| ABSTRACT |
The work integrates DQN-based feature grouping with multifactorial genetic optimization for high-dimensional
imbalanced feature selection. The framework first applies train-only preprocessing to reduce the original feature space
to a candidate set. A DQN agent then organizes the candidate features into learned groups using relevance, class
contribution, redundancy, similarity, and group-size information. These learned groups are subsequently converted into
task structures for a Multifactorial Evolutionary Algorithm with Genetic Algorithm operations (MFEA-GA).
The multifactorial stage searches several related feature-selection spaces simultaneously. The full candidate feature set
is retained as one task, while additional tasks are created from the validation-best DQN group and unions of highly
ranked DQN groups. A dynamic task-capacity mechanism is used so that total grouping capacity remains sufficient
when the number of candidate features changes. MFEA-GA then searches the task structures using a fitness function
that combines F1-score, recall, balanced accuracy, and feature reduction.
|
