Direct Optimization of Evaluation Measures in Learning to Rank Using Particle Swarm

One of the central issues in Learning to Rank (L2R) for Information Retrieval is to develop algorithms that construct ranking models by directly optimizing evaluation measures used in IR such as Precision at n, Mean Average Precision and Normalized Discounted Cumulative Gain. In this work we propose a new learning-to-rank method, referred as RankPSO. This algorithm is based on Particle Swarm Optimization. It builds a ranking model able to directly optimize evaluation measures used in Information Retrieval. To evaluate performance of RankPSO, we have compared it with other methods referenced in literature. We have carried out an experimental study using Letor OHSUMED dataset. The obtained results were analyzed statistically, demonstrating that RankPSO has significant improvement in precision compared to RankSVM, RankBoost and Regression methods; nevertheless, it does not have significant differences with AdaRank-MAP, AdaRank-NDCG, ListNet and FRank. The results show the advantages to use Particle Swarm Optimization as bio-inspired algorithm for learning to rank.