A cross-domain recommendation method and system based on adversarial learning and graph neural architecture search

Abstract

This invention belongs to the field of cross-domain recommendation and discloses a cross-domain recommendation method and system based on adversarial learning and graph neural architecture search. This method enhances the ability to represent interactive information by finely constructing cross-domain heterogeneous graphs; it designs cross-domain customized hypernetworks and automatically generates optimal GNN structures for each domain based on graph neural architecture search, achieving end-to-end repetitive training optimization; and it combines a behavior importance perceptron with adversarial learning to mitigate negative transfer from two levels: instance selection and distribution alignment, thereby improving cross-domain recommendation performance while preserving the heterogeneity of user preferences. Experimental results show that this invention improves evaluation metrics such as AUC, LogLoss, Recall, and ReleaseImpr on the Amazon and Douban datasets.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of cross-domain recommendation technology, and particularly relates to a cross-domain recommendation method and system based on adversarial learning and graph neural architecture search. Background Technology

[0002] In the information age, the internet has become the core channel for people to access various resources. It not only provides abundant information resources but also drives the rapid expansion of data scale. However, the massive accumulation of data has also brought the challenge of information filtering; when faced with massive amounts of complex data, users often find it difficult to quickly find content that matches their interests.

[0003] Against this backdrop, recommender systems emerged. In 1994, the GroupLens research group built a news recommendation system, a pioneering attempt to alleviate information overload by leveraging the collective wisdom of users. In 1997, Resnick and his team proposed a theoretical framework for recommender systems, defining it as an intelligent information filtering system that provides personalized content recommendations by analyzing user behavior data. With continuous technological advancements, recommender systems have permeated all aspects of life, from intelligent product recommendations in e-commerce to personalized playlist customization on music platforms, with increasingly diverse application scenarios and a constant stream of related research findings. Among these, collaborative filtering algorithms, as a representative of early traditional recommender algorithms, have had a profound impact on the field due to their simple principles and stable performance. This algorithm is based on users' past interaction data and achieves personalized recommendations by analyzing the similarities between users or items.

[0004] Traditional recommendation systems are often limited to data from a single domain, facing two core problems: first, data sparsity, meaning the user-item interaction matrix has many missing values, especially for long-tail items and niche user groups; insufficient effective interaction records directly reduce the accuracy of preference modeling; second, the cold start problem, where the recommendation effect drops significantly when new users register or new items are launched due to a lack of historical data support, making it difficult to meet users' immediate needs and achieve platform business goals. Cross-Domain Recommendation (CDR), as a key technology to solve these problems, lies in breaking down data barriers between domains. By mining the knowledge connections between the source and target domains, effective information from the source domain is transferred to the target domain, such as mature user behavior patterns and item characteristic rules, thereby enriching the target domain's data and optimizing user preference representation. For example, in an e-commerce scenario, using toys as the source domain and musical instruments as the target domain, users' purchasing preferences for children's toys can be transferred to musical instruments, recommending suitable children's electronic keyboards. This improves recommendation accuracy while further expanding user needs.

[0005] From an application perspective, CDR technology is adaptable to a wide range of scenarios: on content platforms, it enables cross-domain recommendations for various content types such as short videos, live streams, and text / image content; in the local services sector, it provides cross-domain recommendations for service industries such as catering, accommodation, and cultural tourism; and in government services, it can push appropriate service guides based on user behavior. CDR technology is particularly important for small and medium-sized platforms and emerging fields with insufficient data accumulation. By leveraging knowledge transfer from mature domains, basic recommendation capabilities can be quickly built, reducing operational costs and user churn risks during the cold start phase.

[0006] While existing cross-domain recommendation research has achieved some results, there is still considerable room for improvement in the accuracy of interest transfer.

[0007] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0008] Existing cross-domain recommendation methods based on GNNs mostly rely on fixed architectures designed with domain-specific expert knowledge and fail to reasonably evaluate the value of user interaction information from the source domain, easily leading to negative transfer. Furthermore, they have shortcomings in handling heterogeneous graph structures across multiple domains and aligning feature spaces between domains, making it difficult to achieve accurate transfer of effective information from the source domain and significant improvement in recommendation performance in the target domain. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention provides a cross-domain recommendation method and system based on adversarial learning and graph neural architecture search.

[0010] This invention is implemented as follows: a cross-domain recommendation method based on adversarial learning and graph neural architecture search, characterized in that the cross-domain recommendation method based on adversarial learning and graph neural architecture search specifically includes:

[0011] S1: Construct a cross-domain heterogeneous graph;

[0012] Construct a heterogeneous graph that includes users from multiple domains, projects from multiple domains, and interaction relationships. The domain attributes of the nodes in the heterogeneous graph are explicitly defined, and the edge weights are refined by fusing ratings and comment texts, forming a semantically rich cross-domain heterogeneous graph structure.

[0013] S2: Design cross-domain customized hypernetwork modules;

[0014] The cross-domain customized hypernetwork module is a graph neural network architecture that is customized to adapt to the heterogeneous interaction mode of each domain, and dynamically adjusts the message passing strategy according to the richness of edge information.

[0015] S3: Introduces an adversarial learning module;

[0016] The architectural features output by the cross-domain customized hypernetwork module are input into the domain discriminator, and the feature distributions of the source domain and the target domain are aligned through domain adversarial training.

[0017] S4: Design the behavioral importance sensor module;

[0018] By leveraging the graph structure of user-item interactions, the contribution of source domain interaction behaviors to the target domain is dynamically evaluated during model training, and the behavior weights are adaptively adjusted by combining domain fit calibration and comment semantic gating mechanisms.

[0019] Furthermore, in S1, a scenario with N users and M items is considered, where the interaction data between users and items is represented as a heterogeneous graph. ,here, For a set of n users, Let m be the set of m items. It is a set of user-item interactions, each interaction Each element is labeled with its domain d∈D (D represents the set of domains) and assigned a click label. ,in =1 indicates a click (rating greater than or equal to 3 points). =0 means no click (rating less than 3 points). For each interaction edge, there is a set of edge attributes. Has edge properties , It is composed of a concatenation of a domain vector, a click label vector, and a text vector. The domain vector and click label vector are trainable embedding vectors, and the text vector is obtained by encoding comment text using a pre-trained BERT model (with zeros padded when there are no comments). This augmented graph structure provides richer interactive semantics for subsequent modules. To adapt to graph neural network processing, the above heterogeneous graph is numerically represented as follows:

[0020] (1) Node feature matrix: Each user and item node is initialized as a trainable embedding vector, or a concatenation of node type embeddings;

[0021] (2) Edge index matrix: records the index of the source node (user) and target node (project) of each edge;

[0022] (3) Edge feature matrix: The attributes of each edge are obtained through embedding and concatenation. Specifically, the domain is mapped to a trainable embedding vector, the click label is mapped to an embedding vector, and finally the domain vector, the click label vector, and the text vector are concatenated;

[0023] (4) Boundary label: The index of the domain to which each edge belongs, used for subsequent domain customization processing;

[0024] (5) Node domain label: Each node is determined based on the majority of the domains it interacts with (if a node appears in multiple domains at the same time, the domain with the most occurrences is selected), which is used for domain discrimination in adversarial learning.

[0025] Furthermore, in S2, the cross-domain customized hypernetwork module, based on the One-shot graph neural architecture search, has been improved as follows: A continuous search space containing differentiated graph operations is constructed, and each operation in each layer is assigned learnable weights. During training, gradient descent is used to jointly optimize the operation weights and network parameters. The cross-domain customized hypernetwork contains two layers of GNNs, and the search space covers multiple types of differentiated graph operations: basic node feature operations (GCN, GAT, GraphSAGE, LightGCN, linear layers) and GAT-edge operations that explicitly fuse edge features. GAT-edge, as the core differentiated operation, is used to capture the impact of edge semantic information such as comments on user-item interactions, enabling the hypernetwork to adaptively select the optimal graph operation based on the information content of the edge features.

[0026] Furthermore, in S3, the global user representation output by the hypernetwork is... An input adversarial module is used to learn the homogeneity and heterogeneity of user preferences across different domains. First, a generator network G is used as a mapping function from the global user embeddings to the shared embedding latent space, i.e.:

[0027] (7)

[0028] The generator learns various behavioral patterns in the global user embedding, so that the resulting user-shared embedding can represent complete user preferences.

[0029] Next, it is input into the domain discriminator network D to further share the heterogeneity of user embeddings in different domains. The generator G and the domain discriminator D are trained simultaneously during the optimization process. On the one hand, the parameters of G need to be updated to generate more robust user shared embeddings that cannot be distinguished by D. This step is to minimize the conditional generation loss of G:

[0030] (8)

[0031] On the other hand, to update the parameters of D and enhance its ability to distinguish between real and fake inputs generated by G in the previous step, this step involves minimizing the following cross-entropy loss:

[0032] (9)

[0033] In summary, this adversarial structure follows a two-player mini-maximum game, where the conditional value function... Determined by the following minimum and maximum losses:

[0034] (10)

[0035] in It is a loss of the supernetwork. These are the hyperparameters that balance the hypernetwork module and the adversarial module.

