A Perceptual Computing Approach for Learning Interpretable Unsupervised Fuzzy Scoring Systems

Scoring the driver's behavior through the analysis of his/her road trip data is an active area of research. However, such systems suffer from a lack of explainability, integration of expert bias in the calculated score, and ignoring the semantic uncertainty of various variables contributing to the score. To overcome these limitations, we have proposed a novel perceptual computing based unsupervised scoring system. The prowess of the proposed system has been exemplified in a case study of driver's scoring from telemetry data. Our proposed approach yields scores that showed a higher significant separability between drivers performing responsible or irresponsible (aggressive or drowsy) driving behaviors, than the formal method of computing these scores (a
$p$
value of
$3.94\boldsymbol{\times}10^{-4}$
and
$3.42\boldsymbol{\times}10^{-3}$
, respectively, in a Kolmogorov–Smirnov test). Further, the proposed method displayed higher robustness in the bootstrap test (where 30% of original data were omitted at random) by providing scores that were 90% similar to the original ones for all results within a confidence interval of 95%.