped on terminal 101, and is used to provide computing and application service support for the resource recommendation platform on terminal 101 by interacting with terminal 101. Server 103 may also be the backend server that the resource recommendation platform for the first domain corresponds to, and may provide first domain data for model training by communicating data with server 102. The servers (e.g., server 102 and server 103) may be independent physical servers, server clusters consisting of multiple physical servers, or distributed systems, or further, cloud servers that provide basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Terminal 101 and the servers may be connected directly or indirectly via wired or wireless communication, but this application is not limited thereto.

[0035] Server 102 further includes a database 1021 which may be used to store all resource data, objects, and object interaction data (user-item) between objects and resource data included in the resource recommendation platform of the second domain (i.e., the target domain). For example, under a video recommendation scene, the video platform includes object 1 and object 2, and further, if 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 objects and videos, such as object 1-video 1 (representing that object 1 triggered video 1), object 1-video 2 (representing that object 1 triggered video 2), object 2-video 1 (representing that object 2 triggered video 1), and object 2-video 2 (representing that object 2 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 domain (i.e., the source domain). The specific contents of the object interaction data can be found in the relevant description of the second domain above, but will not be explained in detail here.

[0036] There are two long-standing problems in resource recommendation systems: the problem of sparse data and the problem of cold starts. The so-called problem of sparse data refers to the fact that there is very little object interaction data for objects and resource data, making it very difficult to capture the interests of objects or the characteristics of resource data based on a small amount of object interaction data. The so-called problem of cold starts refers to the situation where objects newly registered or logged into the resource recommendation platform, or resource data newly uploaded or issued to the resource recommendation platform, have no object interaction data. Due to the degree of data sparseness and the scarcity of object interaction data under cold starts, conventional recommendation algorithms that make recommendations based on object interaction data for objects and resource data have difficulty making appropriate recommendations under these two conditions. In contrast, the embodiment of this application combines transfer learning to train the interest alignment model, that is, it utilizes the rich knowledge and information of the first domain to support improved performance in the second domain, reducing the number of training samples required in the second domain, thereby relatively well solving the problem of a relatively small number of training samples that can lead to data sparseness and cold starts in recommendation scenes. Thus, when the interest alignment model obtained through cross-domain training is applied to the second domain, it is still possible to generate accurate feature representations for objects where object interaction data is absent or scarce, making it easier to recommend resource data (i.e., resources included in the target domain) for the user based on accurate feature representations. As can be seen from the above, the model training solution provided by the embodiment of this application can be applied to resource recommendation scenes, and is particularly effective when applied to model training under sparse data scenes or cold start scenes.

[0037] The model training solution provided by the embodiment of the present application may be executed by either terminal 101 or server 102 in the architecture shown in Figure 1a, or both. In other words, the electronic device that is the execution body of the embodiment of the present application may be at least one of terminal 101 or server 102. The training process of the interest alignment model mentioned above is executed by server 102, and as shown in Figure 1a, the trained interest alignment model may be directly equipped in server 102. Thus, each time a resource is delivered, the trained interest alignment model is invoked by server 102 to realize resource recommendation. Under this implementation, the electronic device that is the execution body used to execute the solution provided by the embodiment of the present application is server 102. The trained interest alignment model may also be equipped in terminal 101. As shown in Figure 1b, after the interest alignment model is trained in server 102, the trained interest alignment model is sent to terminal 101, and terminal 101 is equipped with the trained interest alignment model. In this case, the electronic device that is the execution unit of the model training solution provided by the embodiment of the present invention includes a terminal 101 and a 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 on the terminal 101, in which case the electronic device that is the execution unit of the model training solution provided by the embodiment of the present invention includes the terminal 101.

[0038] Figures 1a and 1b above are merely schematic diagrams of the architecture of an exemplary resource recommendation system provided by the embodiment of the present application. In practical applications, the architecture may undergo adaptive changes. For example, the number of first fields provided by the embodiment of the present application may be multiple, meaning the embodiment can be adapted to multiple source domains. In this way, information from more domains can be used to support the 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 operated in a specific product or technology, for example, when making resource recommendations to resource data, it is unavoidable to obtain attribute information of the target object (e.g., information such as age, gender, or name), and therefore permission or consent from the target object must be obtained. In addition, the collection, use, and processing of relevant data must comply with relevant laws, regulations, and standards of the relevant countries and regions.

[0039] As can be seen from the model training solution described above, the model training solution provided by the embodiment of the present application mainly relates to two aspects: one aspect is to train a model to obtain a trained interest alignment model, and the other aspect 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 later by the embodiment of the present application, we first combine the schematic diagram of the model structure of the interest alignment model provided by the embodiment of the present application, given in Figure 2. As shown in Figure 2, the interest alignment model (or object interest alignment model) provided by the embodiment of the present application may include a single-domain feature extraction module, a cross-domain alignment module, and an intra-domain alignment module.

[0040] A single-domain feature extraction module is used to extract features from domain data under a single domain (e.g., a field) and obtain a domain feature representation of that single domain. Depending on the number of fields in a practical application, the number of single-domain feature extraction modules may vary. For example, under a case involving one source domain (i.e., the first field) and one target domain (i.e., the second field), 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 extract features from first field data under the first field and obtain a first field feature representation, and the second feature extraction module is used to extract features from second field data under the second field and obtain a second field feature representation. It is worth noting that the first feature extraction module corresponding to the first field is obtained through pre-training; that is, in the pre-training stage, the first feature extraction model may be pre-trained to learn the object's interests and preferences in the first field. In this way, with the help of object interaction data from the first domain, the first feature extraction module and the feature representations of objects 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, the embodiment of this application uses the already trained first feature extraction module as an initial point based on the idea of ​​transfer learning, and reuses it in the process of developing a model for task learning in the second domain. In this way, the feature representations of objects produced based on the already trained first feature extraction module can be used as actual tags for supervised learning to optimize the interest alignment model.

[0041] A cross-domain alignment module, also known as a cross-domain interest alignment module or an inter-domain alignment module, may be used to perform interest alignment processing on the interest relationship between two objects in a second domain (which may be called the second degree of interest similarity in the embodiments of this application) with respect to the interest relationship between two corresponding objects in a first domain (which may be called the first degree of interest similarity in the embodiments of this application). In other words, the embodiments of this application design an inter-domain alignment module to align the interest similarity of objects in a second domain with the interest similarity in a first domain. In this way, the interest relationship of objects in a first domain can be transferred to a second domain, thereby guiding the interest learning process in the second domain.

[0042] An intradomain alignment module, also called an intradomain interest alignment module, may be used to perform interest alignment on an object graph of a first domain (an image containing only objects) against an object resource graph of a first domain (containing objects and resource data of interest to those objects), and to perform interest alignment on an object graph of a second domain against an object resource graph of a second domain. In other words, the embodiments of the present application design an intradomain alignment module to align interest relationships between objects within a single domain, bringing the feature representations of objects with similar interests or hobbies within a single domain closer together, thereby enabling the subsequent use of an interest alignment model to produce an accurate feature representation for an object based on the accurate feature representation of other objects, even if an object has no object interaction data or very little object interaction data, by analyzing the interest relationships between that object and other objects that have object interaction data.

[0043] In the model training solution provided by the embodiments of the present invention, the object interest alignment may include interdomain interest alignment and intradomain interest alignment, where interdomain interest alignment is the alignment of the interest similarity between two objects in a second domain with respect to the similarity between the same two objects in a first domain, and intradomain interest alignment is, for a single domain, the alignment of the interest similarity between two predicted objects with respect to the actual interest similarity between those two objects. In practical applications, the model training solution may include at least one of the two types of interest alignment given above, and for the sake of convenience, the subsequent embodiments will describe, in particular, the case in which the model training solution includes both interdomain and intradomain interest alignment.

[0044] Based on a brief introduction to the model structure of the interest alignment model by combining it with Figure 2 above, the following will introduce the method steps of the more detailed model training method proposed by the embodiment of this application, specifically the model training process and the model application process included in the model training method, by combining the drawings.

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

[0046] S301: Retrieve the data set.

