Cross-domain recommendation method based on large model semantic structure enhancement

Abstract

The invention discloses a cross-domain recommendation method based on large model semantic structure enhancement, which belongs to the technical field of recommendation systems and comprises the following steps: S1, data acquisition; s2, intra-domain knowledge transfer: based on historical interaction information of users and projects shared in a source domain and a target domain, respectively constructing user-project interaction diagrams of the two domains; s3, semantic embedding representation generation: encoding input unstructured texts such as user comments and item descriptions through ChatGLM (ChatGLM); s4, construction of a user / project homogeneous matrix: generating a homogeneous graph structure reflecting user behavior tendency similarity and project content semantic relevance by calculating semantic representation similarity between users / projects; s5, semantic structure enhancement; and S6, algorithm prediction and optimization. According to the method, the universal capability of a large model in the aspects of semantic comprehension and knowledge migration is fully played, and high-level feature representation is extracted from multi-source semantic information, so that a more detailed user interest portrait with generalization capability is constructed.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation systems, and in particular relates to a cross-domain recommendation method based on large model semantic structure enhancement. Background Art

[0002] Modern information technology, particularly core technologies such as computers and the internet, has profoundly driven the digitization of various industries, accelerated information flows, and significantly transformed society's organizational structure and lifestyles. The widespread application of these technologies in real life has significantly improved the efficiency and convenience of various activities. However, with the rapid expansion of data volumes, the amount of information users are exposed to has increased dramatically, leading to increasingly prominent problems such as difficulty in information screening and reduced decision-making efficiency. Against this backdrop, recommender systems (RSs) have emerged and have become a crucial tool for alleviating information redundancy and improving the quality of information acquisition. By deeply mining and modeling user behavior patterns and item attributes, recommender systems can intelligently identify individual preferences and deliver personalized content, enhancing the user experience while also strengthening the platform's information scheduling capabilities.

[0003] Collaborative filtering is a classic recommendation algorithm that predicts users' potential preferences by mining similarities in historical behavior patterns between users or items. Collaborative filtering is mainly divided into user-based collaborative filtering and item-based collaborative filtering. User collaborative filtering methods identify user groups with interests most similar to the target user and use the interaction information of these users to model and predict the target user's preferences, thereby achieving personalized recommendations and alleviating the data sparsity problem to a certain extent. Item collaborative filtering focuses on mining similarities between items, recommending items to users that are similar to their historical interaction items. Although collaborative filtering has shown certain advantages in improving recommendation accuracy, it still has obvious limitations when dealing with data sparsity and cold start problems.

[0004] Cross-domain recommendation systems offer a new research path to alleviate data sparsity and cold-start problems. By fusing user behavior data and knowledge from multiple heterogeneous domains, these methods can exploit potential connections between domains. Combined with transfer learning, these methods effectively incorporate prior knowledge from the source domain to improve recommendation performance when data in the target domain is insufficient. These strategies not only alleviate the information shortage problem caused by sparse user behavior but also demonstrate greater adaptability and robustness in new domain recommendation tasks, becoming a key approach to improving the performance of recommendation systems. However, most existing cross-domain recommendation methods rely primarily on historical user interaction records across domains for modeling and reasoning. This strategy, based solely on behavioral data, has limitations in practical applications. Recommendation data often contains rich semantic information, such as text descriptions, review content, tags, and category attributes, which is crucial for understanding user intent and item characteristics. Summary of the Invention

[0005] To address the practical limitations of most existing cross-domain recommendation methods, this paper provides a cross-domain recommendation method based on semantic structure enhancement of a large model. This method fully leverages the general capabilities of the large model in semantic understanding and knowledge transfer, extracts high-level feature representations from multi-source semantic information, and constructs a more detailed and generalizable user interest profile.

