Virtual Semantic Knowledge Graphs and Negative Feedback Enhance Recommender System Accuracy

Published in Data Science and Engineering, the study resolves structural limitations in recommender systems by capturing latent semantic associations across knowledge graph relations and modeling negative user feedback. Across extensive evaluations on three real-world benchmark datasets spanning movies, books, and commercial businesses, the proposed architecture outperformed competitive baseline models. On MovieLens-1M, the framework improved top-twenty Recall by 23.29 percent and top-twenty Normalized Discounted Cumulative Gain by 25.86 percent. By combining semantic relation clustering with signed topological matrix factorization, the system establishes a robust mathematical foundation for mitigating data sparsity and preference bias in modern recommendation platforms.

In digital platforms confronting exponential information growth, automated recommender systems serve as critical tools for filtering catalog inventories. Early collaborative filtering models extracted latent features from interaction records, yet these methods struggle when interaction data is sparse or when newly introduced items lack prior engagement, creating the cold-start dilemma. To resolve these challenges, researchers incorporated knowledge graphs as auxiliary structured data, connecting real-world entities through defined relation types. Early techniques relied on first-order translational embedding algorithms like the Trans series or mapped long-range entity paths across heterogeneous networks, but path-based approaches demand excessive memory on expansive graphs. Recent graph neural networks aggregate multi-hop neighbor messages across stacked layers, yet existing architectures process different relational edge types in isolation, ignoring underlying semantic similarities between disparate relation categories.

Handling graph relations independently introduces representational bias because distinct relational edges often express equivalent semantic meaning. In movie recommendation, a system observing that a user watched Dune might suggest Little Women based strictly on a shared actor relation, but fail to recommend Incendies because the shared director and screenwriter relationship lacks an exact edge match. Because directors and screenwriters exhibit strong semantic similarity as creative authors, recognizing relational equivalence allows algorithms to uncover relevant items that isolated relation graphs overlook. Simultaneously, existing models focus almost exclusively on positive interactions, ignoring diagnostic signals embedded in negative feedback like low ratings. Disliking a movie like Wonka indicates that its prequel should not be recommended, while also signaling decreased affinity for its lead actor, demonstrating that ignoring negative interactions distorts preference modeling.

To formalize these dynamics mathematically, the researchers defined discrete user sets, item sets, and historical interaction records. Each interaction entry pairs a user and item with a feedback label, where positive feedback is assigned one, negative feedback negative one, and unobserved pairs zero. In alignment with established sign-aware recommendation conventions, explicit five-star ratings were converted into implicit interaction graphs by treating top ratings of five as positive feedback and ratings from one to four as negative feedback. This procedure yields two bipartite interaction graphs: a positive graph containing favorable interactions and a negative graph containing unfavorable connections. Alongside interaction data, an external knowledge graph encodes entities and relations as structural triples, where catalog items align directly with corresponding knowledge graph entities to supply contextual metadata that enriches collaborative learning.

The SRVSKG architecture operates through four integrated modules: a virtual semantic subgraph collaborative representation module, a sign-aware learning module, a multi-feature fusion prediction module, and a joint optimization module. Within the virtual semantic module, the authors introduced virtual semantic relation dividing to automatically cluster knowledge graph relations. The model defines a hyperparameter representing the quantity of virtual semantic relations and initializes a virtual relation matrix using normally distributed vectors. For each original relation, the framework calculates cosine similarity against every virtual relation vector, assigning the original edge type to the highest-scoring virtual relation via an argmax function. This reorganization divides the knowledge graph into discrete virtual semantic subgraphs, while gradient descent jointly refines the virtual relation matrix during training to ensure optimal grouping of semantically related edge types.

Building upon these partitioned subgraphs, the system generates item embeddings using graph attention hierarchical feature extraction. For each entity in a virtual semantic subgraph, message aggregation weights neighbor contributions based on topological relevance, positing that structurally similar neighbors transmit more informative signals. The model calculates unnormalized attention scores through an inner product between the entity embedding and the hyperbolic tangent activation of its neighbor embedding, normalizing scores using a softmax function. Weighted neighbor embeddings are accumulated and combined with the target node embedding across multiple propagation iterations. Repeating this process across all virtual subgraphs produces distinct semantic representations fused via learnable attention weights to establish first-order item embeddings. Stacking multiple graph neural network layers extends propagation across multi-hop paths to capture deep high-order contextual semantics.

To formulate comprehensive user representations, the virtual semantic module incorporates a dedicated user-attribute preference graph alongside direct interaction aggregation. User embeddings are initially learned by aggregating representations of positively interacted items. Recognizing that users demonstrate varying affinities for specific item characteristics, the authors constructed an auxiliary preference graph linking users directly to first-order attribute entities connected to positively rated items in the knowledge graph. Executing message aggregation over this user-attribute graph generates an attribute-level user embedding, which is summed with the direct collaborative embedding to yield the final user representation. Because items, entities, and attributes share unified feature embeddings, structural information propagates concurrently across interaction and knowledge networks across stacked graph layers.

