Multi-Objective Clustering with Data - and Human-Driven Metrics

Abstract

While the usefulness of clustering analysis for knowledge discovery and information management is well accepted, it is difficult to objectively estimate the quality of clusters due to the fact that resulting clusters are heavily dependent on which and how many variables are used for clustering. Further, there are a number of heuristic criteria that can be used to estimate the quality of clusters and none of them completely dominates the others. To overcome these shortcomings, this study takes a multi-objective genetic algorithm (MOGA) approach to explore possible combinations of input variables to form clusters and evaluate cluster quality in a multi-objective space. In particular, each candidate cluster in MOGA is evolved and evaluated by both data-driven and human-driven metrics developed for this study. Note that human-driven metrics are introduced to measure managerial quality of clusters for efficient information management and effective decision making. The proposed system also allows the decision maker to navigate non-dominated cluster solutions and to choose one of them as the final solution. Experimental results on both synthetic and real data show promise in finding non-dominated cluster solutions with significant managerial insights for decision making.

• Multi-objective clustering
• cluster interestingness
• cluster relevance
• visualization
• Pareto optimization