[0006] The technical solution adopted by the cross-domain recommendation method based on large model semantic structure enhancement of the present invention is:

[0007] A cross-domain recommendation method based on large model semantic structure enhancement includes the following steps:

[0008] S1. Data Acquisition: Analyze the historical interaction data of the source and target domains respectively to extract the overlapping users between the two domains; construct the user-item interaction subsets of the source and target domains based on the overlapping users; and obtain the semantic information of the corresponding users and items by combining the historical interaction records of the users;

[0009] S2. Intra-domain knowledge transfer: Based on the historical interaction information of users and items shared in the source and target domains, user-item interaction graphs of the two domains are constructed respectively;

[0010] S3, semantic embedding representation generation: Encode the input unstructured text such as user comments and project descriptions through ChatGLM;

[0011] S4. Construction of user / item homogeneity matrix: By calculating the similarity of semantic representations between users / items, a homogeneous graph structure is generated that reflects the similarity of user behavior tendencies and the semantic relevance of item content;

[0012] S5. Semantic structure enhancement: Graph neural networks are used to model homogeneous graph structures and further inject semantic structure information from similar users or similar projects;

[0013] S6. Algorithm prediction and optimization: The representation obtained based on the intra-domain interaction graph is integrated with the semantic representation obtained through homogeneous graph structure enhancement. By calculating the inner product between the user and item representations, the matching degree is estimated to generate personalized recommendation results. Structural consistency loss and cross entropy loss are introduced to jointly guide the training process and improve recommendation accuracy.

[0014] A further improvement of the technical solution of the present invention is that: the step S2 is specifically,

[0015] Based on historical user interaction information, we construct user-item bipartite graphs in the source and target domains, respectively. Using graph neural networks as a feature propagation and aggregation mechanism, we iteratively transfer and fuse information from adjacent nodes within the graph structure to dynamically update and enhance the semantics of user and item representations within each domain. The specific representation is as follows:

[0016]

[0017] Among them, represents the embedding representation of the user node at layer l+1, represents the embedding representation of the project node at the l+1 layer, N(u) represents the set of neighbor nodes of the user, and N(v) represents the set of neighbor nodes of the project; the source domain and the target domain are consistent in the process of message transmission and aggregation; in order to clearly distinguish the feature representation of the nodes after aggregation in the two domains, we express it as in Is the initialization ID feature.

[0018] A further improvement of the technical solution of the present invention is that: the step S3 is specifically,

[0019] With the help of the powerful general language modeling capabilities of the large model ChatGLM, unstructured semantic information such as project descriptions and titles is encoded, thereby extracting high-quality feature representations containing rich semantic information; the specific modeling process of the project representation encoding process is shown below.

[0020]

[0021] in, Represents the semantic feature embedding representation; the specific user encoding process is as follows.

[0022]

[0023] A further improvement of the technical solution of the present invention is that: the step S4 is specifically,

[0024] By calculating the similarity of semantic representations between users / items, a homogeneous matrix reflecting the similarity of behavioral tendencies and the semantic relevance of item content is generated. The similarity calculation process is as follows;

[0025]

[0026] in, and It is expressed as the similarity between user / item i and user / item j; since the calculated similarity value is usually located in s i,j ∈(0,1), in order to facilitate subsequent modeling or simplify relationship representation, the similarity is binarized; specifically, a similarity threshold is set. When the similarity between users is greater than the threshold, they are considered to have a potential association relationship, and the corresponding value is set to s i,j =1; otherwise, it is considered that there is no obvious correlation between the two and it is set to s i,j =0.

[0027] A further improvement of the technical solution of the present invention is that: the step S5 is specifically,

[0028] Graph neural networks are used to model homogeneous graphs to achieve structural information enhancement of user and item semantic representations. Since the features generated by different semantic encoding models may differ in distribution and semantic space, direct similarity calculation or feature aggregation may lead to inconsistent feature space or insufficient information expression. Therefore, before structural enhancement, the semantic feature embedding representation is first input into a fully connected network (MLP) with a Tanh activation function to perform nonlinear transformation on the features. The specific transformation process is shown in Formulas (6) and (7).

