| ABSTRACT |
The work presents a Deep Q-Network-based feature grouping framework (DQN-FG) for organizing input features into compact groups before classification. Unlike conventional feature selection methods that immediately retain or discard individual features, DQN-FG treats grouping as a sequential decision process. Each feature is examined by the DQN agent and assigned to one of several feature groups according to the current state of the grouping environment.
The state representation incorporates information on feature relevance, class contribution, feature usage, redundancy with existing group members, similarity to a group, and the relative size of the group. The reward encourages the agent to place informative and class-sensitive features into coherent groups while penalizing redundancy and excessive group growth. After training, the learned policy generates the final groups. Each non-empty group is evaluated using validation F1-score and the best-performing group is retained as the compact feature subset for final classification.
|