Effective Bitmap Indexing for Non-metric Similarities

An increasing number of applications include recommender systems that have to perform search in a non-metric similarity space, thus creating an increasing demand for efficient yet flexible indexing techniques to facilitate similarity search. This demand is further fueled by the growing volume of data available to recommender systems. This paper addresses the demand in the specific domain of music recommendation. The paper presents the Music On Demand framework where music retrieval is performed in a continuous, stream-based fashion. Similarity measures between songs, which are computed on high-dimensional feature spaces, often do not obey the triangular inequality, meaning that existing indexing techniques for high-dimensional data are infeasible. The most prominent contribution of the paper is the proposal of an indexing approach that is effective for non-metric similarities. This is achieved by using a number of bitmap indexes combined with effective bitmap compression techniques. Experiments show that the approach scales well.