Dynamic feature weighting for multi-label classification problems

Abstract

This paper proposes a dynamic feature weighting approach for multi-label classification problems. The choice of dynamic weights plays a vital role in such problems because the assigned weight to each feature might be dependent on the query. To take this dependency into account, we optimize our previously proposed dynamic weighting function through a non-convex formulation, resulting in several interesting properties. Moreover, by minimizing the proposed objective function, the samples with similar label sets get closer to each other while getting far away from the dissimilar ones. In order to learn the parameters of the weighting functions, we propose an iterative gradient descent algorithm that minimizes the traditional leave-one-out error rate. We further embed the learned weighting function into one of the popular multi-label classifiers, namely ML-kNN, and evaluate its performance over a set of benchmark datasets. Moreover, a distributed implementation of the proposed method on Spark is suggested to address the computational complexity on large-scale datasets. Finally, we compare the obtained results with several related state-of-the-art methods. The experimental results illustrate that the proposed method consistently achieves superior performances compared to others.