Using Self-Organizing Feature Maps and Data Mining to Analyze Liability Authentications of Two-Vehicle Traffic Crashes

Abstract

In this study, both self-organizing feature maps and data mining models were generated for analysis of liability authentications of two-vehicle crashes, respectively, to evaluate their applicability on such an application. Factors critical to crash liability attributions were theoretically selected through Chi square test and F test. Meanwhile, factors considered important to committee members of government founded authentication committees in Taiwan were identified using fuzzy Delphi process. Factors commonly appeared in both theoretical results and practical opinions were considered truly critical in this study. Data mining models and self- organizing feature maps models were then generated with identified critical factors and associated liability attribution results of selected crash cases, for frontal, side, and rear collisions of two vehicle crashes, in an attempt to provide appropriate tools for decision support on crash liability authentications. Nine factors, viz. right-of-way, perception, speeding, lane changing, signal status, maneuver, irregularity, mutual position, and perception distance were identified critical. Data mining models can give 60 ~ 83% accurate liability authentications, yet are incapable of giving specific liability attributions corresponding to crash cases. As to SOM models, acceptable silhouette coefficient indicate that generated models can allocate cases to adequate clusters. Meanwhile, qualities of clustering were confirmed by calculated high grey relational coefficients. Although with small data size, a combination of data mining and SOM models were considered to be able to give reasonably good liability attributions predictions and references on given crash cases.