[0029]

[0030] Among them, W U and W V is a trainable parameter, b U and b V is the corresponding bias value; the same MLP structure and parameters are used in both the source and target domains for feature transformation; then, the constructed user homogeneity matrix is ​​used Item Homogeneity Matrix Perform attribute structure information enhancement. The specific enhancement process is shown in formula (8).

[0031]

[0032] in, and represents the matrix representation of the original homogeneous matrix after adding the self-ring, and D is the corresponding degree matrix used for normalization.

[0033] A further improvement of the technical solution of the present invention is that: the step S6 is specifically,

[0034] The representation learned from the in-domain interaction graph is fused with the semantic representation enhanced by the homogeneous graph structure, and the inner product between the fused user and item representations is calculated to estimate their matching degree, thus generating personalized recommendation results.

[0035] The method of calculating the inner product is shown in formula (9);

[0036]

[0037] Among them, e u and e v Respectively represent the user and project representations used for fusion, is the model’s prediction of the user u’s interest score in the project v, σ() is the softmax activation function, The result is mapped to (0,1);

[0038] After convolution enhancement, the structural consistency loss is introduced. The loss constrains the structural enhancement features after convolution to maintain consistency with the interaction matrix. The specific loss function is defined as shown in formula (10);

[0039]

[0040] Among them, the user-level and item-level losses are shown in formula (11) and formula (12).

[0041]

[0042] in, and represents the embedding representation of the i-th user and item after three layers of convolution; n u and n v represent the number of users and items respectively, and Represents the structural information extracted from the two-hop user-item raw interaction relationship, which is used as a constraint to maintain the consistency of user and item structures;

[0043] The binary cross entropy loss function is also introduced as the optimization objective of the recommendation task; the specific formula definition is shown in (13);

[0044]

[0045] Among them, y u,v is the true rating of item v by user u, is the model's pre-score. |θ|2 is the regularization term;

[0046] The joint optimization objective is defined on the source domain and the target domain, and the overall loss function is shown in formula (14);

[0047]

[0048] Due to the adoption of the above technical solution, the technical advancements achieved by the present invention include:

[0049] In step S3 of the present invention, a pre-trained large model such as ChatGLM is used to perform deep semantic encoding on texts such as user comments and project descriptions. Compared with traditional methods such as TF-IDF, Word2Vec or BERT, ChatGLM is better at modeling contextual semantics and hierarchical relationships of intent, and can extract more abstract and generalized user interests and project features, providing a high-quality semantic foundation for knowledge transfer in cross-domain recommendations.

[0050] Step S5 of the present invention fully leverages the high-quality semantic embedding representation generated by ChatGLM in step S3. Based on the user-user and item-item homogeneous graphs constructed in step S4, the unstructured semantic similarity relationships are explicitly mapped as connecting edges in the graph structure, thereby completing the structural gaps caused by the sparsity of behavior in the original user-item interaction graph. Furthermore, by enhancing the attribute structure of nodes through the homogeneous graph, the stability of the node representation in the graph and the efficiency of information dissemination are significantly improved, effectively alleviating the cold start and data sparsity problems, and enhancing the robustness and generalization ability of the recommendation model in cross-domain scenarios.

[0051] Step S6 of the present invention introduces a structural consistency loss function to ensure that attribute structure information can effectively align with the collaborative structure information reflected in the user-item interaction relationship during the high-level modeling process. This loss term constrains the consistency between the attribute structure information after convolution enhancement and the original interaction matrix, guiding the attribute semantic features to more closely align with the user's historical behavior patterns during the optimization process, thereby achieving an effective fusion of collaborative signals and semantic similarity. This mechanism not only significantly alleviates the problem of feature distribution differences between the source domain and the target domain in cross-domain scenarios, but also further enhances the model's generalization ability and recommendation performance for unseen samples, constituting another key technological advancement of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a cross-domain recommendation method based on large model semantic structure enhancement of the present invention;

[0053] Figure 2 This is an architectural diagram of a cross-domain recommendation method based on large model semantic structure enhancement in the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. In the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concept of the present invention.

