Compressive sensing based feature selection: A case study for commuter behaviour modelling

We present a novel algorithm to address the problem of feature selection using the compressive sensing (CS) model. In the proposed algorithm, the candidate feature set is first regarded as a basis dictionary in CS, and then features that minimize the output error are selected. As a result, the selected features have a direct correspondence to the target samples, thereby achieving a better prediction performance. Compared to traditional dimensionality reduction algorithms, the proposed algorithm neither uses the problem-dependent parameters nor requires additional computation for the eigenvalue decomposition. Experimentally, the proposed algorithm is evaluated using a real-life problem for commuter behaviour modelling. The experimental results show that the proposed method is better or competitive when compared to existing feature selection methods.