Non-relevance Feedback Document Retrieval using Large Data Set

In interactive document retrieval, we need to find relevant documents to our interest from a large data set of documents, within a few iterations of judgement on retrieved documents. In each iteration, a comparatively small batch of documents is evaluated to establish their relevance to user’s interest. This method is also called relevance feedback, and it requires both of relevant and non-relevant documents. However, the documents initially presented for user’s judgement do not always include relevant documents. Thus we have proposed a feedback method using information on non-relevant documents only and named this method “non-relevance feedback”. Nonrelevance feedback selects a set of documents which are discriminated not non-relevant region and are near the discriminant hyperplane based on learning result by One-class Support Vector Machine (One-class SVM). We conducted experiments using large data sets including over 500,000 newspaper articles and confirmed that the proposed method outperformed other methods.