Discrete Particle Swarm Optimization Based Filter Feature Selection Technique for the Severity of Road Traffic Accident Prediction

Abstract

The rate of traffic accident is increasing worldwide, which resulted to severe injuries, fatalities and economic loses. More than 1.25 million people lose their lives every year and 50 million are non-fatally wounded in road accident worldwide. The severity of road traffic accident are attributed to many factors which include human (age, gender), road (road alignment, roads surface), and environment (moisture, weather) among others. Nevertheless, some factors are more significant than others in increasing the severity level of injuries sustained by victims of accident. In previous studies, few feature selection techniques were used to get the most relevant and non-redundant features prior to prediction of severity of road traffic accident. Among the techniques of feature selection which includes filter, wrapper, embedded and hybrid, only filter technique has been considered. However, the filter feature selection techniques (filter correlation, mutual information) are inadequate at finding optimal features. Additionally, the redundant feature doesn’t only affect the performance of the classification algorithm but also computational time. To this end, this paper, Proposed a hybrid discrete Particle Swarm Optimization with Fast Correlation Based Feature Selection (DPSO_FCBF) technique for selecting the most significant features for the prediction of severity of road traffic accident. In this technique, the discrete particle swarm optimization (DPSO) search for the optimum values and the fast correlation based feature selection (FCBF) evaluate the selected optimum values to select the most relevant features, then DPSO update the personal best (Pbest) and global best (Gbest) for the optimal solution. The proposed technique has been evaluated using benchmark road traffic accident dataset. The experimental result are compared with the result obtained from DPSO and FCBF individually, the proposed technique showed better result in terms of number of features selected, prediction accuracy and time complexity.