AUFS-Samba: Adaptive Unsupervised Feature Selection for Mixed Data Type Using Simulated Annealing and Multi-Armed Bandit

Unsupervised feature selection (UFS) for mixed-type data is challenging because there are no labels, and traditional methods cannot handle heterogeneous feature spaces effectively. Many existing approaches often rely on data transformation, such as encoding or discretization, which may distort semantic meaning and cause evaluation bias, such as the Bias of Criterion Values to Dimension (BCVD). Additionally, most methods lack adaptive selection mechanisms and do not jointly consider clustering quality and feature redundancy, limiting their applicability to real-world heterogeneous data. In this paper, we propose Adaptive Unsupervised Feature Selection using Simulated Annealing and Multi-Armed Bandit (AUFS-Samba), which operates directly on mixed-type data without requiring transformation. AUFS-Samba combines adaptive exploration using a Multi-Armed Bandit (MAB) with global optimization via Simulated Annealing (SA). To the best of our knowledge, AUFS-Samba is among the first methods to combine MAB and SA for UFS on mixed-type data without transformation. Our method uses entropy-corrected k-Most Similar Neighbor Correlation (kMSNC) during MAB exploration to reduce redundancy. It also applies the Gower-based Silhouette Score during SA optimization to improve cluster coherence. Experiments on eight real-world UCI datasets show that AUFS-Samba consistently achieves higher clustering quality than representative state-of-the-art baselines, while selecting compact, interpretable, and relevant feature subsets. These findings demonstrate the effectiveness and robustness of AUFS-Samba and its practical potential for unsupervised learning on heterogeneous and unlabeled data.