[0036] Furthermore, in S4, the graph-based behavior importance perceptron is an external module that guides the learning process by assigning weights to the loss of each sample in the source domain. Comment semantic gating is introduced to this module to refine the evaluation of interaction value for each source domain interaction. The semantic embeddings of the comment text were extracted using a pre-trained BERT model. (If there are no comments, fill with a zero vector). Based on this embedding, design a comment gating mechanism. To measure the contribution of the comment to user preference modeling:

[0037] (12)

[0038] in, The function is a sigmoid function, and the MLP is a two-layer fully connected network. Comment gating provides fine-grained control at the content level: when the comment has low information content, the gating value approaches 0, thus suppressing the contribution of that interaction to the loss; when the comment is rich in semantics, the gating value amplifies its impact, effectively enhancing the ability to filter high-value transfer information and suppressing the risk of negative transfer.

[0039] A domain fit calibration mechanism is introduced, which adjusts the fusion strategy of source domain weights in real time based on the output of the adversarial discriminator. When the domain fit is high, the model prioritizes the global importance weights of the project, strengthening the transfer of common cross-domain knowledge; when the domain fit is low, it automatically switches to user-specific importance weights, focusing on user personalized preferences, thus improving domain fit. Defined as:

[0040] (14)

[0041] in To counter the discriminator, For the user-shared embedding output by the generator, then, The calculation is as follows:

[0042] (13)

[0043] in, It is a normalization function. It is the domain importance weight of domain s. Items on the cross-domain interaction graph G Global importance weights, This item Compared to users User-specific importance weights.

[0044] Another objective of this invention is to provide a cross-domain recommendation system based on adversarial learning and graph neural architecture search. This system employs a two-layer collaborative optimization structure to achieve end-to-end training, specifically including:

[0045] The heterogeneous graph construction module is used to build heterogeneous graphs that include users from multiple domains, projects from multiple domains, and interaction relationships. By explicitly defining node domain attributes and integrating rating and comment text to refine edge weights, a semantically rich cross-domain heterogeneous graph structure is formed.

[0046] A cross-domain customized supernetwork, located in the inner layer, automatically generates the optimal GNN architecture for each domain (source and target domains) through architecture search. The output of the supernetwork is used as the input for adversarial learning. Through adversarial training, the feature extractor learns domain-invariant user representations, thereby aligning the distributions of the source and target domains in the feature space to alleviate the domain shift problem.

[0047] The behavioral importance perceptron, located in the outer layer, dynamically evaluates the importance of source domain interactions through three-layer weight modeling, filters high-value transfer information, and guides the optimization of the inner supernetwork, thereby suppressing the risk of negative transfer.

[0048] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0049] This invention does not simply superimpose existing GNN recommendation, cross-domain transfer, and NAS methods, but rather achieves synergistic innovation at the mechanism level: First, by constructing a cross-domain heterogeneous graph that integrates rating and comment semantics, it achieves a higher-dimensional information expression than traditional methods based solely on interaction relationships; second, it introduces a cross-domain customized supernetwork, enabling the graph neural architecture to dynamically generate interaction patterns for different domains, rather than relying on a fixed structure, thus solving the problem that existing methods struggle to adapt to multi-domain differences; third, it uses adversarial learning for the distribution alignment of architectural feature layers, rather than traditional embedding alignment, improving cross-domain generalization ability; finally, through a behavior importance perception mechanism, it achieves dynamic modeling of the contribution of source domain behavior to the target domain, overcoming the problem of static or coarse-grained cross-domain transfer weights in existing methods.

[0050] This invention not only employs graph neural architecture search but also achieves structural adaptive optimization through a cross-domain customized hypernetwork and a dynamic message passing strategy driven by edge information. This deeply couples architecture search with graph semantic features, rather than being a simple NAS application. Secondly, it uses comment text semantics and ratings together for refined edge weight modeling, constructing a semantically enhanced heterogeneous graph to improve cross-domain relationship expression. Thirdly, based on adversarial learning, it further combines a behavior importance perceptron with a semantic gating mechanism to achieve dynamic quantification and control of source domain behavior contributions, breaking through the existing single adaptation mode that only performs distribution alignment. The synergistic effect of these multiple modules enables the model to not only solve the distribution offset problem but also take into account structural adaptation and behavior selection optimization, resulting in a synergistic improvement in overall technical performance.

[0051] This invention enhances the ability to express interactive information by constructing cross-domain heterogeneous graphs with fine detail; it designs cross-domain customized supernetworks and automatically generates the optimal GNN structure for each domain based on graph neural architecture search, achieving end-to-end repetitive training optimization; and it combines behavioral importance perceptrons and adversarial learning to alleviate negative transfer from two levels: instance selection and distribution alignment, thereby improving cross-domain recommendation performance while preserving the heterogeneity of user preferences.

[0052] To verify the effectiveness of the proposed model, multiple experiments were designed based on publicly available datasets from Amazon and Douban to demonstrate that the model outperforms comparative methods in metrics such as AUC, LogLoss, and Recall. Specifically, the model's predictive superiority was verified through a CTR comparison experiment; its ranking capability was analyzed using Top-K recommendation experiments; the effectiveness of core components was verified using ablation experiments; the robustness of the model was tested through sparse scenario experiments; and the influence of key parameters was analyzed and the optimal range was determined using the controlled variable method.

[0053] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0054] Improve recommendation accuracy: By constructing a refined cross-domain heterogeneous graph, a cross-domain customized hypernetwork, and an adversarial learning module, effective information from the source domain is accurately transferred, significantly improving the accuracy of user preference modeling and recommendation hit rate.

[0055] Reduce cold start losses: By leveraging cross-domain knowledge transfer, effectively alleviate the challenges of data sparsity and cold start for new users, new items, and niche fields, and improve the quality of recommendations during the cold start phase.

[0056] Suppressing negative transfer risk: By using a behavior importance perceptron, comment semantic gating, and domain fit calibration, high-value interaction information is dynamically filtered to avoid invalid noise transfer and ensure the stability of cross-domain recommendations.

[0057] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0058] Enhance platform competitiveness: In e-commerce, content platforms, local services and other scenarios, it can significantly improve click-through rates, conversion rates and user retention, creating core technological advantages for the platform.

[0059] Increase user engagement: High-quality cross-domain recommendations can accurately match users' potential interests, extend user engagement time, and enhance user loyalty and platform activity.

[0060] Reduce R&D costs: End-to-end automatic modeling is achieved based on One-shot graph neural architecture search, eliminating the reliance on manually designed GNN architectures and reducing the cost of expert intervention and repeated training.

[0061] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0062] This invention integrates adversarial learning with graph neural architecture search for cross-domain recommendation, enabling automatic generation of domain-adaptive GNN architectures and alignment with domain feature distributions, thus filling the gap in collaborative modeling of architecture adaptation and transfer learning in cross-domain recommendation.

[0063] To address the shortcomings of traditional GNN cross-domain recommendation, which relies on a fixed architecture and is difficult to adapt to heterogeneous interactions across multiple domains, a cross-domain customized supernetwork is proposed. This network dynamically generates optimal graph operation combinations for different domains, filling the technical gap in automatic graph architecture search in heterogeneous multi-domain scenarios.

[0064] To address the issues of uneven source domain information value and difficulty in controlling negative transfer in cross-domain recommendation, a graph-based behavior importance perceptron is proposed. By combining comment semantic gating and domain fit calibration, interactive weights are adaptively adjusted, filling the technical gap in cross-domain instance-level dynamic weighted filtering.

[0065] This invention explicitly defines node domain attributes in heterogeneous graphs and integrates domain vectors, click tags, and BERT comment text semantic refinement edge weights, filling the gap in multi-dimensional semantic enhancement modeling of cross-domain heterogeneous graphs.

[0066] (4) The technical solution of the present invention overcomes technical bias:

[0067] The challenge of suppressing negative transfer in cross-domain recommendation: Traditional methods cannot effectively evaluate the interaction value of the source domain and are prone to introducing noise that leads to a decline in the performance of the target domain. This invention suppresses negative transfer through a two-layer approach of feature alignment and instance selection, achieving stable and effective transfer.

[0068] The problem of fixed cross-domain recommendation architectures based on GNNs and reliance on expert experience is that manually designed GNN architectures cannot adapt to heterogeneous multi-domain patterns. This invention automatically generates domain-customized optimal architectures through graph neural architecture search, thus overcoming the limitations of manual design.

[0069] The problem of insufficient utilization and coarse expression of semantic information in cross-domain heterogeneous graphs: Traditional methods simply model interaction relationships and ignore key semantics such as ratings, comments, and domain attributes. This invention constructs a highly semantically enhanced heterogeneous graph, which significantly improves the accuracy of feature expression.

