Improved variable and value ranking techniques for mining categorical traffic accident data

The ever increasing size of datasets used for data mining and machine learning applications has placed a renewed emphasis on algorithm performance and processing strategies. This paper addresses algorithms for ranking variables in a dataset, as well as for ranking values of a specific variable. We propose two new techniques, called Max Gain (MG) and Sum Max Gain Ratio (SMGR), which are well-correlated with existing techniques, yet are much more intuitive. MG and SMGR were developed for the public safety domain using categorical traffic accident data. Unlike the typical abstract statistical techniques for ranking variables and values, the proposed techniques can be motivated as useful intuitive metrics for non-statistician practitioners in a particular domain. Additionally, the proposed techniques are generally more efficient than the more traditional statistical approaches.