New Representations in PSO for Feature Construction in Classification

Feature construction can improve the classification performance by constructing high-level features using the original low-level features and function operators. Particle swarm optimisation (PSO) is an powerful global search technique, but it cannot be directly used for feature construction because of its representation scheme. This paper proposes two new representations, pair representation and array representation, which allow PSO to direct evolve function operators. Two PSO based feature construction algorithms (PSOFCPair and PSOFCArray) are then developed. The two new algorithms are examined and compared with the first PSO based feature construction algorithm (PSOFC), which employs an inner loop to select function operators. Experimental results show that both PSOFCPair and PSOFCArray can increase the classification performance by constructing a new high-level feature. PSOFCArray outperforms PSOFCPair and achieves similar results to PSOFC, but uses significantly shorter computational time. This paper represents the first work on using PSO to directly evolve function operators for feature construction.