An Improved Retrievability-Based Cluster-Resampling Approach for Pseudo Relevance Feedback

Cluster-based pseudo-relevance feedback (PRF) is an effective approach for searching relevant documents for relevance feedback. Standard approach constructs clusters for PRF only on the basis of high similarity between retrieved documents. The standard approach works quite well if the retrieval bias of the retrieval model does not create any effect on the retrievability of documents. In our experiments we observed when a collection contains retrieval bias, then high retrievable documents of clusters are frequently retrieved at top positions for most of the queries, and these drift the relevance feedback away from relevant documents. For reducing (retrieval bias) noise, we enhance the standard cluster construction approach by constructing clusters on the basis of high similarity and retrievability. We call this retrievability and cluster-based PRF. This enhanced approach keeps only those documents in the clusters that are not frequently retrieve due to retrieval bias. Although this approach improves the effectiveness, however, it penalizes high retrievable documents even if these documents are most relevant to the clusters. To handle this problem, in a second approach, we extend the basic retrievability concept by mining frequent neighbors of the clusters. The frequent neighbors approach keeps only those documents in the clusters that are frequently retrieved with other neighbors of clusters and infrequently retrieved with those documents that are not part of the clusters. Experimental results show that two proposed extensions are helpful for identifying relevant documents for relevance feedback and increasing the effectiveness of queries.