An information measure for comparing top $k$ lists (1310.0110v1)

Abstract: Comparing the top $k$ elements between two or more ranked results is a common task in many contexts and settings. A few measures have been proposed to compare top $k$ lists with attractive mathematical properties, but they face a number of pitfalls and shortcomings in practice. This work introduces a new measure to compare any two top k lists based on measuring the information these lists convey. Our method investigates the compressibility of the lists, and the length of the message to losslessly encode them gives a natural and robust measure of their variability. This information-theoretic measure objectively reconciles all the main considerations that arise when measuring (dis-)similarity between lists: the extent of their non-overlapping elements in each of the lists; the amount of disarray among overlapping elements between the lists; the measurement of displacement of actual ranks of their overlapping elements.