원문정보
초록
영어
In recent years, multi-label learning has been used to deal with data attributed to multiple labels simultaneously and has been increasingly applied to various applications. As many other machine learning tasks, multi-label learning also suffers from the curse of dimensionality; so extracting good features using multiple labels of the datasets becomes an important step prior to classification. In this paper, we study the problem of multi-label feature selection for classification and have proposed a method based on single label feature selection ReliefF, termed ML-ReliefF, to select discriminant features in order to boost multi-label classification accuracy. Compared to other multi-label feature selection methods that only consider the relationship between pairwise classes, the proposed method introduces the concept of label set to further consider the relationship among more than two labels, modifies the regulation of the nearest neighbors computation reflecting the influence between samples and multiple labels, and considers and adds the similarity between samples to reinforce the effect. With the classifier, ML-kNN, experiments on five different datasets show that the proposed method is effective in removing irrelevant or redundant features and the selected features are more discriminant for classification.
목차
1. Introduction
2. Related Works
2.1. Single Label ReliefF Feature Selection
2.2. Multi-label Difficulties
3. Multi-label ReliefF Feature Selection
3.1. Aspects of ML-ReliefF
3.2. ML-ReliefF Algorithm
4. Experiment
4.1. Comparison of Feature Selection Methods
4.2. Consideration of Parameters
5. Conclusion and Future Work
Acknowledgments
References