[0055] like Figure 1 and Figure 2 As shown, the present invention discloses a cross-domain recommendation method based on large model semantic structure enhancement, which includes the following steps:

[0056] S1: Data Acquisition. Analyze the historical interaction data of the source and target domains to extract overlapping users. Construct user-item interaction subsets for the source and target domains based on these overlapping users. Combine historical user interaction records to obtain semantic information about the corresponding users and items.

[0057] S2: Intra-domain knowledge transfer. Based on historical user interaction information, we construct bipartite user-item graphs in the source and target domains, respectively. Using graph neural networks as a feature propagation and aggregation mechanism, we iteratively transfer and fuse information from adjacent nodes within the graph structure to dynamically update and enhance the semantics of user and item representations within each domain. The specific implementation is as follows.

[0058]

[0059] Among them, represents the embedding representation of the user node at layer l+1, represents the embedding representation of the project node at the l+1 layer, N(u) represents the set of neighbor nodes of the user, and N(v) represents the set of neighbor nodes of the project. The source domain and the target domain are consistent in the process of message transmission and aggregation. In order to clearly distinguish the feature representation of the nodes after aggregation in the two domains, we express it as in Is the initialization ID feature.

[0060] S3: Leveraging the powerful general language modeling capabilities of the ChatGLM model, we encode unstructured semantic information such as project descriptions and titles, thereby extracting high-quality feature representations that contain rich semantic information. The specific modeling process for encoding project representations is shown below.

[0061]

[0062] in, Representing semantic feature embedding. Due to privacy concerns, the user personal information we obtain is limited and insufficient to support the required text requirements. To address the scarcity of user semantic information, this paper collects a collection of items purchased or rated by users through their historical interaction records and calculates the average semantic representation of these items as the user's semantic feature representation. The specific user encoding process is as follows.

[0063]

[0064] S4: By calculating the similarity of semantic representations between users (or items) (such as cosine similarity), a homogeneity matrix reflecting the similarity of behavioral tendencies and the semantic relevance of item content is generated. The similarity calculation process is as follows.

[0065]

[0066] in, and It is expressed as the similarity between user (or item) i and user (or item) j. Since the calculated similarity value is usually located at s i,j ∈(0,1) interval, in order to facilitate subsequent modeling or simplify relationship representation, the present invention performs binary processing on the similarity. Specifically, a similarity threshold is set. When the similarity between users is greater than the threshold, they are considered to have a potential association relationship, and the corresponding value is set to s i,j =1; otherwise, it is considered that there is no obvious correlation between the two and it is set to s i,j =0.

[0067] S5: Semantic structure enhancement. This homogeneous graph is modeled using a graph neural network to enhance the structured information of user and item semantic representations. Because the features generated by different semantic encoding models may differ in distribution and semantic space, direct similarity calculation or feature aggregation may lead to inconsistent feature spaces or insufficient information expression. Therefore, before performing structural enhancement, the semantic feature embedding representations are first input into a fully connected network (MLP) with a Tanh activation function to perform a nonlinear transformation on the features. The specific transformation process is shown in Formulas (6) and (7).

[0068]

[0069] Among them, W U and W V is a trainable parameter, b U and b Vis the corresponding bias value. It should be noted that the same MLP structure and parameters are used in both the source and target domains for feature transformation. The purpose is to maintain the consistency of the feature space of the two domains, thereby reducing the distribution difference between the domains, facilitating the alignment of semantic representations, and promoting the effective transfer of cross-domain knowledge. Then, the constructed user homogeneity matrix is ​​used Item Homogeneity Matrix Perform attribute structure information enhancement. The specific enhancement process is shown in formula (8).

[0070]

[0071] in, and represents the matrix representation of the original homogeneous matrix after adding the self-ring, and D is the corresponding degree matrix used for normalization.

[0072] S6: Algorithm prediction and optimization. The representation learned from the domain interaction graph is fused with the semantic representation enhanced by the homogeneous graph structure. The inner product between the fused user and item representations is calculated to estimate their matching degree, thereby generating personalized recommendation results. The inner product is calculated as shown in formula (9).