[0070] The challenge of balancing cross-domain feature distribution alignment with user preference heterogeneity: Traditional adversarial methods tend to over-align and lose individuality. This invention preserves user preference heterogeneity in domain adversarial scenarios while eliminating domain-specific noise, thus overcoming a long-standing dilemma. Attached Figure Description

[0071] Figure 1 This is a flowchart of a cross-domain recommendation method based on adversarial learning and graph neural architecture search provided in an embodiment of the present invention;

[0072] Figure 2 This is a schematic diagram of a cross-domain customized hypernetwork module provided in an embodiment of the present invention;

[0073] Figure 3 This is an architecture diagram of a cross-domain recommendation system based on adversarial learning and graph neural architecture search provided in an embodiment of the present invention;

[0074] Figure 4 These are comparative experimental results of different models provided in this embodiment of the invention on the Amazon (Bo-Mo) dataset;

[0075] Figure 5 These are comparative experimental results of different models provided in this embodiment of the invention on the Amazon (Bo-CD) dataset;

[0076] Figure 6 This is a visualization of ablation experiment results provided in the embodiments of the present invention;

[0077] Figure 7 This is a visualization of the comparative experiment on the effects of different numbers of graph neural network layers on AUC and LogLoss provided in the embodiments of the present invention. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0079] like Figure 1 As shown, this embodiment of the invention provides a cross-domain recommendation method based on adversarial learning and graph neural architecture search, which specifically includes:

[0080] S1: Construct a cross-domain heterogeneous graph;

[0081] Construct a heterogeneous graph that includes users from multiple domains, projects from multiple domains, and interaction relationships. The domain attributes of the nodes in the heterogeneous graph are explicitly defined, and the edge weights are refined by fusing ratings and comment texts, forming a semantically rich cross-domain heterogeneous graph structure.

[0082] S2: Design cross-domain customized hypernetwork modules;

[0083] The cross-domain customized hypernetwork module is a graph neural network architecture that is customized to adapt to the heterogeneous interaction mode of each domain, and dynamically adjusts the message passing strategy according to the richness of edge information.

[0084] S3: Introduces an adversarial learning module;

[0085] The architectural features output by the cross-domain customized hypernetwork module are input into the domain discriminator, and the feature distributions of the source domain and the target domain are aligned through domain adversarial training.

[0086] S4: Design the behavioral importance sensor module;

[0087] By leveraging the graph structure of user-item interactions, the contribution of source domain interaction behaviors to the target domain is dynamically evaluated during model training, and the behavior weights are adaptively adjusted by combining domain fit calibration and comment semantic gating mechanisms.

[0088] S1 considers a scenario with N users and M items, where the user-item interaction data is represented as a heterogeneous graph. .here, For a set of n users, Let m be the set of m items. It is a collection of user-item interactions. Each interaction Each element is labeled with its domain d∈D (D represents the set of domains) and assigned a click label. ,in =1 indicates a click. =0 means do not click. For each interaction edge, there is a set of edge attributes. Has edge properties , It is composed of a domain vector, a click label vector, and a text vector. The domain vector and click label vector are trainable embedding vectors, and the text vector is obtained by encoding the comment text using a pre-trained BERT model. This augmented graph structure provides richer interactive semantics for subsequent modules. To adapt to graph neural network processing, the above heterogeneous graph is numerically represented as follows:

[0089] (1) Node feature matrix: Each user and project node is initialized as a trainable embedding vector, or a concatenated node type embedding.

[0090] (2) Edge index matrix: records the index of the source node and the target node of each edge.

[0091] (3) Edge feature matrix: The attributes of each edge are obtained through embedding and concatenation. Specifically, the domain is mapped to a trainable embedding vector, the click label is mapped to an embedding vector, and finally the domain vector, the click label vector and the text vector are concatenated.

[0092] (4) Boundary label: The index of the domain to which each edge belongs, used for subsequent domain customization processing.

[0093] (5) Node domain label: Each node is determined based on the majority of the domains it interacts with, and is used for domain discrimination in adversarial learning.

[0094] The goal of cross-domain recommendation is to leverage the resources from one or more source domains. S Supplementary information ∈D is used to predict the user's preference for the target domain d. T The preferences of the items in, where d S ≠d T Given interaction data G from the source domain ds Interaction data G from the target domain dt The task is to learn the function f: U×I×D→R, which estimates the target domain d. T The probability of interaction between users u∈U and terms i∈I.

[0095] This invention employs the classic dual-domain setup for cross-domain recommendation, specifically a dual-domain cross-domain recommendation scenario involving only a single source domain and a single target domain: the source domain provides transferable auxiliary interaction knowledge to the model, while the target domain is the core optimization object for the recommendation task. In multi-domain settings involving multiple source domains, the function f must effectively integrate data from multiple source domains {d}. S1 d S2 , . . , d Sk This invention aims to improve prediction accuracy in the target domain dT, where k represents the number of source domains. The method is applicable to both two-domain and multi-domain CDR, optimizing data importance and architecture to improve target domain recommendations regardless of the number of source domains.

[0096] S2, the domain-customized supernetwork, aims to determine the optimal graph neural network architecture for each domain. The cross-domain customized supernetwork module proposed in this invention improves upon One-shot graph neural network architecture search by constructing a continuous search space containing differentiated graph operations, assigning learnable weights to each operation in each layer, and jointly optimizing operation weights and network parameters through gradient descent during training. This continuous representation makes the architecture search process differentiable, thereby achieving end-to-end joint optimization of supernetwork training and architecture search. After the architecture search is completed, there is no need to retrain the optimal architecture as in traditional methods, because the supernetwork has already incorporated knowledge of all candidate architectures through weight sharing. The weights of the optimal architecture can be directly inherited from the supernetwork; only the learned operation weights need to be selected as the final architecture. This method not only increases flexibility but also simplifies the overall process.

[0097] The specific structure of cross-domain customized hypernetworks is as follows: Figure 2 As shown, where e is embedded (i) This is generated from the i-th layer of the hypernetwork. The cross-domain customized hypernetwork consists of two layers of GNNs, and the search space covers multiple types of differentiated graph operations: basic node feature operations and GAT-edge operations that explicitly fuse edge features. Basic node feature operations include GCN, GAT, GraphSAGE, LightGCN, and linear layers. GAT-edge, as the core differentiated operation, is used to capture the impact of edge semantic information such as comments on user-item interactions, enabling the hypernetwork to adaptively select the optimal graph operation based on the information content of edge features.

[0098] This invention employs a one-shot graph neural architecture search method, utilizing weight sharing within a hypernetwork. The optimization objective is defined as follows:

[0099] (1)

[0100] (2)

[0101] Where 'a' represents a graph neural network architecture, and A represents the search space for all possible architectures. This refers to a heterogeneous graph constructed solely from training data, as described in the problem definition. The parameters of the hypernetwork are denoted as... , The optimized parameters are denoted as , and the architecture with the best performance in the search space is denoted as . . This indicates the AUC performance of the architecture on the validation set. This refers to the loss of architecture a on the training set. To mitigate the losses in the game, To balance the hyperparameters of the hypernetwork module and the adversarial learning module, specifically for the click-through rate (CTR) prediction problem that this invention focuses on, the loss can be calculated as follows:

[0102] (3)

[0103] (4)

[0104] in For the expected user-item pairs in the dataset, To support user-shared embeddings generated by the adversarial learning module, The project features output by the hypernetic network module. Let U be the actual click label for user u on item i, and BCE represent binary cross-entropy loss, making it well-suited for click-through rate prediction tasks.

[0105] To obtain the optimal architecture *a*, we first construct a hypernetwork *S* following the principles outlined in the One-Shot NAS literature. A hypernetwork is an overparameterized model that encompasses all possible graph neural network architectures in the search space *A*, fusing various operations into a continuous space, represented as:

[0106] (5)

[0107] Where x is the input, i.e., the node features of the heterogeneous graph. This is the set of edge attributes adjacent to the current node. The output consists of the user representation and the item representation, and φ is the set of candidate operations. It is an operation. These are the learnable weights of operation o in the i-th layer, optimized using gradient descent. To simplify this process, this invention maintains a set of parameters for each operation across all layers. The optimization formula for the hypernetwork is:

[0108] (6)

[0109] in It is the training loss of architecture a on graph G, which is determined by the weights in the supernetwork. Parameterization To combat learning loss.

[0110] S3 refers to the global user representation output by the hypernetwork. An adversarial module is input to learn the homogeneity and heterogeneity of user preferences across different domains. Specifically, a generator network G is first used as a mapping function from the global user embedding to the shared embedding latent space, i.e.:

[0111] (7)

[0112] The generator learns various behavioral patterns in global user embeddings, enabling the obtained user-shared embeddings to represent complete user preferences. These are then fed into a domain discriminator network D to further share the heterogeneity of user embeddings across different domains. The domain discriminator network D aims to classify real and fake inputs generated by the generator network G in an unsupervised manner. In the context of this invention, real embeddings contain domain-specific information, enabling the discriminator to identify their source domain. Fake embeddings have had domain-specific information removed, making it impossible for the discriminator to distinguish their source domain.

[0113] The generator G and the domain discriminator D are trained simultaneously during the optimization process. On the one hand, it is necessary to update the parameters of G to generate more robust user-shared embeddings that cannot be distinguished by D. This step is to minimize the conditional generation loss of G.

