Active Learning with Positive and Negative Pairwise Feedback

In this paper, we propose a generic framework for active clustering with queries for pairwise similarities between objects. First, the pairwise similarities can be any positive or negative number, yielding full flexibility in the type of feedback that a user/annotator can provide. Second, the process of querying pairwise similarities is separated from the clustering algorithm, leading to more flexibility in how the query strategies can be constructed. Third, the queries are robust to noise by allowing multiple queries for the same pairwise similarity (i.e., a non-persistent noise model is assumed). Finally, the number of clusters is automatically identified based on the currently known pairwise similarities. In addition, we propose and analyze a number of novel query strategies suited to this active clustering framework. We demonstrate the effectiveness of our framework and the proposed query strategies via several experimental studies.