An Efficient Technique for Feature Selection in Learning to Rank Using NSGA-II

Feature Selection (FS) is essential for optimizing Learning to Rank (LTR) models by determining the vital choice of attributes from a complex dataset. This study examines feature selection strategies within the framework of learning to rank, specifically emphasizing a multi-objective evolutionary algorithm utilizing the listwise approach. Conventional ranking algorithms frequently experience issues with superfluous or unnecessary features, undermining their efficacy. The suggested method employs a combination of instance selection, evolutionary optimization, and ensemble techniques to choose feature subsets. The methodology seeks to reduce the feature count while enhancing ranking precision, assessed via the Normalized Discounted Cumulative Gain (NDCG) metric. Experiments were performed using LETOR 4.0 dataset (MQ2007-list and MQ2008-list), evaluating proposed method against established listwise ranking algorithms like ListMLE, ListNet, RankCosine and p-ListMLE. The findings indicate that, although there are performance discrepancies relative to current algorithms, the suggested method attains substantial feature reduction up to 60 percent while maintaining competitive ranking accuracy.