Enhancing GNN Learning on Relational Databases using Semantic Clustering-based Graph Enrichment

Relational databases (RDBs) have been widely utilized for managing and analyzing large-scale business data. However, analyzing RDBs with machine learning techniques requires complex feature engineering, which can be time-consuming, error-prone, and prone to information loss. Recent research in relational deep learning (RDL) attempts to map RDB data to graph structures, enabling direct learning via Graph Neural Networks (GNNs). Despite these efforts, current methods face challenges such as long path lengths between semantically related nodes, which negatively affect model performance. To address these limitations, this paper proposes a novel graph enrichment method that integrates both primary key-foreign key (PK-FK) relationships and semantic information derived from textual descriptions in RDB records. The proposed method utilizes semantic clustering on textual data to generate additional nodes and their corresponding edges. This enriched graph is then incorporated into the GNN model, enhancing the propagation of relevant information by reducing path lengths between semantically related nodes. Furthermore, this paper introduces a weight-based adjustment mechanism for the additionally generated nodes to account for potential uncertainty in clustering results. Experimental results demonstrate that the proposed framework improves the performance of GNNs compared to the RDL method.