Risk-Sensitive Learning to Rank with Evolutionary Multi-Objective Feature Selection

摘要: Learning to Rank (L2R) is one of the main research lines in Information Retrieval. Risk-sensitive L2R a sub-area that tries learn models are good on average while at same time reducing risk performing poorly few but important queries (e.g., medical or legal queries). One way learned by selecting and removing noisy, redundant features, features promote some detriment others. This exacerbated learning methods usually maximize an metric mean precision (MAP) Normalized Discounted Cumulative Gain (NDCG)). However, historically, feature selection (FS) have focused only effectiveness reduction as objectives. Accordingly, this work, we propose evaluate FS for with additional objective mind, namely risk-sensitiveness. We present novel single multi-objective criteria optimize reduction, effectiveness, risk-sensitiveness, all time. also introduce new methodology explore search space, suggesting effective efficient extensions well-known Evolutionary Algorithm (SPEA2) applied L2R. Our experiments show explicitly including criterion crucial achieving more risk-sensitive performance. provide thorough analysis our experimental results.