Classifying Imbalanced Road Accident Data Using Recurring Concept Drift

Abstract

In New Zealand, road accident casualties have been increasing. Factor analyses and time series analyses show what types of accidents result in casualties, but the results from the analysis can become outdated. We propose a stream classification framework with drift detection to signal and adapt when the factors associated with crash casualties change over time. We propose a drift detection framework, G-mean Adaptive drift Detection (GAD), which adapts a classifier threshold to maximise G-mean. This metric rewards maximising accuracy on each class while keeping these accuracies balanced. As a result, GAD can make concept drift in the minority class easier to detect. We also propose a recurring concept classification framework, G-mean Concept Profiling Framework (GCPF), which reuses previously trained classifiers and uses GAD’s approach to drift detection. Through experimentation, we show GAD improves G-mean without increasing false positive drift detection on imbalanced synthetic and real world datasets. We also show GCPF achieves better G-mean than other state-of-the-art stream classification approaches on the NZ crash data set.