Social Knowledge Recommendation Based on Graph Reconstruction

Introduction

The rapid development of the internet has led to an overwhelming amount of information, making it increasingly difficult for users to find relevant content. Recommendation systems have emerged as essential tools to filter and personalize information, helping users navigate through vast datasets efficiently. Traditional recommendation models primarily focus on explicit relationships between users and items, often neglecting the high-order global characteristics of graph structures and the implicit interests of users. To address these limitations, this paper introduces a novel approach called Social Knowledge Recommendation based on Graph Reconstruction (SKRGR), which leverages graph reconstruction techniques to enhance recommendation performance by capturing both global and local structural patterns.

Existing recommendation models, such as collaborative filtering and matrix factorization, rely heavily on explicit user-item interactions. While these methods have shown success, they often fail to account for the complex relationships embedded in social networks and knowledge graphs. Social recommendation models incorporate trust relationships between users, assuming that socially connected users share similar preferences. However, the sparsity of social networks and the complexity of relationships make it challenging to accurately model user interests.

Knowledge graphs, which represent entities and their relationships in a structured format, provide a rich source of contextual information that can enhance recommendation systems. By integrating knowledge graphs with social recommendation models, it becomes possible to capture both semantic and social influences, leading to more accurate and interpretable recommendations. Graph Neural Networks (GNNs) have further advanced recommendation systems by efficiently modeling non-linear relationships through iterative message passing and neighborhood aggregation. Despite their success, GNN-based models often focus on local node features while overlooking the global properties of graph structures.

To overcome these challenges, SKRGR introduces a graph reconstruction framework that decomposes the user-item knowledge graph into multiple subgraphs, each representing different types of relationships. These subgraphs are independently encoded using relation-based graph attention networks and multi-layer GNNs. Additionally, a neighborhood enhancement strategy is employed to refine local feature representations, while a global contrastive learning mechanism ensures consistency between global and local embeddings. Experimental results on benchmark datasets demonstrate that SKRGR outperforms existing models, achieving significant improvements in both recall and NDCG metrics.

Related Work

Social Recommendation

Social recommendation models leverage social relationships to improve recommendation accuracy. The underlying assumption is that users with social connections tend to have similar preferences, a concept known as social homophily. Early approaches, such as TrustMF, integrated social trust relationships into matrix factorization models, ensuring that socially connected users have similar latent representations. More recent methods employ Graph Convolutional Networks (GCNs) to propagate user preferences through social networks. For example, DiffNet++ models social influence and interest diffusion by simulating how user preferences spread across a network. Attention mechanisms have also been incorporated to differentiate the importance of various social relationships, leading to more precise user and item embeddings.

Despite these advancements, social recommendation models still face challenges due to data sparsity and noise in social networks. Some approaches attempt to mitigate these issues by introducing self-supervised signals or contrastive learning techniques to enhance node representations. However, most existing models still struggle to balance local and global structural information effectively.

Knowledge Graph-Based Recommendation

Knowledge graphs provide a structured representation of entities and their relationships, offering valuable contextual information for recommendation systems. Methods like KGAT (Knowledge Graph Attention Network) use attention mechanisms to propagate high-order connectivity information across knowledge graphs. Other approaches, such as CKAN (Collaborative Knowledge-Aware Attentive Network), combine collaborative filtering with knowledge-aware embeddings to enrich node representations.

Recent studies have explored intent-aware recommendation models that leverage knowledge graphs to uncover user preferences. For instance, some frameworks use path aggregation mechanisms to model user-item interactions at an intent level, while others employ contrastive learning to align semantic and collaborative signals. These methods demonstrate the potential of knowledge graphs in improving recommendation diversity and interpretability. However, they often treat knowledge graphs as static structures, failing to exploit their dynamic and hierarchical nature fully.

Contrastive Learning

Contrastive learning, originally developed for computer vision tasks, has gained traction in recommendation systems. The core idea is to maximize the similarity between positive pairs (e.g., different views of the same node) while minimizing the similarity between negative pairs (e.g., unrelated nodes). In recommendation systems, contrastive learning has been used to generate robust node embeddings by creating augmented views of the interaction graph.

Some methods, such as SGL (Self-Supervised Graph Learning), apply random node or edge dropout to construct contrastive views. Others, like SimGCL, argue that graph augmentation is unnecessary and focus instead on optimizing the contrastive loss function. Recent work has extended contrastive learning to heterogeneous graphs, where different types of nodes and edges coexist. These approaches highlight the importance of maintaining consistency between global and local representations, a principle that SKRGR incorporates through its global contrastive learning mechanism.

Methodology

Overview of SKRGR

The SKRGR model consists of four main components: subgraph reconstruction, feature encoding, contrastive learning, and neighborhood enhancement. The framework begins by decomposing the user-item knowledge graph into multiple subgraphs, each representing distinct relationship types. These subgraphs are then independently encoded to capture both global and local structural patterns. A contrastive learning mechanism ensures consistency between global and local embeddings, while a neighborhood enhancement strategy refines node representations by integrating contextual information. Finally, a gated fusion mechanism dynamically combines global and local features to generate the final recommendations.