[0114] (8)

[0115] On the other hand, to update the parameters of D and enhance its ability to distinguish between real and fake inputs generated by G in the previous step, this step involves minimizing the following cross-entropy loss:

[0116] (9)

[0117] In summary, this adversarial structure follows a two-player mini-maximum game, where the conditional value function... Determined by the following minimum and maximum losses:

[0118] (10)

[0119] in It is a loss of the supernetwork. These are the hyperparameters that balance the hypernetwork module and the adversarial module.

[0120] During training, the generator G attempts to increase the loss, meaning the generated shared user embeddings contain behavioral features from each domain, making it difficult for the discriminator D to distinguish the origin domain of the embedding. The discriminator, on the other hand, attempts to decrease the loss, capturing as much heterogeneous behavior as possible across domains to identify the domain to which the shared user embedding belongs. Through the adversarial interaction between the two networks, the shared user embeddings can learn the similarity of preferences across different domains and also capture behavioral heterogeneity within those domains.

[0121] In S4, to optimize the supernetwork, this invention proposes a graph-based behavior importance perceptron, an external module designed to guide the learning process by evaluating the importance of each interaction in the source domain S. This is achieved by assigning weights to the loss of each sample in the source domain. For ease of illustration, only two domains are considered: the source domain S and the target domain T. The training loss is calculated as:

[0122] (11)

[0123] in Let represent the total loss over the target domain T. Here, and Edges from domains S and T respectively and The true label. and These are the corresponding predicted labels, which are user features output by the hypernetwork mapped to shared embeddings by the generator. Then, the project features output by the hypernetwork are input into the MLP for calculation. The domain S is assigned to the interaction Importance weight, This represents the binary cross-entropy (BCE) loss function. To minimize / maximum game loss for the adversarial module, To balance the hyperparameters of the hypernetwork and the adversarial module.

[0124] To achieve dynamic and adaptive weight allocation, this invention introduces comment semantic gating into this module to refine the evaluation of interaction value. For each source domain interaction... The semantic embeddings of the comment text were extracted using a pre-trained BERT model. Based on this embedding, a comment gating system is designed. To measure the contribution of the comment to user preference modeling:

[0125] (12)

[0126] in, The function is a sigmoid function, and the MLP is a two-layer fully connected network. Comment gating provides fine-grained control at the content level: when the comment information content is low, the gating value approaches 0, thereby suppressing the contribution of the interaction to the loss; when the comment is rich in semantics, the gating value amplifies its influence, effectively enhancing the ability to filter high-value transfer information and suppressing the risk of negative transfer.

[0127] Furthermore, this invention introduces a domain fit calibration mechanism for this module, adjusting the fusion strategy of source domain weights in real time based on the output of the adversarial discriminator. When the domain fit is high, the model prioritizes the global importance weights of the project, strengthening the transfer of common cross-domain knowledge; when the domain fit is low, it automatically switches to user-specific importance weights, focusing on user-specific preferences. Domain Fit Defined as:

[0128] (14)

[0129] in To counter the discriminator, For user-shared embeddings output by the generator. Then, The calculation is as follows:

[0130] (13)

[0131] in, It is a normalization function. It is the domain importance weight of domain s. Items on the cross-domain interaction graph G Global importance weights, This item Compared to users User-specific importance weights.

[0132] Domain adaptability during training As a dynamic gating coefficient, it is related to the importance weight of the source domain. The overall importance weight of the project User-specific weights corresponding to the project Perform weighted fusion, then multiply by comment gating. Finally, normalization is used to generate the final importance weight for each source domain interaction. The calculation of the source domain importance weight, the global importance weight of the item, and the user-specific weight corresponding to the item is as follows:

[0133] (1) Source domain importance weights are scalar parameters that are continuously updated during training.

[0134] (2) The global importance weight of the project is represented by project nodes using GAT. And calculation.

[0135] (3) The user-specific weight corresponding to the project is represented by the user-project interaction subgraph. This is determined by applying multiple layers of GraphSAGE. The terms are represented... and users The connection is established and further processed by the MLP.

[0136] In the method of this invention, a graph-based behavior importance perceptron serves as the outer task data scheduler, while the backbone recommendation model comprises a cross-domain customized hypernetwork, adversarial learning, and a click-through rate predictor as the inner model. Together, they constitute a two-layer optimization framework. The outer perceptron is updated using a development dataset obtained from the training set through random recombination. Using the same data for both layers may hinder the outer model's ability to enhance the inner model, potentially leading to weight imbalances in Equation 11. To prevent this, the present invention employs stochastic gradient descent (SGD) with batch optimization, using different batches for the inner and outer models. This strategy avoids crashes and enables efficient two-layer optimization without requiring additional data in the target domain.

[0137] To further optimize the outer perceptron, this invention employs implicit gradients, a method commonly used in bilayer optimization, particularly suitable when direct gradient computation is not feasible due to complexity or high computational cost. The bilayer optimization problem during model training is expressed as:

[0138] (16)

[0139] (17)

[0140] in, These are the parameters of the inner cross-domain recommendation model. These are parameters of the outer sensor. The weights are calculated using formulas 11 and 13. Weighted training loss. Inner layer optimization seeks to make... Minimize parameters , representing the optimal parameters of the cross-domain recommendation model with fixed perceptron parameters. With Changes, Changes will also occur. Outer layer optimization adjustments. To ensure the inner model Development dataset in the target domain To achieve optimal performance, among which By reordering and reusing And derived from this. Specific optimization methods are as follows:

[0141] (1) Inner layer optimization: The inner layer model is optimized with fixed perceptron parameters. In this case, minimization is achieved by using standard methods such as SGD or Adam. .

[0142] (2) Outer layer optimization: Outer layer optimization is more complex because pass Indirectly dependent on Therefore, gradient The chain rule needs to be used:

[0143] (18)

[0144] Although It can be calculated using automatic differentiation, but the calculation... need:

[0145] (19)

[0146] right The gradient was obtained as follows:

[0147] (20)

[0148] Among them, Hessiande's inverse Approximates the Neumann series truncated by K:

[0149] (twenty one)

[0150] Therefore, the implicit gradient of the outer perceptron is:

[0151] (twenty two)

[0152] This can be efficiently computed using the Vector-Jacobian product method. Through this iterative approach, the outer behavioral importance perceptron and the inner model are fine-tuned alternately. The perceptron is updated as soon as the inner model converges with the current perceptron parameters. Repeating this process throughout training ensures the convergence of both models, enabling AGNAS to jointly optimize the importance of the recommendation architecture and behavioral data.

[0153] like Figure 3 As shown, this embodiment of the invention provides a cross-domain recommendation system based on adversarial learning and graph neural architecture search, namely the AGNAS model, used to construct a comprehensive and highly adaptable cross-domain recommendation solution applicable to dual-domain and multi-domain scenarios. The system employs a two-layer collaborative optimization structure to achieve end-to-end training, specifically including:

[0154] The heterogeneous graph construction module is used to build heterogeneous graphs that include users from multiple domains, projects from multiple domains, and interaction relationships. By explicitly defining node domain attributes and integrating rating and comment text to refine edge weights, a semantically rich cross-domain heterogeneous graph structure is formed.

[0155] A cross-domain customized supernetwork, located in the inner layer, automatically generates the optimal GNN architecture for each domain through architecture search, and uses the output of the supernetwork as input for adversarial learning. Through adversarial training, the feature extractor learns domain-invariant user representations, thereby aligning the distribution of the source and target domains in the feature space to alleviate the domain shift problem.

[0156] The behavioral importance perceptron, located in the outer layer, dynamically evaluates the importance of source domain interactions through three-layer weight modeling, filters high-value transfer information, and guides the optimization of the inner supernetwork, thereby suppressing the risk of negative transfer.

[0157] Example 1: Heterogeneous Graph Construction in Cross-Domain Book and Movie Recommendation Scenarios

[0158] This embodiment is applied to a cross-domain recommendation system containing book and movie domains. The system obtains user A's 5-star rating and comment text "A pinnacle of hard science fiction" for "The Three-Body Problem" in the book domain, and a 4-star rating and comment text "A stunning visualization of black holes" for "Interstellar" in the movie domain. The heterogeneous graph construction module treats user A and the two items as nodes. The interaction edge attributes between user A and "The Three-Body Problem" are concatenated from the book domain embedding vector, the 1-click tag vector, and the BERT-encoded text vector; the interaction edge attributes between user A and "Interstellar" are concatenated from the movie domain embedding vector, the 1-click tag vector, and the corresponding BERT-encoded vector. The system explicitly marks the node domain attributes: user A is marked as a book domain node because most of their interactions occur in the book domain. Through this construction, the system forms a semantically rich cross-domain heterogeneous graph structure, laying the foundation for subsequent accurate modeling. Example 2: Cross-domain customized hypernetic search in multi-category e-commerce platform scenarios

[0159] This embodiment is applied to an e-commerce platform that includes two product categories: electronics and apparel. A cross-domain customized hypernetwork module generates customized graph neural network architectures for each domain. For the electronics domain, due to the sparse nature of user interactions and their high dependence on product parameters, the hypernetwork assigns a weight of 0.7 to the GAT-edge operation for mobile phone category nodes in the second layer of the graph neural network to strengthen the fusion of edge features (such as pixel parameters and battery capacity). For the apparel domain, due to frequent user interactions and preferences being greatly influenced by fashion trends, the hypernetwork assigns a weight of 0.8 to the LightGCN operation for dress category nodes to strengthen high-order collaborative filtering. Through gradient descent joint optimization, the electronics domain architecture focuses on attribute matching, while the apparel domain architecture focuses on collaborative signal propagation, accurately meeting the unique modeling needs of each category.

