High Utility Itemset Extraction using PSO with Online Control Parameter Calibration

Abstract

This study investigates the use of evolutionary computation for mining high-value patterns from benchmark datasets. The approach employs a fitness function to assess the usefulness of each pattern. However, the effectiveness of evolutionary algorithms heavily relies on the chosen strategy parameters during execution. Conventional methods set these parameters arbitrarily, often leading to suboptimal solutions. To address this limitation, the research proposes a method for dynamically adjusting strategy parameters using temporal difference approaches, a machine learning technique called Reinforcement Learning (RL). Specifically, the proposed IPSO RLON algorithm utilizes SARSA learning to intelligently adapt the Crossover Rate and Mutation Rate within the Practical Swarm Optimization Algorithm. This allows IPSO RLON to effectively mine high-utility itemsets from the data.The key benefit of IPSO RLON lies in its adaptive control parameters. This enables it to discover optimal high-utility itemsets when applied to various benchmark datasets. To assess its performance, IPSO RLON is compared to existing approaches like HUPEUMU-GRAM, HUIM-BPSO, IGA RLOFF, and IPSO RLOFF using metrics like execution time, convergence speed, and the percentage of high-utility itemsets mined. From the evaluation it is observed that the proposed IPSO RLON perfroms better than the other methodology.