Subgraph Reconstruction

The user-item knowledge graph is a heterogeneous network comprising user and item nodes connected through various relationships, including social links, interactions, and collaborative similarities. To facilitate fine-grained modeling, SKRGR reconstructs this graph into four subgraphs:

User-Item Subgraph: Captures explicit interactions between users and items, such as ratings or purchases.
User Social Subgraph: Represents trust relationships between users.
User Collaboration Subgraph: Models similarities between users based on shared preferences.
Item Collaboration Subgraph: Encodes similarities between items based on co-interactions.

The reconstruction process uses the Jaccard similarity coefficient to measure node similarities. If the similarity between two nodes exceeds a predefined threshold, they are connected in the corresponding collaboration subgraph. This approach ensures that only meaningful relationships are preserved, reducing noise and improving computational efficiency.

Feature Encoding

Each subgraph is encoded using specialized neural architectures tailored to its characteristics.

• User-Item Subgraph: A relation-based graph attention network assigns adaptive weights to neighboring nodes, distinguishing between different interaction strengths. For example, a high-rated item receives more attention than a low-rated one.

• User Social, User Collaboration, and Item Collaboration Subgraphs: These subgraphs are encoded using three-layer GNNs, which aggregate neighborhood information through iterative message passing. The user social and collaboration embeddings are combined to form a unified user representation, while item embeddings are derived solely from the item collaboration subgraph.

Contrastive Learning

To align global and local representations, SKRGR employs a global contrastive learning mechanism. The user-item subgraph provides global embeddings, while the collaboration subgraphs offer local embeddings. These embeddings are projected into a shared space, where contrastive loss is applied to maximize agreement between corresponding global and local representations.

The contrastive loss function is based on the InfoNCE framework, which treats embeddings of the same node as positive pairs and embeddings of different nodes as negative pairs. This approach ensures that the model learns consistent representations across different subgraphs while preserving the unique characteristics of each view.

Neighborhood Enhancement

Local node representations are further refined through a neighborhood enhancement strategy. For users, this involves integrating features from similar users and items they have interacted with. An attention mechanism assigns higher weights to neighbors with stronger relevance, ensuring that the most influential signals dominate.

For items, neighborhood enhancement aggregates features from similar items and the users who interacted with them. A residual connection preserves the original item embeddings, preventing information loss during aggregation.

Gated Fusion

The final step combines global and local embeddings using a gated fusion mechanism. This adaptive strategy dynamically adjusts the contribution of each embedding based on its relevance, ensuring a balanced integration of global structural patterns and local contextual details. The fused embeddings are then passed through a multi-layer perceptron (MLP) to generate the final recommendation scores.

Experiments

Datasets and Baselines

SKRGR was evaluated on two benchmark datasets: Ciao and Epinions. Both datasets contain user-item interactions and social trust relationships, with Epinions being significantly larger and sparser than Ciao.

The model was compared against several state-of-the-art baselines, including:
• BPR and FM: Traditional matrix factorization methods.

• TrustMF: Integrates social trust into matrix factorization.

• NGCF and LightGCN: GNN-based collaborative filtering models.

• DiffNet++: Models social influence diffusion.

• DSL and CLDS: Contrastive learning-based social recommendation models.

Performance Evaluation

SKRGR consistently outperformed all baselines across both datasets. On Ciao, it achieved an average improvement of 11.00% in recall and 15.06% in NDCG. On Epinions, the improvements were 6.54% and 11.74%, respectively. These results highlight SKRGR’s ability to capture both global and local patterns effectively.

Ablation Studies

Ablation studies confirmed the importance of each component:
• Removing neighborhood enhancement led to a significant drop in performance, particularly for user-side enhancements.

• Excluding social relationships reduced recommendation accuracy, especially on Ciao, where social links are denser.

• Experiments with different similarity thresholds for collaboration subgraphs revealed that stricter thresholds for user similarities yielded better results.

• The gated fusion mechanism proved superior to simpler fusion strategies like concatenation or average pooling.

Conclusion

SKRGR represents a significant advancement in social knowledge recommendation by addressing key limitations of existing models. Its graph reconstruction framework enables fine-grained modeling of diverse relationships, while contrastive learning and neighborhood enhancement ensure robust and consistent representations. The model’s superior performance on benchmark datasets demonstrates its potential for real-world applications.

Future work could explore techniques to reduce computational overhead, such as distributed training or subgraph sampling, making SKRGR more scalable for large-scale networks. Additionally, incorporating dynamic graph updates could further enhance its adaptability to evolving user preferences.

doi.org/10.19734/j.issn.1001-3695.2024.05.0145

Was this helpful?

0 / 0