Multi-label Classification with Feature-aware Implicit Encoding and Generalized Cross-entropy Loss

In multi-label classification problems, each instance can simultaneously have multiple labels. Since the whole number of available labels in real-world applications tends to be (very) large, multi-label classification becomes an important challenge and recently label space dimension reduction (LSDR) methods have received attention. These methods first encode the output space to a low-dimensional latent space. Afterwards, they predict the latent space from the feature space and reconstruct the original output space using a suitable decoding method. The encoding method can be implicit which learns a code matrix in the latent space by solving an optimization problem or explicit which learns a direct encoding function. It can be feature-aware which considers predictability of the latent space from the feature space or not feature-aware which obtains the latent space from only the label space. In this paper, we propose FIECE as a feature-aware implicit encoding method that uses a generalized cross-entropy loss function to compute reconstruction error of the label space. Since label vectors in these problems are usually sparse, we use a parameter in the cost function to address the imbalanced classification problem for each label. Extensive experiments on several datasets demonstrate effectiveness of the proposed method compared to some recent multi-label classification methods.