Evaluation of Jensen-Shannon Distance over Sparse Data

摘要

Jensen-Shannon divergence is a symmetrised, smoothed version of Kuellback-Leibler. It has been shown to be the square of a proper distance metric, and has other properties which make it an excellent choice for many high-dimensional spaces in R*. The metric as defined is however expensive to evaluate. In sparse spaces over many dimensions the Intrinsic Dimensionality of the metric space is typically very high, making similarity-based indexing ineffectual. Exhaustive searching over large data collections may be infeasible. Using a property that allows the distance to be evaluated from only those dimensions which are non-zero in both arguments, and through the identification of a threshold function, we show that the cost of the function can be dramatically reduced.