Particle Swarm Optimization for Feature Selection in Emotion Categorization

Emotion categorization plays an important role in understanding human emotions by artificial intelligence systems such as robots. It is a difficult task as humans express many features, which vary over time when showing an emotion. Thus, existing classification techniques are overwhelmed, and the creation of a subset of appropriate features is needed. Feature selection can be used to improve the performance of an emotion categorization task by selecting a subset of features. This removes irrelevant features. Particle swarm optimization (PSO) is a meta-heuristic algorithm which has demonstrated excellent performance in feature selection tasks. However, traditional PSO algorithms often get trapped in local optima as they use their personal best and global best to determine their search direction, which may lead to premature convergence. In this paper, we present a time-based PSO variant by introducing a time-constant into the velocity update function of the PSO algorithm to avoid premature convergence, particularly in an emotion video-frame dataset. The method has been incorporated into binary and continuous PSO, then compared with the two standard versions on an emotion video-frame (CK+) dataset, as well as on static emotional datasets (i.e. the JAFFE and NIMH-ChEFS) to ensure that bias has not been introduced into the algorithm. While the time-based PSO variant (both binary and the continuous PSO) have achieved non-significantly higher performance than the standard PSO algorithms on the JAFFE (77.15% vs 75.61%) and NIMH-ChEFS (71.57% vs 70.53%) dataset, the performance is significantly higher on the CK+ (96.19% vs 94.06%) dataset.