Active image clustering: Seeking constraints from humans to complement algorithms

Abstract

We propose a method of clustering images that combines algorithmic and human input. An algorithm provides us with pairwise image similarities. We then actively obtain selected, more accurate pairwise similarities from humans. A novel method is developed to choose the most useful pairs to show a person, obtaining constraints that improve clustering. In a clustering assignment elements in each data pair are either in the same cluster or in different clusters. We simulate inverting these pairwise relations and see how that affects the overall clustering. We choose a pair that maximizes the expected change in the clustering. The proposed algorithm has high time complexity, so we also propose a version of this algorithm that is much faster and exactly replicates our original algorithm. We further improve run-time by adding heuristics, and show that these do not significantly impact the effectiveness of our method. We have run experiments in two different domains, namely leaf images and face images, and show that clustering performance can be improved significantly.

Interactively Guiding Semi-Supervised Clustering via Attribute-Based Explanations

“…This demonstrates that semi-supervised clustering algorithms do not gain much until a large number of pairs have been labelled [4,11]. The max-entropy baseline starts learning earliest followed by semi-random and pure-random.…”

“…But these approaches are specific to the clustering algorithms they use. Biswas et al [11] present an active clustering algorithm that goes through all possible pairs and then re-clusters the data in order to identify the most informative pair. Hence, they use a simple Minimal Spanning Tree based clustering algorithm that can be run many times in a reasonable amount of time.…”

“…Active clustering works like [11,13] reduce the number of constraints by querying a user on actively chosen pairs instead of completely random pairs. But these approaches are specific to the clustering algorithms they use.…”

“…But each constraint still remains a weak low-level indication of the desired clustering. For instance, active PCK-Means [12] requires 5000 constraints to achieve 30% accuracy on a dataset of 500 face images as reported in [11].…”

“…This is because the pairs of images to be annotated are typically randomly chosen, and are likely to be redundant. Active clustering approaches [11,12,13] reduce the number of required constraints by iteratively and actively selecting a pair to be annotated by a user in the loop. But each constraint still remains a weak low-level indication of the desired clustering.…”