Structured feature selection using coordinate descent optimization.

Area of Science:

• Machine Learning
• Bioinformatics
• Computational Biology

Background:

• Traditional feature selection methods often overlook feature interdependencies and prior structural knowledge.
• This research addresses the challenge of selecting one representative feature from predefined feature groups.
• The goal is to ensure selected features collectively discriminate between classes effectively.

Purpose of the Study:

• To develop a feature selection method that incorporates prior knowledge by grouping features.
• To formulate the feature selection problem as a constrained optimization problem.
• To propose an efficient optimization algorithm for high-dimensional feature selection.

Main Methods:

• The problem is framed as a binary constrained optimization, relaxed into a convex-concave problem.
• This is further transformed into a sequence of convex optimization problems solvable by standard algorithms.
• A specialized block coordinate gradient descent algorithm is developed for efficiency.

Main Results:

• The method was validated using microarray analysis, grouping genes by expression similarity or function.
• Evaluated on five benchmark gene expression datasets, the proposed method demonstrated superior accuracy compared to state-of-the-art techniques.
• The block coordinate gradient descent approach achieved higher average AUC in 13 out of 25 experiments.

Conclusions:

• A novel method for selecting one feature per group is presented, outperforming existing gene selection algorithms.
• The algorithm effectively utilizes gene expression similarity for grouping and selection, yielding more accurate and less redundant feature sets.
• Unlike methods limited to similarity-based grouping, this approach can leverage any feature grouping structure.