[0160] Example 3: Domain Adversarial Training in Cross-Social and Video Platform Scenarios

[0161] This embodiment is applied to a recommendation scenario where users are migrated from a social platform (source domain) to a video platform (target domain). The adversarial learning module receives the user's global representation output by a cross-domain customized hypernetwork. The generator network G maps this representation to shared embeddings, for example, mapping the behavior patterns of social platform users frequently sharing technology news and video platform users frequently watching science videos to similar latent spatial locations. The domain discriminator D attempts to distinguish whether the embedding comes from the social domain or the video domain. During training, the generator G learns to generate shared embeddings that the discriminator D cannot distinguish by minimizing the generator loss, for example, making the probability of the shared embedding of the aforementioned technology users in the output of the discriminator D close to 0.5. This process eliminates the domain differences between the like noise specific to social platforms and the viewing time noise specific to video platforms, preserving the homogeneity of users' preferences for technology content.

[0162] Example 4: Negative Transfer Suppression Based on Comment Semantic Gating in News Recommendation

[0163] This embodiment is applied to a recommendation scenario that transfers user behavior from the news domain to the short video domain. When processing an interaction sample from the source domain (news domain), the behavior importance perceiver module detects a user (B) who clicks on a news article titled "Street Photos of an Actor," but the comment is simply another title with no content. A pre-trained BERT model extracts the semantic embedding of this comment, and the comment gating module calculates a gating value of 0.1 using a two-layer fully connected network. This low gating value effectively suppresses the contribution of this false click interaction to the loss function, preventing the model from transferring the noisy behavior of curiosity about low-quality headlines to the short video domain, thus avoiding negative transfer risk and ensuring recommendation quality.

[0164] Example 5: Dynamic Adjustment of Cross-Domain Weights Based on Domain Fit Calibration

[0165] This embodiment is applied to a scenario where user preferences from a music platform (source domain) are transferred to an audiobook platform (target domain). In the initial training phase, the adversarial discriminator output shows a large difference in user feature distributions between the two domains, with a domain fit value of 0.3, lower than the preset threshold of 0.6. At this point, the behavior importance perceptron automatically switches to user-specific importance weights, focusing on the fine-grained preferences of individual users on the music platform, such as user C's preference for suspense music. As training progresses, adversarial training aligns the feature distributions of the two domains, and the domain fit value rises to 0.8. The system automatically switches to global importance weights based on dependent items, at which point the transfer of suspense content from the music platform is strengthened as high-value global information, effectively guiding the recommendation of suspense content on the audiobook platform.

[0166] Example 6: System-level Example: Cold Start Recommendation for Multi-Domain Video Streaming Platforms

[0167] This embodiment presents a complete system integrating all modules, applied to a multi-domain video streaming platform. When a new user D logs in for the first time, the system only has a small amount of interaction data from their authorized social domain. The heterogeneous graph construction module merges the social domain data with existing film and television domain and variety show domain data within the platform into a unified heterogeneous graph. The inner-layer cross-domain customized hypernetwork, based on the characteristics of dense interaction in the film and television domain and sparse interaction in the social domain, automatically searches for a message passing strategy emphasizing edge features for the social domain and an architecture emphasizing high-order neighbor aggregation for the film and television domain. The behavior importance perceptron dynamically evaluates the contribution of user D's "like" behavior in the social domain to the science fiction category in the film and television domain, and filters out invalid social information using comment semantic gating. The adversarial learning module aligns the user feature distributions between the social domain and the film and television domain. Ultimately, the system successfully recommends movies of interest to user D in the film and television domain, achieving high-quality cold-start recommendation.

[0168] Evidence related to the technical effects obtained by the embodiments of the present invention.

[0169] 1. Experimental Dataset

[0170] 1.1 Introduction to the Dataset

[0171] To comprehensively evaluate the recommendation performance of the proposed model, experiments were conducted using the Amazon Product 5-core dataset and the Douban dataset, as their extensive user interactions across different product categories make them standard choices for cross-domain recommendation research. The Amazon dataset includes 143 million product reviews across 24 categories (e.g., books, clothing, movies), each representing a domain. The Douban dataset covers tens of millions of user interactions across three core cultural domains: movies, books, and music, each representing an independent recommendation domain. Due to limited user overlap between multiple domains, this invention focuses on dual-domain recommendation tasks, using the same domain pairs and splits as in BIAO. Tables 1 and 2 summarize the dataset statistics, including the number of users, items, interactions, and interaction density for each domain.

[0172] Table 1 Amazon Dataset

[0173]

[0174] Table 2 Douban Dataset

[0175]

[0176] 1.2 Dataset Preprocessing

[0177] To improve data quality and ensure the accuracy and performance of the recommendation system, all datasets must undergo rigorous data cleaning and processing before use. The core objective of this process is to guarantee data consistency and integrity, and to transform the raw multi-source data into a heterogeneous graph structure that conforms to the model input definition. In this heterogeneous graph structure, nodes represent users and items, edges represent interactions, and edge attributes include domain labels, click labels, and semantic vectors of comment text. Considering the different characteristics of the Amazon product review dataset and the Douban dataset, this study develops differentiated preprocessing strategies for the features of the two datasets. The invention formulates separate processing strategies for each dataset, with a unified construction process as follows.

[0178] (1) Node set construction. Extract all unique user IDs from the interaction records to form a user node set, and extract all unique project IDs to form a project node set. Records missing core fields such as user ID, project ID, or rating are deleted to ensure node integrity.

[0179] (2) Construction of the interaction edge set. Each valid interaction record corresponds to one edge, with corresponding attributes. Domain labels are uniformly labeled according to the dataset source or project category. The Amazon dataset uses the primary category of the product as the domain classification basis, such as books, electronics, and home furnishings. The Douban dataset uses the type of work as the domain classification basis, such as movies, music, and books, thus clearly distinguishing the domain information of different interactions. Click labels are divided according to a rating threshold; the Amazon dataset considers a rating ≥3 as a click (…). =1), <3 are considered as not clicked ( =0), the Douban dataset considers a rating of ≥4 stars as a click ( =1), otherwise it is not clicked. =0), no rating interaction, directly mapped based on user behavior logs.

[0180] (3) Construction of Edge Attribute Set. A semantic embedding vector is generated as an attribute for each interaction edge, mainly involving three steps. First, comment text cleaning: HTML tags and non-English characters are removed from Amazon English comments and converted to lowercase; special characters are removed from Douban Chinese comments, followed by word segmentation and stop word filtering. Second, text encoding: The cleaned text is uniformly processed into 128 tokens, input into the corresponding pre-trained BERT model, and the output at the CLS position is taken as the semantic vector; no comment interaction is filled with a zero vector. Third, deduplication and merging: Only the latest timestamp of multiple comments from the same user on the same item is retained to avoid duplicate edges.

[0181] After the above processing, an enhanced heterogeneous graph is obtained, in which each user-item interaction edge contains a domain label, click label, and semantic vector of the comment text. This graph structure provides a unified input format for subsequent cross-domain recommendation models and can be directly used for heterogeneous graph representation learning and click-through rate prediction tasks. Meanwhile, the missing values, noise, and inconsistencies present in the two datasets are effectively resolved, and the cleaned data quality is significantly improved, laying a solid foundation for the subsequent development of a movie-book cross-domain recommendation system.

[0182] 2. Evaluation Indicators

[0183] This invention divides the dataset into three parts: a training set, a validation set, and a test set, with a ratio of 6:2:2. To measure the model's recommendation performance, three widely used metrics are employed to evaluate the CTR task: Area Under the Curve (AUC), Log Loss (LOG), and Recall. Furthermore, the relative improvement (RelaImpr) of two metrics is calculated to measure the performance gain relative to the baseline model. To more clearly define these metrics, the concept of a confusion matrix is ​​first introduced. The confusion matrix, also known as the error matrix, is a representation used to evaluate model accuracy, as shown in Table 3.

[0184] Table 3 Confusion Matrix

[0185]

[0186] In the table above, TP represents the number of positive samples successfully identified by the model, TN refers to the total number of samples correctly identified as negative. FP describes the situation where samples that should have been negative were misclassified as positive, while FN represents the number of samples that should have been positive but were incorrectly classified as negative.

[0187] (1) AUC represents the area under the ROC curve, where the ROC curve is obtained by plotting the TPR and FPR at different thresholds. The larger the AUC value, the better the model performance. False positive rate (FPR) is used to measure the misclassification of negative samples by the classification model, that is, the proportion of samples that are actually negative but are incorrectly predicted as positive. In the ROC curve, FPR is used as the horizontal axis index to evaluate the misclassification of the model. Its calculation formula is shown in equation (23).

