A Score Fusion Method Using a Mixture Copula

In this paper, we propose a score fusion method using a mixture copula that can consider complex dependencies between multiple relevance scores in order to improve the effectiveness of information retrieval. The combination of multiple relevance scores has been shown to be effective in comparison with a single score. Widely used score fusion methods are linear combination and learning to rank. Linear combination cannot capture the non-linear dependency of multiple scores. Learning to rank yields output that makes it difficult to understand the models. These problems can be solved by using a copula, which is a statistical framework, because it can capture the non-linear dependency and also provide an interpretable reason for the model. Although some studies apply copulas to score fusion and demonstrate the effectiveness, their methods employ a unimodal copula, thus making it difficult to capture complex dependencies. Therefore, we introduce a new score fusion method that uses a mixture copula to handle the complicated dependencies of scores; then, we evaluate the accuracy of our proposed method. Experiments on ClueWeb’09, a large-scale document set, show that in some cases, our proposed method significantly outperforms linear combination and others existing methods that use a unimodal copula.