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

[0048] The first and second field data in the data set may exist in graph form, where a graph refers to network structure data consisting of nodes and connecting edges (or abbreviated as edges). Nodes in this graph may include objects, first resource data, and second resource data, and connecting edges refer to connecting lines between nodes, specifically connecting object nodes to resource data nodes of interest to the objects. Considering that the data set includes multiple types of data, namely objects, first resource data, and second resource data, i.e., the sum of node types and connecting edge types is greater than 2, the graph constructed in the embodiment of this application is therefore a heterogeneous graph (also called a heterogeneous network). In this heterogeneous graph, the types of nodes and connecting edges are not singular but diverse.

[0049] A schematic diagram illustrating the existence of the first domain data and second neighbor data in the data set in the form of a heterogeneous graph can be found in Figure 2 given above, where the data set shown in Figure 2 includes three types of nodes: triangles, circles, and squares. Triangles represent the first resource data nodes in the first domain data, circles represent object nodes (for example, node u1 corresponds to object 1, node u2 corresponds to object 2, node u3 corresponds to object 3), and squares represent the second resource data nodes in the second domain data. A connecting edge between a triangular node and a circular node indicates that the object corresponding to the circular node is interested in the first resource data corresponding to the triangular node to which the connecting edge is connected. Similarly, a connecting edge between a circular node and a square node indicates that the object corresponding to the circular node is interested in the second resource data corresponding to the square node to which the connecting edge is connected.

[0050] To test the online effectiveness of a pre-trained interest alignment model (i.e., a cross-domain recommendation algorithm based on object interest alignment) (for example, if resource recommendations are to be made for objects under a cold-start scene, the recommendation effect of primarily cold-start objects will be observed when testing online), an embodiment of the present invention supports equipping the interest alignment model in the recall module of a resource recommendation platform. Thus, the interest alignment model may be trained using streaming updated object interaction data as sample data. In other words, an embodiment of the present invention acquires streaming updated object interaction data in a resource recommendation platform and uses it as the data set for model training, i.e., the first and second domain data included in the data set are dynamically updated, and the interest alignment model also changes dynamically when training the interest alignment model in accordance with the dynamically updated data set. This allows the interest alignment model to adapt to changes in the interests and preferences of objects in the resource recommendation platform, thereby ensuring that the interest alignment model can always produce accurate feature representations for objects in the resource recommendation platform and meet the dynamically changing personalized resource needs of those objects.

[0051] S302: Call the interest alignment model to extract features from the first domain data and obtain the first domain feature representation, then call the interest alignment model to extract features from the second domain data and obtain the second domain feature representation.

[0052] After obtaining a data set used to train an interest alignment model, the embodiments of the present invention support obtaining a first-domain feature representation of the first domain and a second-domain feature representation of the second domain by employing the data set to perform single-domain feature representation, that is, by performing feature extraction (also called feature representation) on the first domain data and the second domain data, respectively, based on the data set.

[0053] Domain feature representations of any one domain may be realized through embedding vectors, and therefore feature representations are also called embedding representations. An embedding vector represents a feature or attribute of an object using a single low-dimensional vector. The distance between the embedding vectors of any two objects may characterize the similarity between those two objects. For example, if the distance between the embedding vectors of two objects is less than a distance threshold, the two objects are relatively similar (e.g., their features or attributes are similar), and if the distance between the embedding vectors of the two objects is greater than or equal to the distance threshold, the degree of similarity between the two objects is relatively low. In embodiments of the present application, domain features of any one domain may include feature representations of objects in that domain and feature representations of resource data. For example, the first domain features of the first domain may include feature representations of objects and feature representations of first resource data, the object feature representation may be used to characterize the features or attributes of the object (e.g., the object's interests or hobbies within the first domain), and the first resource data feature representation may be used to characterize the features or attributes of the resource data (e.g., the resource type of the resource data). The distance between any two feature representations of objects may be used to characterize the similarity of interests between any two objects.

[0054] The process by which the embodiments of this application perform feature extraction on single-domain field data is similar. That is, the process of performing feature extraction on first-domain data to obtain first-domain feature representations is similar to the process of performing feature extraction on second-domain data to obtain second-domain feature representations. Below, the general feature extraction processes for the first-domain feature representations and second-domain feature representations are given, and then, using the second-domain feature representation as an example, the specific implementation process of information feature extraction on second-domain data is given in detail. The process of performing feature representation on first-field data and obtaining first-field feature representations may include constructing an object resource graph of the first field based on the first-field data, wherein the object resource graph of the first field is a graph in which each object and each first resource data are nodes and the first interest relationships between the objects and the first resource data are connected edges, and having a first interest relationship between an object and the first resource data means that the object produces interaction data with the first resource data; and calling an interest alignment model to perform graph coding on the object resource graph of the first field to obtain first object feature representations for each object and first resource feature representations for each first resource data, and constructing a first-field feature representation using the first object feature representations for each object and first resource feature representations for each first resource data.

[0055] The process of performing feature representation on second-domain data and obtaining second-domain feature representations may include, but is not limited to, steps (1) and (2).

[0056] (1) Construct an object resource graph of the second domain based on the second domain data. The object resource graph of the second domain is a graph in which each object and each second resource data are nodes, and the first interest relationships between objects and second resource data are connecting edges. The object resource graph of the second domain here may refer to a bigraph of the second domain (also called a bipartite graph, which is a special type of heterogeneous graph). As described above, the data set may be represented in the form of a heterogeneous graph, and for any one node in the heterogeneous graph, the nodes that have a connecting edge with it may be called neighbor nodes of that node, and the neighbor nodes of the node u1 shown in Figure 2 include two first resource data and two second resource data that have a connecting edge with the u1 node. Thus, the process of constructing an object resource graph of the second domain based on the second domain data here may include obtaining an object resource graph of the first domain by performing neighbor sampling on the heterogeneous graph of the data set. Here, neighbor sampling may include a process of selecting a portion of neighbor nodes from among all neighbor nodes of an object node in the heterogeneous graph of the data set (in this case, the first resource data node and the second resource data node). Thus, the embodiment of the present application supports performing neighbor sampling on each object node in the heterogeneous graph of the data set, thereby 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 Figure 4, the data set includes object nodes u1, u2, and u3. Furthermore, the neighbor nodes of object node u1 include first resource data node j1, first resource data node j2, second resource data node i1, and second resource data node i2. The neighbor nodes of object node u2 include first resource data node j2, first resource data node j3, and second resource data node i1. The neighbor nodes of object node u3 include first resource data node j3, first resource data node j4, second resource data node i1, and second resource data node i2. Thus, after performing neighbor sampling on the heterogeneous graph of the data set based on the second domain data, the neighbor nodes of object node u1 obtained by sampling are the second resource data node i1 and the second resource data node i2, the neighbor node of object node u2 is the second resource data node i1, and the neighbor nodes of object node u3 are the second resource data node i1 and the 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 of interest to each object node. Here, the connecting edges between an object node and a second resource data node in the object resource graph indicate that there is an interaction between the object corresponding to the object node and the second resource data corresponding to the second resource data node.

[0058] (2) An interest alignment model is invoked to perform graph coding 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 embodiment of the present invention uses a graph encoder to capture the 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 one exemplary graph coding learning architecture (or architecture of the second feature extraction module) may be seen in Figure 5. As shown in Figure 5, the second feature extraction module includes a graph attention network (GAT), and through the attention mechanism included in the graph attention network (GAT), feature representations can be made for each attention node in the input object resource graph of the second domain, where an attention node may refer to any one node in the object resource graph, for example, any one object node or any one 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 called to distinguish the importance of each neighbor node of each attention node, treating each object node and each second resource data node as attention nodes, 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 with which it has a connecting edge. Distinguishing the importance of each neighbor node can mean determining the degree of fondness (or interest or preference) of the object corresponding to the object node towards each second resource data. When the attention node is a second resource data node, the neighbor node refers to the object node with which it has a connecting edge. Distinguishing the importance of each neighbor node can mean determining the degree of fondness of the second resource data that the second resource data node corresponds to by the object corresponding to each object node.

