Abstract
This paper describes an application of support vector machines (SVMs) to interactive document retrieval using active learning. We show that an SVM-based retrieval has an association with conventional Rocchio-based relevance feedback by a comparative analysis. We propose a cosine kernel, which denotes cosine similarity, suitable for an SVM-based interactive document retrieval based on the analysis. We confirm the effectiveness of our approach when we adopt Boolean, TF, and TFIDF to represent the document vectors. Our proposed approach, in particular, TF representation, shows better performance, which is demonstrated experimentally.
