Multiple reference points MOEA/D for feature selection

Feature selection can be considered a multi-objective problem since its two main objectives usually conflict with each other. Many Pareto dominance-based algorithms have been applied to feature selection. However, feature subsets evolved by these algorithms are mostly around the center of the Pareto front. MOEA/D can avoid this issue to some extent, but still needs to be modified when applying it to solve complex feature selection problems. This paper proposes a new decomposition strategy for feature selection called MOEA/D-MRPs which uses multiple reference points instead of multiple weight vectors. The proposed algorithm, is evaluated on eight different datasets and compared with three Pareto dominance-based algorithms and the standard MOEA/D algorithm. Experimental results show that MOEA/D-MRPs can efficiently evolve a more diverse set of non-dominated solutions than three Pareto dominance-based algorithms and achieve better classification performance than the standard MOEA/D algorithm. On large datasets, MOEA/D-MRPs is also the most efficient algorithm.