Incremental Feature Selection in Dynamic Incomplete Ordered Decision Systems

TMLR Paper6604 Authors

Abstract: Incremental feature selection aims to efficiently identify key features from dynamic data. However, existing feature selection algorithms for dynamic incomplete ordered data often rely on upper and lower approximations while overlooking the impact of inter-feature relationships across different decision classes. This can lead to reduced computational efficiency and suboptimal classification accuracy. To address these issues, this paper proposes an incremental feature selection method based on expanded dominance matrices for incomplete ordered decision systems. Firstly, we propose to use non-dominant relationships between classes as a measure of attribute importance, thereby avoiding the computational complexity of traditional lower and upper approximation. Furthermore, to maintain efficiency and accuracy in dynamic data environments which involve frequent object addition and deletion, we propose two matrix-based incremental update mechanisms: matrix-based non-dominance attribute reduction for addition (MNAR-A) and matrix-based non-dominance attribute reduction for deletion (MNAR-D). These mechanisms are crucial for efficiently updating the feature subset when new objects are added or existing objects are removed, ensuring the algorithm remains effective and avoids recomputing from scratch. Experimental results on the UCI dataset showed that the proposed algorithm achieved a 1.3\(\times\) speedup and delivered a 7\% relative accuracy gain compared to the state-of-the-art method on average.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Marcello_Restelli1

Submission Number: 6604