[0073]

[0074] Among them, e u and e v Respectively represent the user and project representations used for fusion, is the model’s prediction of the user u’s interest score in the project v, σ() is the softmax activation function, The result is mapped to (0,1).

[0075] To ensure that the semantic feature embedding representation can fully align the collaborative structure reflected by the user-item interaction relationship in the high-level modeling process, we introduce a structural consistency loss after convolution enhancement. This loss constrains the structural enhancement features after convolution to maintain consistency with the interaction matrix, guiding the semantic features to fit the user's historical behavior pattern during the optimization process. The specific loss function is defined as shown in Formula (10).

[0076]

[0077] Among them, the user-level and item-level losses are shown in formula (11) and formula (12).

[0078]

[0079] in, and Represents the embedded representation of the i-th user and item after three layers of convolution.u and n v represent the number of users and items respectively, and It represents the structural information extracted from the two-hop user-item raw interaction relationship and is used as a constraint to maintain the consistency of user and item structures.

[0080] In order to ensure that the representation of users and items can effectively capture the interactive behavior patterns, we also introduce the binary cross entropy loss function as the optimization objective of the recommendation task. The specific formula definition is shown in (13).

[0081]

[0082] Among them, y u,v is the true rating of item v by user u, is the model's pre-rating. |θ|2 is a regularization term used to prevent the model from overfitting.

[0083] Finally, the present invention defines a joint optimization objective on the source domain and the target domain, and its overall loss function is shown in formula (14).

[0084]

[0085] To verify the effectiveness of the present invention, five different subject areas were selected from the Amazon dataset to form three groups of cross-domain tasks: Sport-Cell, Toys-Groce, and Auto-Cell. The models proposed in the present invention were comprehensively compared from two directions: single-domain recommendation (BPRMF, NeuMF, NGCF, LightGCN) and cross-domain recommendation (CoNet, PPGN, BitGCF, DisenCDR, DRLCDR). The specific introduction of the relevant baseline models is as follows.

[0086] (1) BPRMF is a classic implicit feedback recommendation model that uses matrix decomposition as its basic modeling framework. It characterizes user interests and item features by mapping users and items into low-dimensional vectors in the same latent space. It learns personalized implicit feedback recommendation results by optimizing the user's relative ranking preference for positive and negative samples.

[0087] (2) NeuMF introduces neural networks into the recommendation system model. It combines the advantages of matrix factorization (MF) and multi-layer perceptron (MLP) to simultaneously model the linear features and nonlinear interactions between users and items. Compared with traditional BPR-MF, BPR only considers linear inner products to model user preferences, while NeuMF can learn more complex nonlinear relationships and performs better when dealing with complex behavioral features or high-order feature interactions.

[0088] (3) NGCF is a recommendation model based on graph neural networks. It regards users and items as nodes in the graph and the interactions between users and items as edges in the graph. By aggregating the feature information of neighbor nodes in the multi-layer propagation process through graph neural networks (GNN), it can effectively capture the nonlinear high-order interaction information between users and items, avoiding the excessive reliance of traditional methods on low-order relationships, thereby significantly improving the recommendation performance, especially in long-tail recommendations and cold start problems.

[0089] (4) LightGCN simplifies the network structure of NGCF and reduces the computational overhead and overfitting risk of the model by removing unnecessary operations (such as activation functions and fully connected layers), making the training and inference processes more efficient.

[0090] (5) CoNet uses a cross-connection approach to achieve cross-domain recommendations, effectively sharing and integrating collaborative information about users and projects across multiple domains. Through a joint learning approach, it optimizes data from two different domains, ensuring that the correlations between different domains can be fully mined and utilized.

[0091] (6) PPGN uses parallel graph convolutional networks to simultaneously learn multiple interactive information between users and items, and captures complex user behaviors and item features through multi-level graph information fusion.

[0092] (7) BitGCF models users and items as a bipartite graph structure. It aggregates neighborhood information through multi-layer graph convolution, learns the latent representation of users and items, and captures high-order user-item interactions. By overlapping users, knowledge is transferred between different domains, improving the recommendation effect in multiple domains simultaneously.