[0060] The specific implementation process for calling a graph attention network to determine the feature representation of each attention node is similar, and the embodiment of the present application provides a specific implementation process in which, taking as an example, one of at least one objects included in the second field data is represented as a training object, the graph attention mechanism is called to distinguish the importance of each neighbor node of the training object, thereby providing the training object to obtain the corresponding second object feature representation. The implementation process may include, but is not limited to, steps s11 to s13. Here,

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

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

number

[0063] Here, a ui This indicates the degree of association (i.e., relationship) between the object node u and the second resource data node i, and h u The object node u represents the initial feature representation of the training object, and h i This shows the initial feature representation of the second resource data corresponding to the second resource data node i, and N uk represents the neighbor set of object node u, which includes all second resource data nodes that have a connecting edge with object node u, k represents 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 the LeakReLU function is an activation function.

[0064] As can be seen from formula (1) above, the embodiment of the present application calculates the relationship between a training object and each neighbor second resource data of the training object, for example, the similarity a between the object node u of the training object and the second resource data node i of the second resource data. ui It supports the calculation of the relationship between the object node u and the second resource data node i as the numerator, and the relationship between the object node u and all neighbor second resource data nodes N. u The denominator is the sum of the relationships with the neighbor node. Thus, using the relationship with the second resource data i as the numerator and the sum of the relationships with all neighbor nodes as the denominator is advantageous for quickly determining the proportion of the second resource data that the second resource data node i corresponds to 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: Based on the degree of relevance between the training object and each of the second resource data of interest to the training object, the initial feature representation of the training object in the second field, and the initial feature representation of each of the second resource data of interest to the training object, the first object feature representation of the training object is obtained. In other words, after obtaining the degree of relevance between the training object and each of the neighbor second resource data based on step s12, the degree of preference of the training object for each of the second resource data can be roughly determined, and in order to obtain the first object feature representation of the training object, i.e., the overall interest preference of the training object, it is also necessary to further specify the first object feature representation of the training object based on the degree of relevance between the training object and each of the neighbor second resource data, and the initial feature representation of the second resource data and the training object itself, thereby obtaining the overall interest preference of the training object (e.g., preference for certain types of resource data). The formula used to determine the second object feature representation of the training object is as follows.

number

[0066] Here,

number

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

[0068] The specific implementation process for calling the Graph Attention Network (GAT) to obtain the second resource feature representation for each second resource data node in the object resource graph of the second domain is similar to the specific implementation process for obtaining the first object feature representation described above, except that when calculating the second resource feature representation for the second resource data, the neighbor set consists of 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 for determining the second resource feature representation for the second resource data is not described in detail here. Similarly, the specific implementation process for 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 for the first object feature representation given above, and is not described in detail here.

[0069] The first feature extraction module, which corresponds to the first domain, is pre-trained; that is, during the pre-training phase, it has already been optimized by employing the abundant object interaction data in the first domain. Therefore, during the process of training the interest alignment model, it is not necessary to optimize the first feature extraction module according to the predicted representation of first domain features in the first domain data. However, because the object interaction data in the second domain is relatively scarce, transferring the pre-trained first feature extraction module during the model training process of the interest alignment model helps to better optimize the second feature extraction module.

[0070] Based on this, after obtaining 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 its interests in the second resource data is predicted, and the interest alignment model is trained based on the difference between the predicted affinity of the object for its interests in the second resource data and the actual affinity of the corresponding object for its interests in the second resource data. Specifically, the feature extraction performance of the second feature extraction model is improved by optimizing it, that is, the optimized second feature extraction model can predict more accurate object feature representations for the objects.

[0071] First, the embodiments of the present invention support obtaining a combined feature representation of each object by fusing the first object feature representation in the first field and the second object feature representation in the second field of each object. This enables the formation of a more complementary interest representation of the object by combining the interest features in the first field and the interest features in the second neighborhood of the object. For example, the first object feature representation in the first field of object node u is:

number

number

number

number

number

[0072] Subsequently, by connecting and manipulating the fused feature representation of each object with the second resource feature representation of the second resource data of interest to the corresponding object, we obtain the predicted level of attention (i.e., predicted level of liking) for the corresponding second resource data of interest to each object. The formula for calculating the predicted level of attention for the corresponding second resource data of interest to each object is as follows:

number

[0073] Here,

number

number

number

number

[0074] Finally, the data set includes the actual level of attention each object has to the corresponding second resource data of interest. In this way, the actual level of attention each object has to the corresponding second resource data of interest can be obtained from the data set. Based on the difference between the actual level of attention and the predicted level of attention each object has to the corresponding second resource data of interest, a first loss for the interest alignment model is constructed, which facilitates subsequent training of the interest alignment model based on this first loss. Embodiments of the present invention support learning the interest preferences of objects to the corresponding second resource data of interest through minimizing the squared error, the formula for which the squared error of interest is given as follows.

number

[0075] Here, R shows the object-second resource data interaction matrix,

number

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

number

[0077] S303: Call the interest alignment model and perform the following process: Based on the first and second domain feature representations, perform the interest alignment process between the first and second domains.

[0078] As can be seen from the above description, the embodiments of this application support interdomain object interest similarity alignment and / or object interest similarity alignment within a domain. Below, we will introduce specific implementation methods for these two types of interest similarity alignment.

[0079] (1) Inter-domain interest similarity alignment: The embodiment of this application designs a cross-domain alignment module (also called an inter-domain alignment module) based on the phenomenon that interests and hobbies are similar between the same objects in the inter-domain, that is, the phenomenon that two objects with similar interests in the first domain also exhibit similar interests in the second domain. Specifically, considering that the first domain has abundant object interaction data for the objects, and furthermore, that the first feature extraction model of the first domain is pre-trained (i.e., already has relatively good feature representation performance), the cross-domain alignment module is adopted to realize inter-domain object interest similarity alignment. So-called inter-domain object interest similarity alignment refers to using the interest similarity of two objects in the first domain as actual tags, and performing interest similarity alignment on the interest similarity of the two objects in the second domain relative to the interest similarity in the first domain, thereby guiding the learning of interest similarity in the second domain. In other words, the embodiment of the present application fully considers the interest relationships between objects, transfers the object interest similarity of the first domain to the second domain, that is, aligns the object interest similarity of the second domain to the source domain of the first domain, thereby achieving the objective of guiding interest learning in the second domain.

[0080] In concrete implementation, as can be seen from the steps above, the first domain feature representation includes the first object feature representation of each object in the first domain data, and the second domain feature representation includes the second object feature representation of each object in the second domain data. Below, using Figure 6, we will introduce a concrete implementation process in which the interest alignment model is invoked and interest alignment processing is performed between the first and second domains based on the first and second domain feature representations. In this process, the interest alignment processing includes cross-domain alignment processing, and the cross-domain alignment processing includes, but is not limited to, steps s21 to s22.

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

[0082] To account for differences in the degree of personalization within a single domain, the differences in object interest similarity between different domains are made relatively large. For example, some single domains (e.g., books or movies) have a relatively high degree of personalization (i.e., relatively large differences in resource types of resource data within that single domain), resulting in low object interest similarity (different objects' interests and preferences for the same resource data). Furthermore, for example, some single domains (e.g., news) have a relatively low degree of personalization (i.e., relatively small differences in resource types of resource data within that single domain), resulting in even higher object interest similarity (i.e., different objects' interests and preferences for the same resource data). Therefore, directly aligning the object interest similarity values ​​is influenced by the nature of the domain, thereby failing to reflect the actual interest similarity between objects. Based on this, embodiments of the present invention support eliminating the influence of the nature of the domain by converting the distance between object feature representations of objects into a probabilistic form to indicate interest similarity between objects.

[0083] Here, the specific implementation process for determining the degree of interest similarity between pairs of objects in the first and second fields is similar. Therefore, using the determination of the first degree of interest similarity between pairs of objects in the first field as an example, we will introduce the specific implementation process for determining the degree of first-order similarity between pairs of objects in the first field by calling up an interest alignment model and determining the degree of first-order similarity between pairs of objects in the first field based on the first-order feature representation of each object.

