Hyperdimensional Computing Improves Tabular Data Querying with Interpretable Thresholds

A new approach leverages HyperDimensional Computing (HDC), specifically the Holographic Reduced Representations (HRR) model, to enhance tabular data embeddings for structured querying. Current embedding techniques, while useful for tasks like schema matching and table search, suffer from a fundamental flaw: their similarity scores lack intrinsic meaning, making it difficult to establish reliable thresholds for determining true matches or identifying when no valid answer exists. This research addresses this limitation by exploiting the algebraic properties of HDC operations. The authors derived closed-form expected similarity values for both equality and non-equality retrieval predicates. These values converge to interpretable metrics as dimensionality increases, enabling the setting of principled retrieval thresholds and reliable detection of zero-match scenarios. Evaluations compared HDC against EmbDI, a graph-based baseline, across two real-world datasets with varying table sizes and predicate lengths. The results demonstrated that HDC either matched or surpassed EmbDI in row retrieval performance across all configurations. Furthermore, HDC proved more robust in handling non-equality predicates and achieved perfect attribute projection accuracy at sufficient dimensionality, uniquely offering reliable zero-match detection through its interpretable thresholds.