[0093] (8) DisenCDR decouples the user and item features from different fields for learning, so that the model can capture the unique information of each field and the shared information between fields, thereby improving the recommendation accuracy.

[0094] (9) DRLCDR decouples user representation into two independent parts: domain-specific representation and domain-conditional representation, to better model users' preference differences and sharing patterns in different domains.

[0095] Table 1-1 Comparative experimental results with various baseline methods on the HR@10 indicator

[0096]

[0097]

[0098] As shown in Table 1-1, through HR@10 and NDCG@10 evaluations on multiple domain datasets, we conclude that experimental results demonstrate that the proposed method significantly outperforms existing baseline models in performance evaluations on multiple public datasets, achieving state-of-the-art performance. Specifically, in the Sport-Cell task, the proposed method improves the HR@10 metrics by 30.4% and 18.1% in the source and target domains, respectively, and improves the NDCG@10 metrics by 47.3% and 25.8%, respectively. In the Toys-Groce task, the HR@10 metrics improve by 6.1% and 7.6% in the source and target domains, respectively, and the NDCG@10 metrics improve by 15.4% and 7.0%, respectively. Furthermore, in the Auto-Cell task, the HR@10 metrics improve by 15.1% and 9.9% in the source and target domains, respectively, and the NDCG@10 metrics improve by 5.6% and 9.1%, respectively. The significant improvement in experimental results fully verifies the effectiveness and robustness of this invention in cross-domain recommendation tasks, and further confirms that the semantic enhancement strategy based on large models has important theoretical significance and practical value in improving model performance.

[0099] In the above embodiment, the present invention provides a cross-domain recommendation method based on the semantic structure enhancement of the big model, which fully utilizes the general capabilities of the big model in semantic understanding and knowledge transfer, extracts high-level feature representations from multi-source semantic information, and is used to construct a more detailed and generalized user interest portrait.

[0100] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Any modifications and improvements made to the technical solution of the present invention by a person of ordinary skill in the art without departing from the design concept of the present invention shall fall within the scope of protection of the present invention. The technical content for which protection is sought in the present invention is fully set forth in the claims.

Claims

1. A cross-domain recommendation method based on semantic structure enhancement of a large model, characterized by: The following steps are included: S1. Data acquisition: Analyze the historical interaction data of the source domain and the target domain respectively, and extract the overlapping users between the two domains; Construct user-item interaction subsets of the source and target domains based on overlapping users; obtain semantic information of corresponding users and items by combining historical user interaction records; S2. Intra-domain knowledge transfer: Based on the historical interaction information of users and items shared in the source and target domains, user-item interaction graphs of the two domains are constructed respectively; S3, semantic embedding representation generation: Encode the input unstructured text such as user comments and project descriptions through ChatGLM; S4. Construction of user / item homogeneity matrix: By calculating the similarity of semantic representations between users / items, a homogeneous graph structure is generated that reflects the similarity of user behavior tendencies and the semantic relevance of item content; S5. Semantic structure enhancement: Graph neural networks are used to model homogeneous graph structures and further inject semantic structure information from similar users or similar projects; S6. Algorithm prediction and optimization: The representation obtained based on the intra-domain interaction graph is integrated with the semantic representation obtained through homogeneous graph structure enhancement. By calculating the inner product between the user and item representations, the matching degree is estimated to generate personalized recommendation results. Structural consistency loss and cross entropy loss are introduced to jointly guide the training process and improve recommendation accuracy.

2. The cross-domain recommendation method based on large model semantic structure enhancement according to claim 1, characterized in that: The step S2 is specifically as follows: Based on historical user interaction information, we construct user-item bipartite graphs in the source and target domains, respectively. Using graph neural networks as a feature propagation and aggregation mechanism, we iteratively transfer and fuse information from adjacent nodes within the graph structure to dynamically update and enhance the semantics of user and item representations within each domain. The specific representation is as follows: Among them, represents the embedding representation of the user node at layer l+1, represents the embedding representation of the project node at the l+1 layer, N(u) represents the set of neighbor nodes of the user, and N(v) represents the set of neighbor nodes of the project; the source domain and the target domain are consistent in the process of message transmission and aggregation; in order to clearly distinguish the feature representation of the nodes after aggregation in the two domains, we express it as in Is the initialization ID feature.

