Exploiting Pairing Attribute-Based VDM for Enhanced Similarity Learning

The value difference metric (VDM) is a well-established similarity measure for nominal attributes in classification tasks. However, it suffers from a critical limitation: it assigns a zero distance to differing attribute values with identical class distributions, reducing discriminatory power. To address this, we propose the pairing attribute value difference metric (PAVDM), which enhances similarity evaluation by jointly considering pairs of attribute values. While PAVDM improves discrimination, it introduces higher computational costs. To mitigate this, we introduce two optimization strategies: CSPAVDM, which leverages Cramér’s V for correlation-based pairing, and ASPAVDM, which employs AdaBoost to prioritize impactful attributes. Results show that PAVDM and its optimized variants outperform classical VDM in accuracy, precision, F۱-score, and ROC AUC under a fair evaluation protocol.

کلیدواژه ها:

Value difference metric ، Similarity criterion ، Learning metric ، Similarity metric