[0084] First, a distance operation is performed on the first object feature representations of pairs of objects from among the multiple objects to obtain distance information between the first object feature representations of pairs of objects. Specifically, the distance operation here is a Euclidean distance operation, meaning that 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 Figure 6, at least one object contains object u1, object u2, and object u3, and furthermore, the first object feature representation of object u1 is

number

number

number

number

number

number

number

number

number

number

number

number

[0085] Subsequently, the distance information between the first object feature representations of pairs of objects is transformed into a probability form to obtain the first degree of interest similarity between pairs of objects. In other words, by transforming the Euclidean distance between the object feature representations of objects into a probability form to show the degree of interest similarity, we can avoid the problem that the alignment results are affected by the degree of domain personalization and do not actually reflect the actual degree of interest similarity between objects. Here, the formula for transforming the Euclidean distance into a probability form to show the degree of interest similarity is as follows.

number

[0086] Here,

number

number

number

[0087] s22: Based on the specific implementation process shown in step s21 described above, the first degree of attentional similarity between pairs of objects in the first field can be obtained, and the first degree of attentional similarity between any two objects is used to characterize the similarity of interests and hobbies of those two objects in the first field. Based on the first degree of attentional similarity between each object in the first field, the attentional similarity distribution P of the first field can be determined, and this attentional similarity distribution P is a probability distribution composed of the first degree of attentional similarity between pairs of objects in the first field. Similarly, based on the specific implementation process shown in step s21 described above, the second degree of attentional similarity between pairs of objects in the second field can be obtained, and the second degree of attentional similarity between any two objects is used to characterize the similarity of interests and hobbies of those two objects in the second field. Based on the second degree of interest similarity between each object in the second field, the interest similarity distribution Q of the second field can be determined, which is a probability distribution composed of the second degree of interest similarity between pairs of 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 gap) between the attention similarity distribution P in the first domain and the attention similarity distribution Q in the second domain. In an embodiment of the present application, KL divergence is used as a loss function, and based on the difference between the attention similarity distribution P in the first domain and the attention similarity distribution Q in the second domain, the cross - domain alignment loss of the interest alignment model is constructed. The cross - domain alignment loss is used as the second loss, and thereby, by minimizing the cross - domain alignment loss, the optimization of the interest alignment model is supported. Here, KL divergence, also known as relative entropy, is a measure of the asymmetry of the difference between two probability distributions. In an embodiment of the present application, the calculation formula for obtaining the cross - domain alignment loss of the interest alignment model by using KL divergence as a loss function is as follows.

Equation

[0089] Here, P represents the attention similarity distribution in the first domain, and p uiuj represents the first attention similarity degree between the object ui and the object uj in the first domain. Q represents the attention similarity distribution in the second domain, and q uiuj represents the second attention similarity degree between the object ui and the object uj in the second domain.

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

[0091] (2) Intradomain interest similarity alignment: The embodiments of the present invention primarily involve aligning the interest similarity of objects within a domain using a graph decoding target, thereby bringing the feature representations of objects with similar interests or hobbies within a single domain closer together. This is intended to ensure the accuracy of the object feature representation of a cold-start object when applying a model, even if the object of the resource data to be delivered is a cold-start object, by analyzing the object feature representations of other objects with similar interests or hobbies to the cold-start object.

[0092] Considering that the domain data (in bigraph format) input to the single-domain feature extraction module includes interest similarity relationships for each object within a single domain, as shown in Figure 5, object nodes u1 and u2 are both connected to the 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 closer, the embodiment of this application supports alignment of intra-domain object interest similarity. Alignment of so-called intra-domain object interest similarity refers to using the input information (single-domain domain data) of the single-domain feature extraction module as actual tags. Specifically, the interest similarity between two objects in the domain data station input to the single-domain feature extraction module is the actual tag, and the predicted interest similarity between the same two objects within a single domain is aligned to it. In other words, the interests of intra-domain objects are aligned, making the object feature representations of objects with similar interests or hobbies within the domain closer. Based on this, embodiments of the present application hope to be able to reconstruct interest relationship graphs between objects based on the fused feature representation of objects, including an interest relationship graph between objects in a first domain and an interest relationship graph between objects in a second domain, wherein the interest relationship graph is an object-object graph containing only object nodes, thereby enabling intra-domain alignment learning based on the reconstructed object-object graph and the object resource graph input to the model.

[0093] The following describes a specific implementation process in which the interest alignment model is invoked using Figure 7, and interest alignment processing is performed in the first and second domains based on the first and second domain feature representations, respectively. In this process, 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 the fused feature representation of each object, the object resource graph of the first domain, and the object resource graph of the second domain. Here, the fused feature representation of any one object is obtained by fusing the first object feature representation in the first domain and the second object feature representation in the second domain of that object. By fusing the first and second object feature representations of an object, a complementary feature representation of the object can be formed, reducing the problem that leads to 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, and the related contents of the object resource graph may be described in the related description above and will not be explained in detail here.

[0095] s32: The interest alignment model (specifically, the intra-domain alignment module included in the interest alignment model) is called to perform graph decoding on the first domain and graph decoding on the second domain for each object's fused feature representation, thereby obtaining the object graph for the first domain and the object graph for the second domain. The object graph here is the object-object graph described above, and this object graph is a graph in which each object is a node and the second interest relationship between the objects is a connecting edge. Having a second interest relationship between objects can mean that two objects produce interaction data for the same resource data.

[0096] As an example, to reconstruct the object graphs of the first and second domains, the present invention supports obtaining the graph decoding weights of the first domain and calling an interest alignment model to reconstruct the 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 obtaining the graph decoding weights of the second domain and calling an interest alignment model to reconstruct the object graph of the second neighborhood based on the graph decoding weights of the second domain and the fused feature representation of each object. In other words, the embodiment of the present invention supports reconstructing a single-domain object graph using a graph decoder, thereby predicting whether two objects will produce interaction data for the same resource data within a single domain through the object graph, and if the two objects produce interaction data for the same resource data, it is hoped that a connecting edge exists between the two objects in the object graph, in which case the two objects are recognized as having similar interests or tastes, and thereby the object feature representations of these two objects become closer (i.e., the distance between the embedding vectors becomes closer).

[0097] Each domain corresponds to a different graph decoder, specifically, this manifests as different graph decoding weights for the graph decoders. Here, the graph decoder is

number

number

[0098] Here,

number

number

number

number

number

number

[0099] As shown in Figure 8, after the first domain inputs the fused feature representation of the object into the corresponding graph decoder, the object graph G of the first domain is generated. SThis can be reconstructed and obtained, and as can be seen by comparing the object resource graph and object graph of the first domain, the quadratic relationships of objects in the first domain are converted to linear relationships (i.e., the metapath in the object resource graph: object-first resource data-object is converted to the metapath in the object graph: object-object. A metapath can refer to a specific path connecting two or more nodes, and may be used to describe complex relationships between nodes, and the definitions explicitly stated by different metapaths are different), thereby enabling the description of the similarity of interest between two objects in the first domain. Similarly, after the feature fusion representation of the objects is input into the corresponding graph decoder of the second domain, the object graph G of the second domain is obtained. T It can be obtained by reconstructing the object graph G T Related information can be found in the object graph G above. S You may refer to related descriptions, which will not be explained in detail here.

[0100] s33: Intra-domain alignment is achieved by performing alignment processing on the object graph of the first domain with respect to the object resource graph of the first domain, and on the object graph of the second domain with respect to the object resource graph of the second domain. Specifically, after reconstructing the object graphs of the first and second domains based on the steps described above, the object graphs are used as predicted object-object relationships, and the object resource graphs of the corresponding domains are used as actual object-object relationships. This supports training the interest alignment model according to the difference between the actual object-object relationships and the predicted object-object relationships.

