Dual dimensionality reduction on instance-level and feature-level for multi-label data

Abstract

The training data in multi-label learning are often high dimensional and contains a quantity of noise and redundant information, resulting in high memory overhead and low classification performance during the learning process. Therefore, dimensionality reduction for multi-label data has become an important research topic. Existing dimensionality reduction methods for multi-label data focus on either the instance-level or the feature-level; few studies have achieved both. This paper proposes a novel two-stage method to reduce dimensionality for both instances and features on multi-label data. In the dimensionality reduction stage of instances, the original training data are converted into single-label data utilizing binary relevance. The learning vector quantization technique is employed to perform prototype selection on the transformed data and generate new instance-level low-dimensional multi-label data on the ground of the nearest neighbor information of the selected prototypes. Next, a filter-based feature selection method is proposed to choose discriminative features for each class label in the feature reduction phase. The number of retained features is determined according to the preset proportion parameters to achieve the feature-level dimensionality reduction. Experimental results on seven benchmarks verify the effectiveness of the proposed method.