| dc.description.abstract | This study aims to apply autoencoder feature selection to supervised tasks, investigate its prediction performance and stability compared to relevance feature selection, and further ana-lyze the impact of homogeneous ensemble and the proposed two-phase combination on feature selection effectiveness to establish a better feature selection method.
We constructed an autoencoder feature selection method based on the Gedeon method and compared it with four relevance feature selection methods: Impurity, Anova, ReliefF, and Mutu-al Information. The experimental results showed that the autoencoder feature selection per-formed poorly without architectural improvements.
In the homogeneous ensemble experiment, relevance feature selection achieved better overall evaluation by sacrificing a small amount of prediction performance in exchange for im-proved stability. The autoencoder feature selection improved stability and prediction perfor-mance, outperforming relevance feature selection in prediction performance. In the two-phase combination, using autoencoder feature selection as the first-phase is the optimal combination order. Combining two different evaluation feature selections in this order, it outperforms all non-ensemble and homogeneous ensemble feature selection methods in prediction performance.
Based on the experimental results, this study suggests that feature selection should be cho-sen based on different application scenarios, either using a homogeneous ensemble or the two-phase combination, to enhance the effectiveness of feature selection. The homogeneous ensemble focuses on improving stability. In contrast, the two-phase combination effectively im-proves prediction performance and maintains good stability by applying a homogeneous en-semble to the feature selection in both phases. | en_US |