Semi-supervised Feature Selection For Multi-view Data

Lots of features in high-dimensional data are redundant or irrelevant. They pose a challenge to learning tasks. To tackle this problem, the concept of feature selection has been introduced. Feature selection is one effective means to identify relevant features for dimension reduction. Feature selection is an important problem in machine learning and data mining field.In the meantime, many problems in machine learning involve data sets that are comprised of multiple views. Multi-view learning is the one of the hotspots in machine learning. In general, multiple views can be complementary and, when used together, can help improve the performance of learners. With the development of information technology, there are abundant unlabeled data while the number of labeled examples is limited, because labeling the examples requires human efforts. In this paper, we investigate how to exploit relations among views to help each other select relevant features with minimum redundancy, and propose a semi-supervised feature selection and clustering framework for multi-view data with a limited number of labels. To remove redundant and irrelevant features, the relations among views and the relations among features in each view are exploited, and a limited number of labeled data are used to help select features.The main contributions of this paper are summarized below:(1) Defining the novel problem of semi-supervised feature selection for multi-view data with a limited number of labels;(2) Proposing to exploit the relations among views and the relations among features in each view in formulating multi-view feature selection;(3) Evaluating the proposed framework on5multi-view datasets and the results demonstrate the SSMVFS can get better performance.