A New Many-Objective Optimization Approach to Association Rule Mining: The NSGA-II/DE-ARM Algorithm.

Area of Science:

• Data Mining
• Optimization
• Machine Learning

Background:

• Traditional association rule mining (ARM) uses single-metric filtering, which is insufficient for capturing multi-criteria rule quality.
• There is a need for advanced methods to uncover complex relationships in large datasets considering multiple quality aspects.

Purpose of the Study:

• To formulate ARM as a many-objective optimization problem.
• To propose a hybrid algorithm, NSGA-II/DE-ARM, for simultaneous optimization of four key rule-quality measures: support, confidence, lift, and NetConf.

Main Methods:

• Developed a hybrid algorithm (NSGA-II/DE-ARM) integrating NSGA-II with binary differential evolution operators.
• Incorporated adaptive operator selection, lift-weighted tournament selection, and a dynamic minimum support threshold.
• Evaluated performance on a SIPRI-World Bank dataset and the UCI Mushroom benchmark dataset.

Main Results:

• NSGA-II/DE-ARM significantly outperformed the Apriori baseline on both datasets across all four metrics (support, confidence, lift, NetConf).
• Achieved substantial improvements with large effect sizes (Cohen's d = 0.93-6.16) and high hypervolume values (HV = 3.231 for SIPRI-World Bank, HV = 6.262 for Mushroom).
• Generated a diverse set of 68 Pareto-optimal rules in a representative run, offering a balanced multi-criteria solution set.

Conclusions:

• NSGA-II/DE-ARM effectively addresses the limitations of single-metric filtering in ARM.
• The proposed algorithm provides decision-makers with a broader and more balanced set of high-quality association rules.
• This multi-objective approach enhances the discovery of latent relationships in complex datasets.