3. The cross-domain recommendation method based on large model semantic structure enhancement according to claim 1, characterized in that: The step S3 is specifically as follows: With the powerful general language modeling capabilities of the large model ChatGLM, unstructured semantic information such as project descriptions and titles is encoded to extract high-quality feature representations containing rich semantic information. The specific modeling process of the project representation encoding process is shown below. in, Represents the semantic feature embedding representation; the specific user encoding process is as follows; 4. The cross-domain recommendation method based on large model semantic structure enhancement according to claim 1, characterized in that: The step S4 is specifically as follows: By calculating the similarity of semantic representations between users / items, a homogeneous matrix reflecting the similarity of behavioral tendencies and the semantic relevance of item content is generated. The similarity calculation process is as follows; in, and It is expressed as the similarity between user / item i and user / item j; since the calculated similarity value is usually located in s i,j ∈(0,1), in order to facilitate subsequent modeling or simplify relationship representation, the similarity is binarized; specifically, a similarity threshold is set. When the similarity between users is greater than the threshold, they are considered to have a potential association relationship, and the corresponding value is set to s i,j =1; otherwise, it is considered that there is no obvious correlation between the two and it is set to s i,j =0.

5. The cross-domain recommendation method based on large model semantic structure enhancement according to claim 1 is characterized in that: The step S5 is specifically as follows: Graph neural networks are used to model homogeneous graphs to achieve structural information enhancement of user and item semantic representations. Since the features generated by different semantic encoding models may differ in distribution and semantic space, direct similarity calculation or feature aggregation may lead to inconsistent feature space or insufficient information expression. Therefore, before structural enhancement, the semantic feature embedding representation is first input into a fully connected network (MLP) with a Tanh activation function to perform nonlinear transformation on the features. The specific transformation process is shown in Formulas (6) and (7). Among them, W U and W V is a trainable parameter, b U and b V is the corresponding bias value; the same MLP structure and parameters are used in both the source and target domains for feature transformation; then, the constructed user homogeneity matrix is ​​used Item Homogeneity Matrix Perform attribute structure information enhancement. The specific enhancement process is shown in formula (8); in, and represents the matrix representation of the original homogeneous matrix after adding the self-ring, and D is the corresponding degree matrix used for normalization.

6. The cross-domain recommendation method based on large model semantic structure enhancement according to claim 1, characterized in that: The step S6 is specifically as follows: The representation learned from the in-domain interaction graph is fused with the semantic representation enhanced by the homogeneous graph structure, and the inner product between the fused user and item representations is calculated to estimate their matching degree, thus generating personalized recommendation results. The method of calculating the inner product is shown in formula (9); Among them, e u and e v Respectively represent the user and project representations used for fusion, is the model’s prediction of the user u’s interest score in the project v, σ() is the softmax activation function, The result is mapped to (0,1); After convolution enhancement, the structural consistency loss is introduced. The loss constrains the structural enhancement features after convolution to maintain consistency with the interaction matrix. The specific loss function is defined as shown in formula (10); Among them, the user-level and item-level losses are shown in formula (11) and formula (12); in, and represents the embedding representation of the i-th user and item after three layers of convolution; n u and n v represent the number of users and items respectively, and Represents the structural information extracted from the two-hop user-item raw interaction relationship, which is used as a constraint to maintain the consistency of user and item structures; The binary cross entropy loss function is also introduced as the optimization objective of the recommendation task; the specific formula definition is shown in (13); Among them, y u,v is the true rating of item v by user u, is the model's pre-score; |θ|2 is the regularization term; The joint optimization objective is defined on the source domain and the target domain, and the overall loss function is shown in formula (14);

Images