Index compression through document reordering

Abstract

An important concern in the design of search engines is the construction of an inverted index. An inverted index, also called a concordance, contains a list of documents (or posting list) for every possible search term. These posting lists are usually compressed with difference coding. Difference coding yields the best compression when the lists to be coded have high locality. Coding methods have been designed to specifically take advantage of locality in inverted indices. Here, we describe an algorithm to permute the document numbers so as to create locality in an inverted index. This is done by clustering the documents. Our algorithm, when applied to the TREC ad hoc database (disks 4 and 5), improves the performance of the best difference coding algorithm we found by fourteen percent. The improvement increases as the size of the index increases, so we expect that greater improvements would be possible on larger datasets.