[0101] As an example, the intra-domain alignment loss for the first domain is obtained based on the differences 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 differences between the object graph of the second domain and the corresponding object resource graph. Then, considering that there are object graphs for multiple domains, the intra-domain alignment loss for the first domain and the intra-domain alignment loss for the second domain are merged in this way to obtain the intra-domain alignment loss of the interest alignment model, and this intra-domain alignment loss is taken as the second loss. As described above, graph reconstruction is a binary classification task, and in this way, the adoption of the binary classification cross-entropy loss as the graph reconstruction loss is supported. The intra-domain alignment loss of the interest alignment model constructed under the multi-domain (i.e., first and second domains) case is as follows.

number

[0102] Here,

number

number

number

number

number

[0103] Here,

number

number

number

number

number

number

number

number

number

number

number

number

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

[0105] It is worth noting that, considering the relatively large number of objects and resource data contained in a data set (or first and second domain data), the embodiments of this application support dividing single-domain domain data into multiple data subsets (batches) to reduce computational complexity. Each data subset contains some of the objects and resource data, and thus an object graph can be constructed of some of the object nodes rather than constructing an object graph based on all objects. This reduces computational complexity and improves the efficiency and speed of model training. For example, assuming that one data subset batch contains n objects, n ≤ N, and N is all objects, the computational complexity is O(N) compared to constructing an object graph based on all objects 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 and performing the following processes, namely, an interest alignment process between the first and second domains based on the first and second domain feature representations, can be achieved through the following technical solutions: Calling the interest alignment model and performing the following processes, namely, a cross-domain alignment process between the first and second domains based on the first and second domain feature representations. Calling the interest alignment model and performing the following processes, namely, an intra-domain alignment process between the first and second domains based on the first and second domain feature representations.

[0107] The implementation method for performing a cross-domain alignment process between the first and second domains based on the first and second domain feature representations may refer to the implementation methods of steps s21 to s22, and the implementation method for performing an intra-domain alignment process between the first and second domains based on the first and second domain feature representations may refer to the implementation methods 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: The interest alignment model is trained in a manner that reduces 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 target of the model training method provided in the embodiment of the present application can be divided into three parts: a) Learning of a recommendation target in the second domain, i.e., learning the interest preferences of an object in the second domain in accordance with object interaction data between the object and the second resource data. This can be obtained by constructing the first loss of the interest alignment model under the learning of this target. b) Learning of 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 to the first domain, thereby achieving the objective of guiding interest learning in the second domain. This can be obtained by constructing the cross-domain alignment loss of the interest alignment model under the learning target. c) Learning of an intra-domain interest alignment target, i.e., using a graph decoding target to align the interest preferences of objects within a domain, with the intention of making the object feature representations of objects with similar interests within the domain closer together. This can be obtained by constructing the in-domain alignment loss of the interest alignment model under the learning of this target.

[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 processing process may include at least one of the cross-domain alignment loss and the intra-domain alignment loss. It needs to be explained that, depending on the differences in the modules included in the interest alignment model, the process of training the interest alignment model differs in a direction that reduces the first loss in the feature extraction process and the loss in the interest alignment processing process.

[0112] The interest alignment model may include only a single-domain feature extraction module and a cross-domain alignment module. Under this implementation, the process of training the interest alignment model in a direction that reduces the first loss of the feature extraction process and the loss of the interest alignment process may include joint training of the interest alignment model by adopting the first loss of the interest alignment model and the cross-domain alignment loss of the interest alignment model obtained based on the steps described above. More specifically, the first loss of the interest alignment model and the cross-domain alignment loss of the interest alignment model are added together to obtain the target loss of the interest alignment model, and the interest alignment model is trained in a direction that reduces the above target loss. Here, the formula for calculating the target loss is as follows.

number

[0113] L represents the overall loss of the interest alignment model.

number

number

[0114] The interest alignment model may include only a single-domain feature extraction module and an intra-domain alignment module. Under this implementation, the process of training the interest alignment model in a direction that reduces the first loss of the feature extraction process and the loss of the interest alignment process may include joint training of the interest alignment model by adopting the first loss of the interest alignment model and the intra-domain alignment loss of the interest alignment model obtained based on the steps described above. More specifically, the first loss of the interest alignment model and the intra-domain alignment loss of the interest alignment model are added together to obtain the target loss of the interest alignment model, and the interest alignment model is trained in a direction that reduces the target loss. Here, the formula for calculating the target loss is 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. Under this implementation, the process of training the interest alignment model in a direction that reduces the first loss of the feature extraction process and the loss of the interest alignment process may include joint training of the interest alignment model by adopting 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 based on the steps described above. More specifically, the first loss of the interest alignment model, the cross-domain alignment loss, and the intra-domain alignment loss are added together to obtain a target loss of the interest alignment model, and the interest alignment model is trained in a direction that reduces the target loss. Here, the formula for calculating the target loss is as follows.

number

[0117] As can be seen from the above, the embodiment of the present application supports training an interest alignment model by adopting a joint training method, and specifically, by adjusting the model parameters of each module included in the interest alignment model, the predictive performance of the interest alignment model after the model parameters have been adjusted is made 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, since the embodiment of the present application is intended to guide the learning of the second domain using the object interaction data of the first domain, in the process of training the interest alignment model, it supports interrupting gradient backpropagation for the first feature extraction module based on the cross-domain alignment loss, using the attention similarity distribution P of the first domain as an anchor, in accordance with the direction of decrease in the first loss of the feature extraction process and the loss of the interest alignment processing process. In other words, it does not 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 intradomain alignment loss so that the first domain feature representations of the first domain contribute to the first loss of the second domain.

[0118] Embodiments of the present invention employ an object interest similarity alignment method, taking into full consideration the interest similarity between objects, to train an interest alignment model, thereby enabling the interest alignment model to produce relatively accurate feature representations for objects even when applied to domains with relatively little interaction data. For example, considering that the interaction data in the first domain is relatively abundant, the feature representations of the first domain extracted based on the abundant interaction data become more accurate. Therefore, the interest alignment model is trained by employing a method that aligns the interest similarity of two objects in a second domain with relatively little interaction data to the interest similarity of the two objects in the first domain with relatively abundant interaction data. Even when the interaction data in the second domain is relatively small, it is ensured that the trained interest alignment model can produce relatively accurate feature representations for objects, thereby ensuring that the resource data recommended for objects based on accurate feature representations meets the object personalization needs and achieves high-precision resource recommendations. The embodiment of this invention employs a method of joint training the interest alignment model using the first loss in the feature extraction process and the loss in the interest alignment processing process, thereby enriching the overall learning targets of the interest alignment model and ensuring the performance of the interest alignment model through multi-target learning.

[0119] The embodiment shown in Figure 3 above provides a detailed explanation of the model training portion of the interest alignment model, and the model application portion of the interest alignment model will be introduced below. In concrete implementation, after obtaining a trained interest alignment model based on the steps described above, the embodiment of this application supports equipping the trained interest alignment model in a resource recommendation platform, and specifically applies it to the resource recall stage of the resource recommendation platform. Therefore, the interest alignment model may also be called a resource recall model. The resource recall stage mainly facilitates the subsequent high-precision sorting of the selected resource data from a large resource database that is of interest to an object, and recommends it to the object. A schematic diagram of one exemplary flow for applying the interest alignment model to perform 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 a resource delivery need, the object features (or attribute information, e.g., age, gender, etc.) of the target object can be 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. Subsequently, in the resource recall stage, the trained interest alignment model in the embodiment of this application is employed to generate feature representations for the target object and for the resource data in the resource recommendation platform. The feature representations of the target object and the feature representations of each resource data are matched, specifically calculating the similarity between the feature representations of the target object and the resource data. Then, based on the similarity results, multiple resource data are selected and placed into a precisely sorted candidate pool for use in the subsequent sorting stage.

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

[0121] S901: Retrieve the data set.

[0122] S902: Call the interest alignment model to extract features from the first domain data and obtain the first domain feature representation, then call the interest alignment model to extract features from the second domain data and obtain the second domain feature representation.

[0123] S903: Call the interest alignment model and perform interest alignment processing in the first and second domains based on the first and second domain feature representations.

[0124] S904: The interest alignment model is trained in a manner that reduces 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 process shown in steps S901 to S904 may be described by referring to the relevant description of the specific implementation process shown in steps S301 to S304 in the embodiment shown in Figure 3 above, and will not be explained in detail here.

[0126] S905: Call up a pre-trained interest alignment model to extract features from the target object to be delivered and obtain a fused feature representation of the target object.

[0127] After equipping the resource recommendation platform with a trained interest alignment model, if a target object for the resource data to be delivered exists in the resource recommendation platform, and if the target object is an object that has just been registered in the resource recommendation platform, the object attributes of the target object for the resource data to be delivered can be obtained. These object attributes may include relevant attribute information entered when the target object was registered in the resource recommendation platform (e.g., a set nickname, age, gender, or selected resource type tag). Subsequently, the trained interest alignment model is invoked to perform feature extraction on the object attributes of the target object to obtain a fused feature representation of the target object.

