Efficient Index for Retrieving Top-k Most Frequent Documents

Abstract

In the document retrieval problem [9], we are given a collection of documents (strings) of total length D in advance, and our target is to create an index for these documents such that for any subsequent input pattern P, we can identify which documents in the collection contain P. In this paper, we study a natural extension to the above document retrieval problem. We call this top-k frequent document retrieval, where instead of listing all documents containing P, our focus is to identify the top k documents having most occurrences of P. This problem forms a basis for search engine tasks of retrieving documents ranked with TFIDF metric.

A related problem was studied by [9] where the emphasis was on retrieving all the documents whose number of occurrences of the pattern P exceeds some frequency threshold f. However, from the information retrieval point of view, it is hard for a user to specify such a threshold value f and have a sense of how many documents will be outputted. We develop some additional building blocks which help the user overcome this limitation. These are used to derive an efficient index for top-k frequent document retrieval problem, answering queries in O(P + logD loglogD + k) time and taking O(DlogD) space. Our approach is based on novel use of the suffix tree called induced generalized suffix tree (IGST).

This work is supported in part by Taiwan NSC Grant 96-2221-E-007-082-MY3 (W. Hon) and US NSF Grant CCF–0621457 (R. Shah).