Active Ranking from Pairwise Comparisons with Dynamically Arriving Items and Voters

We initiate the study of ranking from pairwise comparisons where the items to be ranked and (potentially malicious/random) voters who provide comparisons appear dynamically over time. We present DARPC-TOP, a general algorithmic framework for this problem. A detailed experimental study on simulated data-sets under the standard Bradley-Terry-Luce assumption for generating comparisons reveals that DARPC-TOP adapts very well to various distributions of voter and item arrivals. Furthermore, DARPC-TOP also is able to focus on ranking items well at the top of the list thus achieving the dual goal of ranking well at the top while adapting to item and voter arrivals.