Pattern Mining and Predictive Inference on Short-Term Weather and Collision Time Series Data

Abstract

Collisions are rare random events. Many traffic safety indexes with a small-sized temporal or spatial unit, e.g., daily collisions of a city or a regional highway network, are highly random and fiercely fluctuated. The descriptive and inferential analyses for this type of short-term collision time series data, abbreviated as SCTS data in this paper, are still not well-established yet. This paper is to tackle this issue by a newly emerging approach—pattern mining combined with data mining methods. Based on a collision database, calendar information, and historical weather records, the approach of descriptive statistics was employed to illustrate correlations between all data items and to identify main affecting factors for a SCTS response, with respective to single variable pattern, variable pair and multiple variable correlations. Then the structure and flow-chart of the major attributes led to different SCTS outputs were further investigated by means of decision tree method. The established decision tree structure was then utilized to predict SCTS values of future days as consequence from their calendar characters and weather forecasts. The approaches of description and inference of SCTS data developed in this paper filled in the methodological vacancy of discovering SCTS data pattern and to infer their attributes. The study of this paper also provided a viable solution to predict SCTS and therefore help to pre-schedule safety countermeasures for practitioners.