[0188] (twenty three)

[0190] The True Positive Rate (TPR) measures a model's ability to identify true positives; that is, the proportion of samples that are actually positive that are correctly classified as positive. In the ROC curve, TPR corresponds to the change in value on the vertical axis. Its calculation formula is shown in equation (24).

[0191] (twenty four)

[0193] Therefore, the mathematical definition of AUC is:

[0194] (25)

[0196] (2) Log loss (LogLoss), also known as binary cross-entropy loss, is used to measure how well the click probability output by the model matches the true label. The lower the value, the more accurate the probability prediction of the model, that is, the better the model performance.

[0197] The log loss for a single sample is defined as:

[0198] (26)

[0199] The overall logarithmic loss is obtained by averaging all N test samples:

[0200] (27)

[0201] Among them, when the real label is a positive example (click, When =1), the loss is due to leading, The closer to 1, the smaller the loss; when the true label is negative (not clicked), When =0), the loss is due to leading, The closer it is to 0, the smaller the loss.

[0202] (3) RelaImpr calculates the relative improvement of the target model's AUC and LOG loss relative to the benchmark. Since the return of the random prediction is 0.5, for AUC, The calculation formula is:

[0203] (28)

[0204] For log loss The calculation formula is:

[0205] (29)

[0206] (4) Recall rate refers to the proportion of samples that the user is actually interested in among all the samples they like in the top K recommended results. A higher value indicates that, given K recommendation results, the model has a stronger ability to identify samples that users like, and can more comprehensively cover users' potential interests, effectively reducing the probability of high-quality content being missed. The formula is shown in Equation 30.

[0207] (30)

[0209] 3. Experimental environment and parameter settings

[0210] The AGNAS model experiments were conducted on a Windows operating system. The main hardware and software configuration included: Python 3.9 programming language, PyTorch deep learning framework, NVIDIA RTX 4090 GPU (24GB VRAM), 32GB RAM, and an i7-11800H CPU. To comprehensively verify model performance, the experiments did not fix a single target domain but instead constructed multiple sets of source-target domain cross-domain task pairs. Different tasks were defined by configuring the combination of source and target domain categories. All tasks enabled the source domain information transfer module and activated the hypernetwork architecture search mechanism to achieve adaptive feature customization of the model for different target domains.

[0211] Regarding model parameter configuration, the base graph convolution operator for the supernetwork is GraphSAGE, which outperforms GCN and GAT in sampling efficiency and sparse graph adaptability, making it more suitable for user-item interaction graph modeling in cross-domain scenarios. The number of graph neural network layers is set to 2 by the num-layers parameter, employing a shallow structure to avoid oversmoothing of features in cross-domain scenarios and ensure the effectiveness of knowledge transfer. The hidden layer dimension of the supernetwork is controlled by the meta-hidden-dim parameter, which constrains the number of model parameters, avoids the risk of overfitting, and meets the feature representation requirements of different target domains. In the adversarial learning module, both the generator and discriminator use two-layer MLPs, with the hidden layer dimension consistent with that of the supernetwork.

[0212] In terms of optimization and training details, the optimizer Adam was chosen, and a differentiated learning rate configuration was adopted: the basic learning rate of the model backbone network was set by the lr parameter, and a dedicated learning rate conv-lr was configured for the graph convolutional layers to adapt to the optimization characteristics of graph structure feature learning. Weight-decay and dropout regularization mechanisms were introduced to suppress overfitting and improve cross-domain generalization ability. A cosine annealing strategy was used for learning rate scheduling, with the annealing period controlled by the T-max parameter and the decay step size set by the descent-step parameter to ensure convergence stability. The learning rate of the hyperparameter optimization module was configured by the hpo-lr parameter, and the meta-learning parameter update interval was adjusted by the meta-interval parameter to balance optimization efficiency and training cost. An early stopping strategy was adopted during training, with a maximum of 200 training epochs, and 20 epochs of warm-up training were performed before formal optimization (i.e., the first 20 epochs only trained the inner hypernetwork and did not update the outer perceptron) to stabilize the two-layer optimization process.

[0213] To ensure the reliability and statistical significance of the experimental results, each experimental configuration was run independently using five different random seeds. The final performance index was calculated as the average value, and the influence of randomness on the results was eliminated through a significance test. The specific parameter settings for this model are shown in Table 4.

[0214] Table 4 Parameter Settings

[0215]

[0216] 4. Experimental Comparison Model

[0217] To verify the recommendation performance of the improved model, this invention selects several methods to compare and evaluate the performance of the AGNAS model, including single-domain and cross-domain recommendation methods. Since the method of this invention is based on graph neural network (GNN) recommendation algorithms and adversarial learning, most of the comparison models are selected from leading GNN-based models and adversarial learning-based models.

[0218] (1)NGCF: This model is a collaborative filtering model based on graph neural networks. It captures high-order collaborative signals by propagating and embedding features on the user-item interaction graph, thereby improving the accuracy of recommendations.

[0219] (2) LightGCN: This model is a simplified and improved version of NGCF. It removes complex nonlinear transformations and feature transformations by retaining only the neighborhood aggregation operation in the graph convolutional layer. It can efficiently capture high-order collaborative signals in the user-item interaction graph with a lighter structure, achieving a better balance between efficiency and performance in recommendation tasks.

[0220] (3) Conet-B: This model is a BIAO enhanced variant of Conet. It introduces an auxiliary learning mechanism on the basis of cross-domain cross splicing and attention information balance of the original model, optimizes the efficiency of information transmission between domains, and improves cross-domain recommendation performance.

[0221] (4) Minet-B: This model is a BIAO enhanced variant of Minet. It continues the core idea of ​​multi-domain hybrid interest modeling of the original model, and combines auxiliary learning to improve the information interaction effect between domains, thereby enhancing the ability to represent user preferences and predict click rates in cross-domain scenarios.

[0222] (5) DisenCDR: This model uses advanced techniques such as heterogeneous graph modeling, untangling and domain alignment to achieve superior performance across multiple domains.

[0223] (6) CDR-HA: This model uses a combination of heterogeneous information networks and adversarial learning to alleviate the sparsity of domain target data and improve the accuracy of recommendations.

[0224] (7) M-scan: This model incorporates the idea of ​​causal inference into cross-domain recommendation. The model's scene-aware collaborative attention mechanism calculates the relevance score between the current scene and other scene behaviors, incorporating user behaviors from other scenes that match the current scene's interests into the calculation, thereby extracting user interests more accurately. The scene bias eliminater uses causal counterfactual reasoning to quantify and eliminate the bias caused by the scene directly affecting click behavior.

[0225] (8) ACDR: This model uses adversarial learning to further extract user preference heterogeneity. By applying independent generators to different domains for domain-specific embedding, it effectively represents users.

[0226] 5. Experimental Results and Analysis

[0227] (1) Comparison of CTR under different models

[0228] In the click-through rate (CTR) prediction task, AUC and LogLoss were used as evaluation metrics for the model. Experiments were conducted on two datasets to compare the proposed model and the baseline model. By comparing the experimental results on different datasets, the generalization performance of the model was more comprehensively measured. Table 5 shows the experimental comparison results between the CTR prediction task and single-domain methods, and Table 6 shows the experimental comparison results between the CTR prediction task and cross-domain methods. The best results are marked in bold, and the second-best results are indicated by underline.

[0229] Table 5. Experimental Results of CTR Prediction (Comparison with Single-Domain Methods)

[0230]

[0231] Analysis of the experimental results in the table above shows that AGNAS significantly outperforms NGCF and LightGCN single-domain recommendation methods in the CTR prediction task on both the Amazon and Douban datasets, verifying the superiority of the proposed model in cross-domain feature fusion, knowledge transfer, and CTR prediction accuracy.

[0232] Table 6. Experimental Results of CTR Prediction (Comparison with Cross-Domain Methods)

[0233]

[0234] Analysis of the experimental results in the table above shows that AGNAS accurately filters high-value user interaction behaviors through the behavior importance perceptron, captures the deep dependency relationship between users and items by combining GNN, and aligns inter-domain shared information with domain-specific information by using adversarial learning, ultimately achieving a more accurate recommendation effect, while the ACDR method achieves suboptimal results in most tasks.

[0235] Furthermore, in all tasks of this experiment, the cross-domain recommendation (CDR) method outperformed the single-domain model. Although the single-domain model utilizes complete data from both the source and target domains, it lacks the ability to transfer information between domains, resulting in a significantly lower performance ceiling than the CDR method. Further comparison of different types of CDR methods reveals that DisenCDR, based on GNNs, outperforms MLP-based methods (such as MiNet-B and CoNet-B). This is attributed to the ability of GNNs to capture higher-order associations through neighbor aggregation, making them more suitable for mining potential user preferences. Among the adversarial learning-based CDR methods, CDR-HA, while combining graph structures and adversarial learning to alleviate the sparsity of target domain data, did not further mine fine-grained features of user preferences. While ACDR captures heterogeneous relationships between users and items through adversarial learning, its lack of a behavioral importance perception module results in insufficient precision in filtering user and item features. The M-scan method performed exceptionally well in experiments, ranking only behind AGNAS and ACDR, and even surpassing ACDR in some tasks. M-scan's causal-driven design allows it to capture user behavior patterns more precisely, thus achieving results that are close to or even surpass those of state-of-the-art adversarial methods in cross-domain recommendation.

