CAST-FC: Context-Aware Spatio-Temporal Feature Selection for Classification in Heterogeneous Urban Environments

Abstract

The growth in complexity of traffic systems in urban settings has rendered crash severity prediction as an increasingly difficult task mainly due to high heterogeneity of data in space and time. Conventional feature selection algorithms often make an assumption that the relevance of certain features holds true in the whole data set without any consideration for possible heterogeneities that exist within different portions of it. To overcome this limitation, this paper proposes a novel Context-Aware Spatio-Temporal Feature Selection (CAST-FS) technique which leverages local contextual information when selecting features from heterogeneous urban traffic data sets. In particular, the data set is partitioned into homogeneous groups according to its geographic and temporal characteristics by applying K-means clustering algorithm, and then a LightGBM classifier is applied locally within each group to estimate local feature importances. The local feature importance scores obtained in the process are integrated using cluster size-based weights to produce a global ranking of features. After that, selected feature subset which consists of the highest-ranking features is used to build a final classifier. The experiment in this study is conducted on the publicly available New York City Motor Vehicle Collisions dataset. It uses a binary classification approach for the task of crash severity prediction. In order to evaluate the effectiveness of the proposed approach, both threshold-independent and threshold-dependent evaluation metrics, such as AUC-ROC, Average Precision, accuracy, precision, recall, and F1-score, are employed. Results show that compared to global LightGBM-based and mutual information approaches CAST-FS demonstrates higher performance in AUC-ROC and Average Precision while shrinking the number of used features from 61 to 12. At the same time, imbalanced data classes lead to poor recall results when applying a fixed threshold.