Semi-supervised fuzzy-rough feature selection

With the continued and relentless growth in dataset sizes in recent times, feature or attribute selection has become a necessary step in tackling the resultant intractability. Indeed, as the number of dimensions increases, the number of corresponding data instances required in order to generate accurate models increases exponentially. Fuzzy-rough set-based feature selection techniques offer great flexibility when dealing with real-valued and noisy data; however, most of the current approaches focus on the supervised domain where the data object labels are known. Very little work has been carried out using fuzzy-rough sets in the areas of unsupervised or semi-supervised learning. This paper proposes a novel approach for semi-supervised fuzzy-rough feature selection where the object labels in the data may only be partially present. The approach also has the appealing property that any generated subsets are also valid (super)reducts when the whole dataset is labelled. The experimental evaluation demonstrates that the proposed approach can generate stable and valid subsets even when up to 90% of the data object labels are missing.With the continued and relentless growth in dataset sizes in recent times, feature or attribute selection has become a necessary step in tackling the resultant intractability. Indeed, as the number of dimensions increases, the number of corresponding data instances required in order to generate accurate models increases exponentially. Fuzzy-rough set-based feature selection techniques offer great flexibility when dealing with real-valued and noisy data; however, most of the current approaches focus on the supervised domain where the data object labels are known. Very little work has been carried out using fuzzy-rough sets in the areas of unsupervised or semi-supervised learning. This paper proposes a novel approach for semi-supervised fuzzy-rough feature selection where the object labels in the data may only be partially present. The approach also has the appealing property that any generated subsets are also valid (super)reducts when the whole dataset is labelled. The experimental evaluation demonstrates that the proposed approach can generate stable and valid subsets even when up to 90% of the data object labels are missing.P