[0236] In summary, AGNAS has achieved a comprehensive improvement over single-domain methods and mainstream CDR methods in the CTR prediction task, providing a more efficient solution for cross-domain recommendation.

[0237] (2) To quantify the optimization magnitude of the target model relative to the best-performing benchmark model ACDR on core evaluation metrics and to eliminate scale bias and scenario dependence of the absolute values ​​of the metrics, this study introduces Relative Improvement (RelaImpr) as an auxiliary evaluation metric to standardize the model improvement effects on AUC (area under the curve) and LogLoss (log loss). Based on the CTR comparison experimental data mentioned above, the RelaImpr calculation results relative to the best-performing benchmark model ACDR are shown in Table 7, to more intuitively demonstrate the relative improvement in model performance.

[0238] Table 7 RelaImpr Calculation Results

[0239]

[0240] Analysis of the calculation results in the table above shows that the AGNAS model achieves stable performance improvements on most cross-domain recommendation tasks. The relative improvement rates of its AUC and LogLoss are mostly positive, verifying the universal effectiveness of this method in improving ranking ability and probability prediction accuracy. Furthermore, the consistent performance of the model on both the Amazon and Douban datasets demonstrates its good cross-scenario generalization ability.

[0241] (3) In the Top-K recommendation task, K values ​​of 2, 5, 10, 20, and 50 were used to observe the recommendation performance of different models. Experiments were conducted on the Bo-MO and Bo-CD tasks of the Amazon dataset, with Recall used as the evaluation metric. To more comprehensively evaluate the performance of the proposed model, it was compared with CoNet-B, MiNet-B, DisenCDR, CDR-HA, M-scan, and ACDR recommendation models. The reason for this comparison is that these five methods significantly outperform other single-domain comparison methods in Top-k prediction performance, and this comparison better highlights the advantages and effectiveness of the proposed method. Experimental results are as follows: Figure 4 and Figure 5 As shown.

[0242] As can be seen, in both the Bo-MO and Bo-CD tasks on the Amazon dataset, compared with the six recommendation models, the AGNAS model maintains the best performance for all K values, which indicates that AGNAS has stable and excellent recall performance in the TOP-K recommendation task.

[0243] (4) Ablation Experiments: To analyze the impact of each module within the AGNAS model on recommendation performance, this invention uses AUC as the evaluation metric and designs five ablation experiments using four tasks from the Amazon dataset as examples. Different modules are removed in each experiment, and the changes in model performance are compared: AGNAS-S represents not using source domain data to evaluate the role of cross-domain information in target domain recommendation; AGNAS-O represents removing adversarial modules and using the hypernetwork output directly as recommendation input; AGNAS-SN represents deleting the cross-domain customized hypernetwork and assigning equal weights to all operations to evaluate the effect of architecture customization; AGNAS-B represents removing the graph-based behavior importance perceptron and treating all source domain user behaviors equally to evaluate the impact of dynamic interaction weights; AGNAS-F represents using a fixed, optimal, manually designed GNN architecture, uniformly applied across all domains without dynamic customization. The experimental results are shown in Table 8 below:

[0244] Table 3-8 Ablation Experiment Results

[0245]

[0246] To more intuitively demonstrate the performance differences between various ablation schemes, this invention visualizes the comparative results of AGNAS and its variants on the CTR prediction task, such as... Figure 6 As shown in the figure, this bar chart clearly presents the changes in the model's AUC metric after removing different core modules, intuitively revealing the degree of influence of each module on the final recommendation effect.

[0247] Through the Figure 6 Analysis shows that the complete AGNAS consistently outperforms all alternatives across all tasks, demonstrating the importance of each module in AGNAS. Removal of source domain information leads to the most significant decrease in average model performance, indicating that source domain knowledge transfer is the core support for alleviating the target domain data sparsity problem and the foundation for improved cross-domain recommendation performance. Removal of the adversarial module and the cross-domain customized supernetwork also results in significant performance degradation, validating the crucial roles of adversarial training in aligning domain distribution and the supernetwork in generating personalized cross-domain features; both jointly ensure the model's generalization ability across scenarios. Removal of the behavior importance perceptron or the use of a fixed GNN architecture results in the slightest performance degradation, but still exhibits a stable degradation trend. This indicates that the behavior importance perceptron mechanism and the adaptive GNN architecture can effectively capture fine-grained preferences in user behavior, providing continuous performance gains for the model. In summary, ablation experiments successfully validated the importance of each module in the model.

[0248] (5) Comparative experiments in sparse scenes

[0249] To analyze the recommendation performance of AGNAS in sparse data scenarios, experiments were conducted on the Bo-MO and Bo-CD tasks of the Amazon dataset, using AUC as the evaluation metric. While maintaining the original ratio of the test set to the training set, the proportion of the training set was gradually increased from 20% to 100% to simulate training data environments with different sparsity levels. Subsequently, AGNAS was compared with the following baseline models, and the specific experimental results are shown in Tables 9 and 10.

[0250] Table 9. Comparative Experiments on the Impact of Different Training Set Sizes on AUC in the Amazon (Bo-Mo) Dataset

[0251]

[0252] As shown in the table above, the performance of all models improves with increasing training data. The AGNAS proposed in this invention achieves the best AUC value across all proportions, demonstrating a stable advantage. In particular, in sparse data scenarios (training proportion = 20%), AGNAS significantly outperforms the second-best model, indicating its ability to effectively learn user and item features with limited data and possessing strong sparsity adaptability. When the training proportion increases to 100%, AGNAS's AUC further improves, widening its lead and validating its ability to fully extract information and continuously improve recommendation performance even with abundant data.

[0253] Table 10 Comparative Experiments on the Impact of Different Training Set Sizes on AUC in the Amazon (Bo-CD) Dataset

[0254]

[0255] On the Bo-CD task, all models also exhibited a performance improvement trend with increasing training data. AGNAS outperformed other comparative methods across all data proportions from 20% to 100%, further validating the model's stability and effectiveness under different sparsity levels. Even under conditions of extremely scarce data, AGNAS still maintains strong feature representation and cross-domain transfer capabilities, outperforming other cross-domain recommendation methods.

[0256] (6) Experiment on the analysis of the number of layers in a graph neural network

[0257] The number of num-layers in a graph neural network determines the aggregation hops of node features, which has a crucial impact on the model's ability to capture higher-order neighborhood information. Too few layers may fail to fully exploit collaborative signals in the interaction graph, while too many layers can easily lead to oversmoothing of features, causing the representations of different nodes to become homogeneous and impairing the effectiveness of cross-domain knowledge transfer. To verify the rationality of the two-layer graph neural network structure used in this invention, comparative experiments were conducted on the Bo-Mo and Bo-CD tasks on the Amazon dataset and the Mo-Bo and Mo-Mu tasks on the Douban dataset. AUC and LogLoss were used as evaluation metrics, and the experimental results are shown in Tables 11 and 12.

[0258] Table 11 Comparison of the impact of different numbers of layers in graph neural networks on AUC

[0259]

[0260] As shown in Table 11, the AUC on both tasks initially increases and then decreases with increasing layer count. The model achieves optimal performance on both Amazon and Douban with 2 layers. Performance slightly decreases with 3 layers. A significant decrease in AUC with 4 layers indicates that an excessively deep network structure leads to oversmoothing, making node features difficult to distinguish and thus weakening the cross-domain adaptation effect. Experiments comparing the impact of different graph neural network layer counts on AUC verify the rationality of the 2-layer graph neural network structure used in this invention.

[0261] Table 12 Comparison of the impact of different numbers of layers in graph neural networks on LogLoss

[0262]

[0263] LogLoss is a loss metric; a smaller value indicates higher model prediction accuracy and better performance. Table 12 shows that on all four datasets, the LogLoss reaches its minimum when the number of layers is 2, indicating that a 2-layer graph neural network can most effectively extract features and avoid overfitting or underfitting in the current task. Too many or too few layers will lead to performance degradation. When the number of layers increases from 1 to 2, LogLoss decreases significantly (performance improves); when further increasing to 3 or 4 layers, LogLoss gradually rises again (performance deteriorates), showing a clear trend of first decreasing and then increasing.

[0264] To present the performance change trend more intuitively, the above results are visualized, such as... Figure 7 As shown in the figure, this graph uses a dual-axis format to simultaneously illustrate the changes in AUC (left axis, larger is better) and LogLoss (right axis, smaller is better). With two layers, the model achieves both peak AUC and trough LogLoss on all datasets, realizing the optimal balance between classification performance and prediction accuracy. As the number of layers deviates from two, both metrics worsen simultaneously, further validating the robustness and universality of the two-layer structure.

[0265] (7) Hypernetwork Learning Rate Analysis Experiment

[0266] The supernetwork is responsible for generating customized GNN architectures for each domain, and its learning rate (hpo-lr) directly affects the convergence speed of the architecture parameters and the quality of the final structure. Since the supernetwork and the main network use different learning rates, examining the sensitivity of hpo-lr helps verify the robustness of the parameter settings.

