An Incremental Parallel Particle Swarm Approach for Classification Rule Discovery from Dynamic Data

Classification is a supervised learning technique that predicts the classes of unobserved data by employing a model built from available data. One of the efficient ways to represent this predictive model is to express it as an optimal set of classification rules to provide comprehensibility and precision, simultaneously. In this paper, we propose a novel incremental parallel Particle Swarm Optimization (PSO) approach for classification rule discovery. Our proposed method separates the training data into a set of data chunks regarding the classes and extracts optimal set of classification rules for each chunk in a parallel manner. In order to extract the rules from data chunks, we introduce an incremental PSO algorithm in which the previously extracted rules are directly employed to initialize the swarm population. Moreover, in each generation of the swarm, a tournament method is employed to substitute the weak individuals with strong extracted knowledge. To support the parallelism, we assign a PSO thread for each data chunk. As soon as all the PSO threads are completed, the extracted rules are integrated into a rule-base to construct a classification model. The evaluation results of the proposed approach on six datasets suggest that the classification precision of our proposed framework is competitive with offline learning methods and is 35% faster than its counterpart offline PSO approach.