Differential Evolution-Based Fusion and Its Properties for Web Search

In recent years, data fusion has been applied to many different application areas such as neural networks, classification, multi-sensor systems, image processing, information retrieval, Web search among others. Linear combination is a popular data fusion method due to its flexibility. Proper weight assignment is a key issue for its success. In this paper, we apply the differential evolution optimization method to find suitable weights in the search space. Experiments are carried out with authoritative TREC data and we find it is a good method for the task and can improve fusion performance significantly than the best component results and other heuristic data fusion methods. Moreover, we have two findings. One finding is compared with other fusion methods, differential evolution based method performs better when more component search engines are involved in the fusion process. The second is a relatively large number of queries (e.g. over 100 queries) should be used as training data in order to obtain reliable weights.