Learning to enumerate

Abstract

The Learning to Enumerate problem is a new variant of the typical active learning problem. Our objective is to find data that satisfies arbitrary but fixed conditions, without using any prelabeled training data. The key aspect here is to query as few as possible non-target data. While typical active learning techniques try to keep the number of queried labels low they give no regards to the class these instances belong to. Since the aim of this problem is different from the common active learning problem, we started with applying uncertainty sampling as a base technique and evaluated the performance of three different base learner on 19 public datasets from the UCI Machine Learning Repository.