DeepRank: A Novel Rank Aggregation Method Using Deep-Learning Ensembles

Rank-aggregation takes multiple rankings of the same items and combines them into a unified ranking. Despite the existence of different rank-aggregation algorithms, they might struggle with capturing the intricate connections between different rankings. Deep-learning algorithms, with their ability to learn complex patterns, can effectively identify these relationships and leverage them for better aggregation. This paper introduces DeepRank, a novel rank-aggregation method that uses deep-learning ensembles to improve the ranking performance. It employs multiple deep-learning models as base rankers, which generate individual rankings. DeepRank captures relationships between the base rankers. This algorithm is evaluated on the MQ2007-agg and the MQ2008-agg as the rank-aggregation part of the LETOR4.0 benchmark dataset, a widely used dataset in the learning-to-rank area. DeepRank achieves improvements of 7.04% in P@1 and 4.3% in NDCG@1 compared to baseline algorithms on the MQ-2008 dataset. This gain demonstrates the potential of deep-learning ensembles for superior ranking performance.