In parallel with knowledge graph modeling, the sign-aware learning module captures topological signals from positive and negative interaction records using spectral graph theory. Spectral properties derived from graph Laplacian matrices offer effective mathematical characterizations of global topology. The authors combined the Laplacian matrices of the positive and negative interaction graphs into a unified sign-aware Laplacian matrix modulated by a weighting hyperparameter, alpha, bounded between negative one and positive one. Positive alpha values push recommendations away from disliked items, whereas negative alpha values permit controlled proximity to negatively rated content. Decomposing this combined Laplacian reveals its eigenvalues and eigenvectors, where the eigenvector corresponding to the smallest eigenvalue captures the global topological structure of the signed graph. Multiplying this minimal eigenvector by its transpose and a learnable scaling parameter yields a sign-aware feature matrix encoding macro-level structural relationships.

To dynamically integrate these sign-aware spectral features into node representations, the authors deployed a multilayer Transformer architecture. Transformers utilize self-attention mechanisms to compute similarities between users and items while aggregating context from connected nodes. Diverging from standard Transformer designs that utilize distinct projection matrices for query, key, and value transformations, SRVSKG directly assigns concatenated user-item embeddings to all three attention roles, markedly reducing training complexity. Within each Transformer layer, the sign-aware spectral feature matrix is added directly to the scaled query-key dot product inside the softmax function, aligning the attention mechanism with signed graph topology. After generating sign-aware embeddings, the multi-feature fusion module sums user and item vectors from the virtual semantic subgraph module with corresponding vectors from the sign-aware module, calculating final recommendation scores through standard inner products.

Model training employs Bayesian Personalized Ranking loss functions within a multi-task learning framework to ensure that observed interaction items receive higher predicted scores than unobserved candidates. The training procedure establishes separate positive and negative triplet sample sets, alternating between a loss function optimizing the virtual semantic subgraph module on positive samples and a second loss function optimizing the sign-aware module across both positive and negative samples. The overall loss sums both components alongside an L2 regularization penalty controlled by a regularizer hyperparameter, optimized via the Adam algorithm.

The authors evaluated SRVSKG on three real-world benchmarks: MovieLens-1M containing one million ratings from 6,040 users aligned via Microsoft Satori, Amazon-Book filtered under a 10-core threshold with Freebase mappings, and Yelp filtered under a 20-core threshold using commercial business metadata. Experiments utilized an 80-20 train-test partition, measuring top-twenty Recall, top-twenty Normalized Discounted Cumulative Gain, and Area Under the Receiver Operating Characteristic Curve for Click-Through Rate prediction. Comparative baselines spanned factorization models like FM and NFM, graph neural network models including KGAT, KGIN, CG-KGR, VRKG, and KGTN, and signed graph architectures like SLGNN and SIGformer. SRVSKG outperformed all baselines on MovieLens-1M and Yelp, ranking second only on Amazon-Book Area Under the Curve while leading in retrieval ranking. Specifically, MovieLens-1M exhibited relative gains of 23.29 percent in Recall and 25.86 percent in Normalized Discounted Cumulative Gain, while Amazon-Book showed 6.82 percent Recall improvements and Yelp demonstrated consistent gains across all metrics.

Ablation studies and parameter investigations confirmed the specific necessity of each architectural component. Variants removing the virtual semantic knowledge graph module, the sign-aware module, or user-attribute preferences showed substantial performance degradation across all datasets, with negative feedback exclusion causing the steepest declines on sparse datasets like Amazon-Book and Yelp. Parameter analysis revealed an inverted U-shaped performance curve for the number of virtual relations, peaking at three for MovieLens-1M and four for the larger Amazon-Book and Yelp datasets, showing that excessive clustering fragments semantic concepts. The negative influence weight alpha demonstrated an optimal setting around negative 0.2 before stabilizing. Finally, a qualitative case study on user fifty-six in MovieLens-1M verified that writer and director relations clustered into a single virtual relation, while low predicted scores were assigned to items connected by shared directors to previously disliked movies, proving the framework successfully suppresses undesired recommendations.

Subject of Research: Graph neural network recommender systems incorporating virtual semantic knowledge subgraphs and sign-aware feedback modeling

Article Title: Sign-aware Recommendation Based on Virtual Semantic Knowledge Graph

Article References: Ju, S., Zhang, Y., Cai, T., Zhao, R., & Sun, J. (2026). Sign-aware Recommendation Based on Virtual Semantic Knowledge Graph. Data Science and Engineering, 11(1), 100-115. https://doi.org/10.1007/s41019-025-00319-5

Image Credits: AI Generated

DOI: 10.1007/s41019-025-00319-5

Keywords: recommender systems, knowledge graphs, sign-aware learning, graph neural networks, graph attention networks, virtual semantic subgraphs, negative feedback modeling, Laplacian matrix factorization, Transformer architecture, collaborative filtering, user-attribute preference graph, Bayesian Personalized Ranking, MovieLens-1M, Amazon-Book dataset, Yelp dataset, relation clustering, spectral graph theory, click-through rate prediction