[0128] Specifically, the first feature extraction module included in the interest alignment model is called to predict the first object feature representation in the first domain of the target object, then the second feature extraction module included in the interest alignment model is called to predict the second object feature representation in the second domain of the target object, and then the first and second object feature representations of the target object are merged to obtain the merged feature representation of the target object. Here, the specific implementation process by which the first feature extraction module predicts the first object feature representation of the target object (or the second feature extraction module predicts the second object feature representation of the target object) can be described by referring to the relevant description in the embodiment shown in Figure 3 above, and will not be explained in detail here.

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

[0130] S907: Candidate resource data whose similarity comparison result is greater than the comparison result threshold will be used as the second resource data to be delivered to the target object.

[0131] In steps S906 to S907, based on the above steps, a fused feature representation of the target object of the resource data to be distributed can be obtained, and this fused feature representation may be used to characterize the attributes or features of the target object (for example, it can be used to characterize that the target object is interested in a certain type of resource data). In this way, a resource feature representation of the candidate resource data to be distributed in the second domain 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, it can be used to characterize the case of a trigger within the historical time of the candidate resource data).

[0132] Considering that candidate resource data whose resource feature representation is closer to the fused feature representation of the target object is more likely to be resource data of interest to the target object, the system supports matching the similarity between the fused 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. 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 the resource recommendation rule is designated as the resource data to be delivered to the target object. When the interest alignment model is equipped in the resource recall stage of the resource recommendation platform, candidate resource data whose similarity comparison result is greater than the comparison result threshold can be placed in a precisely sorted candidate pool as resource data to be delivered, thereby facilitating further recommendations by subsequent precise sorting. Here, the resource recommendation rule may be customized by the operator according to their business needs. For example, the resource recommendation rule includes designating candidate resource data whose similarity comparison result is greater than the result threshold as the resource data to be delivered to the target object. Furthermore, for example, the resource recommendation rule may include sorting the result values ​​from largest to smallest and selecting the first k candidate resource data as the resource data to be delivered associated with the target object. The embodiments of this application are not limited to the specific content of the resource recommendation rule.

[0133] For an example schematic diagram of matching the fused feature representation of the target object with the resource feature representation of each candidate resource data, see Figure 10. As shown in Figure 10, the candidate resource data to be distributed in the second domain includes candidate resource data 1, candidate resource data 2, candidate resource data 3, and candidate resource data 4. The similarity comparison result between the fused feature representation of the target object and the resource feature data of candidate resource data 1 is 60% (or expressed in numerical form such as a fraction or decimal), the similarity comparison result between the fused feature representation of the target object and the resource feature data of candidate resource data 1 is 50%, the similarity comparison result between the fused feature representation of the target object and the resource feature data of candidate resource data 1 is 20%, and the similarity comparison result between the fused feature representation of the target object and the resource feature data of candidate resource data 1 is 70%. If the resource recommendation rule includes sorting the result values ​​from largest to smallest and selecting the first k (k=2) candidate resource data as the resource data to be delivered to the target object, then a sort among the 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 this sort can be selected as the resource data to be delivered to the target object.

[0134] In practical applications, a trained interest alignment model of the embodiment of this application may be equipped in the recall module of a resource recommendation platform and used to perform resource recommendations for millions of online objects in the resource recommendation platform. Here, the results of evaluating the model effect of the interest alignment model using key metrics for determining whether the model is accurately recommended in the industrial sector may be found 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 metric is obtained by comparing the number of objects with click behavior with the number of objects that visit the resource data. After recommending resource data using the interest alignment model provided in the embodiment of this application, if the number of objects that visit the resource data is 100 and the number of users with click behavior for the resource data is 13, then 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, and as shown in Table 1, X may take values ​​of 180, 60, or 30 seconds. Pctr (Page-click-through-rate) may be called "page click-through rate." dau (daily-active-users) may be abbreviated as "daily active users," and may be used to reflect the case of objects that are active every day in the resource recommendation platform. As described above, after equipping the resource recommendation platform with the trained interest alignment model of the embodiment of this application, it can produce object feature representations for objects with higher accuracy, thereby recommending resource data for objects with higher accuracy and further promoting the widespread adoption of the resource recommendation platform.

[0137] It is understood that, in the embodiments of this application, relevant data such as user information is involved, and when the embodiments of this application are applied to specific products or technologies, it is necessary to obtain the user's permission or consent. Furthermore, 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 describes in detail the method of the embodiment of the present application, and in order to facilitate the better implementation of the above method of the embodiment of the present application, the apparatus of the embodiment of the present application is provided below accordingly.

[0139] Figure 11 shows a schematic diagram of the structure of one type of model training device provided by one exemplary embodiment of the present application, which may be a single computer executable instruction (including program code) operated on an electronic device. The model training device may be used to perform some or all of the steps in the embodiment of the method shown in Figures 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, the data set including first field data and second field data, wherein the first field data includes at least one object and first resource data of interest within the first field of each object, and the second field data includes at least one object and second resource data of interest within the second field of each object. The processing unit 1102 is configured to call an interest alignment model to extract features from the first domain data and obtain a first domain feature representation, and then call the interest alignment model to extract features from the second neighbor data and obtain a second domain feature representation. The processing unit 1102 is further configured to call the interest alignment model to perform the following processing, namely, an interest alignment process between the first and second domains based on the first and second domain feature representations. The processing unit 1102 is further configured to train the interest alignment model in a direction that reduces the first and second losses and obtain a trained interest alignment model. Here, the first loss is the loss corresponding to feature extraction, the second loss is the loss corresponding to the interest alignment process, and the trained interest alignment model is used to recommend resource data for target objects in the second domain.

[0140] In some embodiments, the processing unit 1102 is further configured to construct a first-field object resource graph based on first-field data, where each object and each first-field resource data is a node, and the first interest relationships between the objects and the first-field resource data are connected edges. An interest alignment model is invoked to perform graph coding on the first-field object resource graph to obtain the first object feature representation for each object and the first resource feature representation for each first-field resource data. The first-field feature representation is constructed using the first object feature representation for each object and the first resource feature representation for each first-field resource data.

[0141] In some embodiments, the processing unit 1102 is further configured to construct a second-field object resource graph based on second-field data, where each object and each second-field resource data is a node, and the first interest relationships between the objects and the second-field resource data are connected edges. An interest alignment model is invoked to perform graph coding on the second-field object resource graph to obtain the second-object feature representation for each object and the second-source feature representation for each second-field resource data. The second-field feature representation is constructed using the second-object feature representation for each object and the second-source feature representation for each second-field resource data.

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

[0143] In some embodiments, the data set includes the actual level of attention each object has to the corresponding second resource data of interest, and the process of obtaining the first loss of the feature extraction process includes: fusing the first object feature representation and the second object feature representation of each object to obtain a fusing feature representation of each object; connecting and manipulating the fusing feature representation of each object with the second resource feature representation of the corresponding second resource data of interest of the object to obtain a predicted level of attention each object has to the corresponding second resource data of interest; and constructing the first loss of the interest alignment model based on the difference between the actual level of attention and the predicted level of attention each object has to the 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 the first object feature representation of each object in the first domain data, and the second domain feature representation includes the second object feature representation of each object in the second domain data. The processing unit 1102 is further configured to call an interest alignment model to determine the first degree of interest similarity between pairs of objects in the first domain based on the first object feature representation of each object, and to call an interest alignment model to determine the second degree of interest similarity between pairs of objects in the second domain based on the second object feature representation of each object, and to perform an alignment process on the second degree of interest similarity between two objects in the second domain with respect to the first degree of interest similarity between corresponding two objects in the first domain.

[0145] In some embodiments, the processing unit 1102 is further configured to perform distance calculations on the first object feature representations of pairs of objects from among a plurality of objects, obtain distance information between the first object feature representations of pairs of objects, perform probability transformation on the distance information between the first object feature representations of pairs of objects, and obtain a first degree of similarity of interest between pairs of objects.

[0146] In some embodiments, the process of obtaining the loss of the interest alignment process includes determining the attention similarity distribution of the first domain based on the first attention similarity between pairs of objects in the first domain, determining the attention similarity distribution of the second domain based on the second attention similarity between pairs of objects in the second domain, 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 taking the cross-domain alignment loss as the second loss.

