Extreme Multi-Label Classification of Images via Multiscale k-Nearest Neighbour

Abstract

We consider the extreme multi-label classification (XMC) problem, which aims at finding positive labels of a query from an extreme variety of labels, e.g., diverse text-tags of images posted on social networking services. k-nearest neighbour (k-NN) can be applied to XMC problem: k-NN predicts the positive label probabilities by averaging the labels of k objects nearest to the query. However, the predicted probability with small k can be unintentionally stick to 0 in many cases, as many labels are often sparse in XMC setting. Conversely, k-NN estimator with large k has large bias, as it leverages the labels of objects distant from the query. For solving these issues, we employ multiscale k-NN, which reduces the bias of the k-NN asymptotically. Through NUS-WIDE dataset experiments, we examine the multiscale k-NN and its modification using a sigmoid function, as a first work of the practical XMC application of the MS-k-NN.