Multi-View Decision Fusion with Association Rules- and Distributions-Based Similarity Matches for Personal Credit Risk Assessment

Abstract

Financial big data has been mushrooming currently in accompany with the accelerative expansion of human demand and market scale, resulting in the emergence of a large number of credit data. Also due to the rapid development of data collection capabilities, the data structures of customer credit characteristics become more and more complex, causing more difficulties in data analysis. This paper presents a novel personal credit risk assessment (PCRA) method characterized by multi-view decision fusion based on association rules- and distributions-based similarity matches. The method considers the similarities between customer credit information reflected from five views: coding difference, numerical difference, direction difference, differences of joint distributions and marginal distributions. Correspondingly, the dissimilarity measures including Hamming distance, Chebyshev distance, Cosine distance, Hotelling  statistics, and Jensen-Shannon divergence are used to quantize respective similarities. In particular, by considering that the identification performance based on the first three distance-based measures tends to be affected by high-dimensionality and information redundancy of original features, association rules are used to train crucial items and item sets for dimension reduction to yield compact features for per subject involved. To utilize the complementarity of different views of decision information that actually exist discriminative power differences, our method forms the final decision via weighted voting for the individual decisions made by five independent matches. The experimental results on the UCI data set show that association rules-derived compact features can markedly improve the classification performance of distance measure-based matches, making the fused decision more discriminative for PCRA.