Optimizing copious activity type classes based on classification accuracy and entropy retention

Abstract

Despite the advantages, big transport data are characterized by a considerable disadvantage as well. Personal and activity-travel information are often lacking, making it necessary to deduce this information with data mining techniques. However, some studies predict many unique activity type classes (ATCs), while others merge multiple activity types into larger ATCs. This action enhances the activity inference estimation, but destroys important activity information. Previous studies do not provide a strong justification for this practice. An objectively optimized set of ATCs, balancing model prediction accuracy and preserving activity information from the original data, becomes essential. Previous research developed a classification methodology in which the optimal set of ATCs was identified by analyzing all possible ATC combinations. However, for the US National Household Travel Survey (NHTS) 2009 data set which comprises 36 ATCs (home activity excluded), this approach is practically impossible in a finite amount of time since there would be 3.82*10^30 unique combinations. The aim of this paper is to optimize which original ATCs should be grouped into a new class, and this for data sets for which it is impossible or impractical to simply calculate all ATC combinations. The proposed method defines an optimization parameter U (based on classification accuracy and 18 information retention) which is maximized in an iterative search algorithm. The optimal set of ATCs for the NHTS 2009 data set was determined. A comparison finds that this optimum is considerably 20 better than many expert opinion activity type classification systems. Convergence was confirmed and performance gains were benchmarked.