[69] This invention was tested on the Bo-MO task of the Amazon dataset and the Mo-Bo task of the Douban dataset, with AUC used as the evaluation metric. The specific experimental results are shown in Table 13 below.

[0267] Table 13 Comparative Experiments on the Impact of Learning Rates of Hypernetworks of Different Sizes on AUC

[0268]

[0269] As shown in the table above, in both tasks, the AUC value initially increases and then decreases with increasing hpo-lr. The best performance occurs around hpo-lr=0.001 (Amazon) and hpo-lr=0.0005 (Douban), indicating that the hypernetwork is sensitive to the learning rate; both excessively large and small learning rates lead to a decrease in architecture search performance. This suggests that there is an optimal learning rate range for the hypernetwork (approximately 0.0005~0.001). Too high a rate (≥0.01) leads to oscillations and non-convergence, while too low a rate (≤0.0005) results in slow updates and getting trapped in local optima. The current parameter selections of hpo-lr=0.001 for the Amazon dataset and hpo-lr=0.0005 for the Douban dataset are both within the optimal range for their respective datasets, validating the rationality of the parameter configuration.

[0270] (8) Computational complexity analysis

[0271] Let |V| and |E| represent the number of nodes and edges in the heterogeneous graph, and d represent the dimension of the hidden representation. Most GNNs have a time complexity of O(|E|d + |V|d). 2 When O is the candidate operation set, the time complexity of the Graph Neural Architecture Search (GNAS) method becomes O(|O|(|E|d+|V|D). 2 In terms of learnable parameters, most GNNs have a complexity of O(D). 2 )

[70] The complexity of GNAS is O(|O|D). 2 Therefore, the use of GNAS only introduces a linear increase in complexity, which adds minimal computational overhead and ensures that the method of the present invention remains effective.

[0272] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0273] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A cross-domain recommendation method based on adversarial learning and graph neural architecture search, characterized in that, The method includes: S1: Construct a cross-domain heterogeneous graph; Construct a heterogeneous graph that includes users from multiple domains, projects from multiple domains, and interaction relationships. The domain attributes of the nodes in the heterogeneous graph are explicitly defined, and the edge weights are refined by fusing ratings and comment texts, forming a semantically rich cross-domain heterogeneous graph structure. S2: Design cross-domain customized hypernetwork modules; The cross-domain customized hypernetwork module is a graph neural network architecture that is customized to adapt to the heterogeneous interaction mode of each domain, and dynamically adjusts the message passing strategy according to the richness of edge information. S3: Introduces an adversarial learning module; The architectural features output by the cross-domain customized hypernetwork module are input into the domain discriminator, and the feature distributions of the source domain and the target domain are aligned through domain adversarial training. S4: Design the behavioral importance sensor module; By leveraging the graph structure of user-item interactions, the contribution of source domain interaction behaviors to the target domain is dynamically evaluated during model training, and the behavior weights are adaptively adjusted by combining domain fit calibration and comment semantic gating mechanisms.

2. The cross-domain recommendation method based on adversarial learning and graph neural architecture search as described in claim 1, characterized in that, The heterogeneous graph is represented as G={U, I, E, T}, where U is the user set, I is the item set, E is the user-item interaction set, and T is the edge attribute set; the attribute of each interaction edge (u, i)∈E is composed of a domain vector, a click tag vector, and a text vector, and the text vector is obtained by encoding the comment text by a pre-trained BERT model; The heterogeneous graph is numerically represented in the following manner: The node feature matrix is ​​used to initialize user nodes and project nodes; An edge index matrix is ​​used to record the source and target node indices of each edge; The edge feature matrix is ​​used to record the attributes of each edge; Edge domain labels are used to record the index of the domain to which each edge belongs; node domain labels are used to record the domain index of each node determined by the majority of domains it interacts with.

3. The cross-domain recommendation method based on adversarial learning and graph neural architecture search as described in claim 1, characterized in that, The cross-domain customized supernetwork module is constructed based on a one-shot graph neural architecture search, which includes a continuous search space for differentiated graph operations and assigns learnable weights to each operation in each layer. During training, the operation weights and network parameters are jointly optimized through gradient descent. The cross-domain customized supernetwork module contains a two-layer graph neural network, and its search space includes basic node feature operations and edge operations that explicitly fuse edge features. The basic node feature operations include graph convolutional networks, graph attention networks, GraphSAGE, LightGCN, and linear layers.

4. The cross-domain recommendation method based on adversarial learning and graph neural architecture search as described in claim 1, characterized in that, In the adversarial learning module, the global user representation input generator network G output by the cross-domain customized supernetwork module is mapped to a user shared embedding in the shared embedding latent space; the user shared embedding is input to the domain discriminator network D; the generator network G and the domain discriminator network D are trained simultaneously during the optimization process, and a minimum-maximum game is formed by minimizing the generator loss and minimizing the discriminator cross-entropy loss.

5. The cross-domain recommendation method based on adversarial learning and graph neural architecture search as described in claim 1, characterized in that, The graph-based behavior importance perceptron module is an external module that guides the learning process by assigning weights to the loss of each sample in the source domain. The comment semantic gating mechanism uses a pre-trained BERT model to extract the semantic embeddings of interactive comment texts in the source domain, and calculates the comment gating value based on the semantic embeddings through a multi-layer fully connected network to measure the contribution of the comment to user preference modeling. The domain fit calibration mechanism adjusts the fusion strategy of the source domain weights in real time based on the output of the adversarial discriminator. When the domain fit is higher than the threshold, the model relies on the global importance weights of the items. When the domain fit is lower than the threshold, the model switches to user-specific importance weights.

6. A cross-domain recommendation system based on adversarial learning and graph neural architecture search, characterized in that, The system for implementing the method according to any one of claims 1 to 5 specifically comprises: The heterogeneous graph construction module is used to build heterogeneous graphs that include users from multiple domains, projects from multiple domains, and interaction relationships. By explicitly defining node domain attributes and integrating rating and comment text to refine edge weights, a semantically rich cross-domain heterogeneous graph structure is formed. Cross-domain customized hypernetworks, located in the inner layer of the system, automatically generate the optimal graph neural network architecture for each domain through architecture search, and use the output results as input for adversarial learning. Through adversarial training, the feature extractor learns domain-invariant user representations. The behavior importance sensor, located in the outer layer of the system, dynamically evaluates the importance of source domain interactions through weight modeling, filters high-value migration information, and guides the optimization of the cross-domain customized hypernetwork described in the inner layer.

7. The cross-domain recommendation system based on adversarial learning and graph neural architecture search as described in claim 6, characterized in that, The heterogeneous graph constructed by the heterogeneous graph construction module is represented as G={U, I, E, T}, where U is the user set, I is the item set, E is the user-item interaction set, and T is the edge attribute set. The attribute of each interaction edge (u, i)∈E is composed of a domain vector, a click tag vector, and a text vector, wherein the text vector is obtained by encoding the comment text using a pre-trained BERT model. The heterogeneous graph is numerically represented in the following ways: a node feature matrix, used to initialize user nodes and item nodes; and an edge index matrix, used to record the source node and target node indices of each edge. The edge feature matrix records the attributes of each edge; the edge domain label records the index of the domain to which each edge belongs; and the node domain label records the domain index of each node determined by the majority of domains it interacts with.

8. The cross-domain recommendation system based on adversarial learning and graph neural architecture search as described in claim 6, characterized in that, The cross-domain customized supernetwork is constructed based on a one-shot graph neural architecture search, which includes a continuous search space for differentiated graph operations and assigns learnable weights to each operation in each layer. During training, the operation weights and network parameters are jointly optimized through gradient descent. The cross-domain customized supernetwork contains a two-layer graph neural network, and its search space includes basic node feature operations and edge operations that explicitly fuse edge features. The basic node feature operations include graph convolutional networks, graph attention networks, GraphSAGE, LightGCN, and linear layers.

9. The cross-domain recommendation system based on adversarial learning and graph neural architecture search as described in claim 6, characterized in that, In the adversarial learning, the global user representation input generator network G output by the cross-domain customized supernetwork is mapped to a user shared embedding in the shared embedding latent space; the user shared embedding is input to the domain discriminator network D; the generator network G and the domain discriminator network D are trained simultaneously during the optimization process, and a minimum-maximum game is formed by minimizing the generator loss and minimizing the discriminator cross-entropy loss.

10. The cross-domain recommendation system based on adversarial learning and graph neural architecture search as described in claim 6, characterized in that, The behavioral importance perceptron is an external module that guides the learning process by assigning weights to the loss of each sample in the source domain. The behavioral importance perceptron includes a comment semantic gating mechanism, which uses a pre-trained BERT model to extract semantic embeddings of interactive comment texts in the source domain, and calculates a comment gating value based on these semantic embeddings through a multi-layer fully connected network to measure the contribution of the comment to user preference modeling. The behavioral importance perceptron also includes a domain fit calibration mechanism, which adjusts the fusion strategy of source domain weights in real time based on the output of the adversarial discriminator. When the domain fit is higher than a threshold, the model relies on the global importance weights of the item; when the domain fit is lower than the threshold, the model switches to user-specific importance weights.

Images