Inverted file search algorithms for collaborative filtering

Abstract

This paper explores the possibility of using a disk based inverted file structure for collaborative filtering. Our hypothesis is that this allows for faster calculation of predictions and also that early termination heuristics may be used to further speed up the filtering process and perhaps even improve the quality of the predictions. In an experiment on the EachMovie dataset this was tested. Our results indicate that searching the inverted file structure is many times faster than general in-memory vector search, even for very large profiles. The Continue termination heuristics produces the best ranked predictions in our experiments, and Quit is the top performer in terms of speed.