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

[0148] In some embodiments, the processing unit 1102 is further configured to obtain graph decoding weights for the first domain, and to call an interest alignment model to reconstruct the object graph for the first domain based on the graph decoding weights for the first domain and the fused feature representation of each object, and to obtain graph decoding weights for the second domain, and to call an interest alignment model to reconstruct the object graph for the second domain based on the graph decoding weights for 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 process includes obtaining an in-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 in-domain alignment loss of the interest alignment model; and setting the cross-domain alignment loss as the second loss.

[0150] In some embodiments, the interest alignment process includes cross-domain alignment and intra-domain alignment, and the processing unit 1102 is further configured to call an interest alignment model to perform the following processes: cross-domain alignment between the first and second domains based on the first and second domain feature representations, and intra-domain alignment between the first and second domains based on the first and second domain feature representations.

[0151] In some embodiments, the process of obtaining a second loss in the interest alignment process includes obtaining a cross-domain alignment loss corresponding to the cross-domain alignment process, 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.

[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. The first feature extraction module is used to extract features from first domain data and is obtained through pre-training. The second feature extraction module is used to extract features from second domain data. The cross-domain alignment module is used to perform interest alignment on the second degree of interest similarity between two objects in the second domain and the first degree of interest similarity between two corresponding objects in the first domain. The intra-domain alignment module is used to perform interest alignment on the object graph of the first domain and on the object resource graph of the first domain, and on the object graph of the second domain and on the object resource graph of 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 for the interest alignment model, and to train the interest alignment model in a direction that reduces the target loss, where the training of the first feature extraction module based on the cross-domain alignment loss is interrupted during the process of training the interest alignment model in a direction that reduces the target loss.

[0154] In some embodiments, the processing unit 1102 is further used to obtain the object attributes of the target object to be distributed and the resource feature representation of each candidate resource data to be distributed in the second domain; to call a trained interest alignment model to extract features based on the object attributes of the target object and obtain a fused feature representation of the target object; to perform a similarity comparison between the fused 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 to set candidate resource data that satisfy the resource recommendation rule as the second resource data to be distributed associated with the target object.

[0155] In some embodiments, each unit in the model training device shown in Figure 11 may be configured by merging each or all of them into one or several other units, or some of the units may be further divided into several functionally smaller units, which can achieve similar operations but does not affect the realization of 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 several units, or the function of several units may be realized by one unit. In other embodiments of the present application, the model training device may include other units, and in practical applications, these functions may be realized in conjunction with other units, or in cooperation with several units. In another embodiment of the present application, a model training apparatus as shown in Figure 11 may be configured and a model training method of the embodiment of the present application may be implemented by running a computer program (including program code) capable of performing each of the relevant steps in the corresponding methods shown in Figures 3 and 9 on a general-purpose electronic device, such as a computer, which includes processing elements and memory elements such as a central processing unit (CPU), a random access storage medium (RAM), and a read-only storage medium (ROM). The computer program may be written on, for example, a computer-readable recording medium, and installed on the electronic device via the computer-readable recording medium and run thereon.

[0156] The embodiments of this application fully consider the similarity of interests between objects and employ an object interest similarity alignment method to support training an interest alignment model, enabling the interest alignment model to produce relatively accurate feature representations for objects even when applied to domains with relatively little interaction data. For example, considering that the first domain has a relatively large amount of interaction data, and that the first domain feature representation extracted based on this rich interaction data is more accurate, the interest alignment model is trained by aligning the similarity of interests between two objects in a second domain with relatively little interaction data to the similarity of interests between the two objects in a first domain with relatively large interaction data. Even when the interaction data in the second domain is relatively small, the trained interest alignment model can produce relatively accurate feature representations for objects, thereby ensuring that the resource data recommended for the objects based on accurate feature representations meets the personalization needs of the objects and achieving high-precision resource recommendations. The embodiment of this invention employs a method of joint training the interest alignment model using the first loss from the feature extraction process and the loss from the interest alignment processing process. This allows for a richer set of learning targets for the entire interest alignment model, and ensures the performance of the interest alignment model through multi-target learning.

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

[0158] Embodiments of the present invention further provide a computer-readable memory medium, which is a storage device in an electronic device used to store programs and data. Understandably, the computer-readable memory medium herein may include embedded memory in an electronic device, but may also include extended memory supported by the electronic device. The computer-readable memory medium provides a memory space in which the processing system of the electronic device is stored. Furthermore, the memory space also stores one or more instructions suitable for loading and executing by the processor 1201, which may be one or more computer programs (including program code). It should be noted that the computer-readable memory medium herein may be a high-speed RAM memory device, a non-volatile memory device, such as at least one magnetic disk memory device, or at least one computer-readable memory medium located away from the aforementioned processor.

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

[0160] Embodiments of the present application further provide a computer program product which includes computer executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer executable instructions from the computer-readable storage medium and causes the electronic device to perform the model training method by executing the computer executable instructions.

[0161] As will be apparent to those skilled in the art, each exemplary unit and algorithmic step described in the embodiments disclosed herein can be combined to be implemented as electronic hardware, or as a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application of the technical solution and the design constraints. While experts in the art may implement the described functions using different methods for each specific application, such implementations should not be considered beyond the scope of this application.

[0162] In the embodiments described above, implementation can be carried out in whole or in part through software, hardware, firmware, or any combination thereof. When implemented using software, it may be carried out in whole or in part in the form of a computer program product. A 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 function, in whole or in part, according to embodiments of the present invention is produced. The computer may be a general-purpose computer, a dedicated 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 a wired (e.g., coaxial cable, optical fiber, digital user line (DSL)) or wireless (e.g., infrared, radio, microwave, etc.) method. The computer-readable storage medium may be any available medium that a computer can access, or it may be a data processing device such as a server or data center that integrates 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 foregoing describes only specific embodiments of the present application; however, the scope of protection of the present application is not limited thereto. Any changes or substitutions that a person skilled in the art could easily conceive of within the scope of the present invention are included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be the same as the scope of protection of the above claims. [Explanation of symbols]

[0164] 101 terminals 102 Server 103 Server 1021 Database 1031 Database 1101 Acquisition Unit 1102 Processing Unit 1201 Processor 1202 Communication Interface 1203 Computer-readable Storage Medium

Claims

1. A model training method, wherein the method is performed by an electronic device. A step of obtaining a data set, wherein the data set includes first field data and second field data, the first field data includes at least one object and first resource data of interest within the first field of each object, and the second field data includes at least one object and second resource data of interest within the second field of each object. The steps include calling up an interest alignment model to extract features from the first domain data and obtain a first domain feature representation, and then calling up the interest alignment model to extract features from the second domain data and obtain a second domain feature representation, The steps include calling the aforementioned interest alignment model and performing the following process, namely, an interest alignment process between the first and second domains based on the first and second domain feature representations, The process includes the step of training the interest alignment model in a direction that reduces the first loss and the second loss, and obtaining a trained interest alignment model. A model training method 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 domain.

2. The step of calling up an interest alignment model, performing feature extraction on the first domain data, and obtaining a first domain feature representation is: A step of constructing an object resource graph of the first domain based on the first domain data, wherein the object resource graph of the first domain is a graph in which each object and each first resource data are nodes and the first interest relationships between the objects and the first resource data are connecting edges. The steps include calling an interest alignment model and performing graph coding 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, The method according to claim 1, comprising the step of constructing a first field feature representation using a first object feature representation of each of the aforementioned objects and a first resource feature representation of each of the aforementioned first resource data.

3. The step of calling up the interest alignment model, performing feature extraction on the second domain data, and obtaining a second domain feature representation is: A step of constructing an object resource graph of the second domain based on the second domain data, wherein the object resource graph of the second domain is a graph in which each object and each second resource data are nodes and the first interest relationships between the objects and the second resource data are connecting edges. The steps include calling the interest alignment model and performing graph coding on the object resource graph of the second domain to obtain the second object feature representation of each object and the second resource feature representation of each second resource data, The method according to claim 1, comprising the step of constructing the second field feature representation with the second object feature representation of each of the aforementioned objects and the second resource feature representation of each of the aforementioned second resource data.

4. The step of representing any one of the at least one of the aforementioned objects as a training object, calling the interest alignment model to perform graph coding on the object resource graph of the second domain, and obtaining a second object feature representation of each of the aforementioned objects, The steps include obtaining the initial feature representation of the training object in the second domain and the initial feature representation of each of the second resource data of interest to the training object, The process involves calling the interest alignment model and performing the following steps: calculating the degree of association between the training object and each of the second resource data of interest to the training object, according to the initial feature representation of the training object and the initial feature representation of each of the second resource data of interest to the training object; The method according to claim 3, comprising the step of determining a second object feature representation of the training object based on the degree of relevance between the training object and each of the second resource data of interest to the training object, the initial feature representation of the training object in the second field, and the initial feature representation of each of the second resource data of interest to the training object.

5. The data set includes the actual level of attention each object has to the corresponding second resource data of interest, and the method is as follows: The steps include: fusing the first object feature representation and the second object feature representation of each of the aforementioned objects to obtain a fused feature representation of each of the aforementioned objects; The steps include: connecting and manipulating the fused feature representation of each of the aforementioned objects with the second resource feature representation of the second resource data of interest to the corresponding object, and obtaining the predicted level of attention for the corresponding second resource data of interest to each of the aforementioned objects; The method according to claim 3, further comprising the step of constructing a first loss of the interest alignment model based on the difference between the actual and predicted levels of attention for the corresponding second resource data of interest for each of the aforementioned objects.

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, 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 and performing the following process, namely, an interest alignment process between the first domain and the second domain based on the first domain feature representation and the second domain feature representation, The steps include calling the interest alignment model and determining a first interest similarity between pairs of objects in the first domain based on the first object feature representation of each of the objects, The steps include calling the interest alignment model and determining the second interest similarity between pairs of objects in the second domain based on the second object feature representation of each object, The method according to claim 1, comprising the step of performing an alignment process on a second similarity of interest between two objects in the second field with respect to a first similarity of interest between two corresponding objects in the first field.

7. The step of calling the interest alignment model and determining a first interest similarity between pairs of objects in the first domain based on the first object feature representation of each of the objects is: The steps include performing a distance calculation on the first object feature representations of two objects from among the multiple objects, and obtaining distance information between the first object feature representations of the two objects, The method according to claim 6, comprising the step of performing a probability transformation on distance information between the first object feature representations of the pairs of objects to obtain a first attention similarity between the pairs of objects.

8. The above method further, A step of determining the attention similarity distribution of the first field based on the first attention similarity between pairs of objects in the first field, The steps include determining the attention similarity distribution of the second field based on the second attention similarity between pairs of objects in the second field, The method according to claim 6, comprising the steps of constructing a cross-domain alignment loss of the interest alignment model based on the difference between the interest similarity distribution of the first domain and the interest similarity distribution of the second domain, and then using the cross-domain alignment loss as the second loss.

9. The interest alignment process includes an intradomain alignment process, and the step of calling the interest alignment model and performing the following process, namely, an interest alignment process between the first and second domains based on the first and second domain feature representations, The steps include obtaining the fused feature representation of each object, the object resource graph of the first domain, and the object resource graph of the second domain, The steps include: calling the interest alignment model and performing graph decoding on the first domain and graph decoding on the second domain for each of the fused feature representations of the objects, respectively, to obtain an object graph of the first domain and an object graph of the second domain, wherein the object graph is a graph in which each object is a node and the second interest relationships between the objects are connecting edges; The method according to claim 1, comprising the steps of performing alignment processing on the object graph of the first field with respect to the object resource graph of the first field, and performing alignment processing on the object graph of the second field with respect to the object resource graph of the second field.

10. The step of calling the interest alignment model and performing graph decoding for the first domain and graph decoding for the second domain on the fused feature representation of each object, respectively, in order to obtain the object graph of the first domain and the object graph of the second domain, is: The steps include obtaining the graph decoding weights of the first domain, and calling the interest alignment model to reconstruct the object graph of the first domain based on the graph decoding weights of the first domain and the fused feature representation of each object, The method according to claim 9, comprising the steps of obtaining graph decoding weights for the second domain, and calling the interest alignment model to reconstruct the object graph for the second domain based on the graph decoding weights for the second domain and the fused feature representation of each of the objects.

11. The above method further, A step of obtaining the 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, A step of obtaining the intra-domain alignment loss of the second domain based on the difference between the object graph of the second domain and the corresponding object resource graph, The method according to claim 9, comprising the steps of merging the intradomain alignment loss of the first field and the intradomain alignment loss of the second field to obtain the intradomain alignment loss of the interest alignment model, and setting the intradomain alignment loss to 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 and performing the following process, namely, an interest alignment process between the first domain and the second domain based on the first domain feature representation and the second domain feature representation, The steps include calling the aforementioned interest alignment model and performing the following process, namely, 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, The method according to claim 1, comprising the steps of calling the interest alignment model and performing the following process, namely, performing an intra-domain alignment process between the first and second domains based on the first and second domain feature representations.

13. The above method further, The steps include obtaining the cross-domain alignment loss corresponding to the cross-domain alignment process, and obtaining the intra-domain alignment loss corresponding to the intra-domain alignment process, The method according to 12, characterized by comprising the step of constituting the second loss with the cross-domain alignment loss and the intra-domain alignment loss.

14. The aforementioned 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 extract features from the first domain data, and the first feature extraction module is obtained through pre-training. The aforementioned second feature extraction module is used to perform feature extraction on the aforementioned second domain data. The cross-domain alignment module is used to perform interest alignment on the second interest similarity between two objects in the second domain and the first interest similarity between two corresponding objects in the first domain. The method according to claim 1, wherein the intradomain alignment module is used to perform interest alignment processing on the object graph of the first domain with respect to the object resource graph of the first domain, and to perform interest alignment processing on the object graph of the second domain with respect to the object resource graph of the second domain.

15. The second loss includes the cross-domain alignment loss obtained in the cross-domain alignment process and the intra-domain alignment loss obtained in the intra-domain alignment process, and the step of training the interest alignment model in a direction that reduces the first loss and the second loss, and obtaining the trained interest alignment model, is: The steps include adding the first loss of the interest alignment model, the cross-domain alignment loss, and the intra-domain alignment loss to obtain the target loss of the interest alignment model, The step includes training the interest alignment model in a direction that reduces the target loss, The method according to claim 14, wherein training of the interest alignment model in a direction that reduces the target loss is interrupted during the process of training the first feature extraction module based on the cross-domain alignment loss.

16. The above method further, The steps include obtaining the object attributes of the target object to be distributed, and the resource feature representation of each candidate resource data to be distributed in the second domain, The steps include calling the pre-trained interest alignment model, extracting features based on the object attributes of the target object, and obtaining a fused feature representation of the target object, The steps include: performing a similarity comparison between the fused feature representation of the target object and the resource feature representation of each candidate resource data, and obtaining the similarity comparison result between the target object and each candidate resource data; The method according to claim 1, comprising the step of making candidate resource data whose similarity comparison results satisfy the resource recommendation rules second resource data to be distributed associated with the target object.

17. A model training device, It includes an acquisition unit and a processing unit, The acquisition unit is configured to acquire a data set, the data set comprising first field data and second field data, wherein the first field data comprises at least one object and first resource data of interest within the first field of each object, and the second field data comprises at least one object and second resource data of interest within the second field of each object. The processing unit is configured to call an interest alignment model to extract features from the first domain data and obtain a first domain feature representation, and then call the interest alignment model again to extract features from the second domain data and obtain a second domain feature representation. The processing unit is further configured to call the interest alignment model and perform the following processing, namely, interest alignment processing between the first and second domains based on the first and second domain feature representations: The processing unit is further configured to train the interest alignment model in a manner that reduces the first loss and the second loss, and to obtain the 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 domain.

18. It is an electronic device, A processor suitable for executing computer executable instructions, Electronic device comprising a computer-readable storage medium in which computer-executable instructions are stored, and which, when executed by the processor, realizes the model training method described in any one of claims 1 to 16.

19. It is a computer program, The computer program includes computer executable instructions, the computer executable instructions, when executed by a processor, implement the model training method described in any one of claims 1 to 16.

Images