Improved Golf Optimization for Feature Selection in Palm Vein and Fingerprint Biometrics (MATLAB)

Hello,

I am currently working on my MPhil research titled:

“Improved Golf Optimization Algorithm for Feature Selection in Multimodal Biometrics.”

My study focuses on feature selection for a multimodal biometric system combining palm vein and fingerprint data, implemented in MATLAB using an Improved Golf Optimization Algorithm (GOA).

I would appreciate guidance on the following:

1. Fitness Function Design for Multimodal Biometrics

• In MATLAB, what is the best way to design a fitness function that:
• maximizes recognition accuracy for palm vein + fingerprint fusion, and
• minimizes the number of selected features?
• Is a weighted-sum fitness function appropriate, or should a multi-objective approach be used?

1. Binary Feature Selection Encoding

• For GOA-based feature selection, is it better to:
• apply a sigmoid (or V-shaped) transfer function to convert continuous positions to binary feature masks, or
• redesign GOA directly as a binary optimizer?
• Any MATLAB examples or recommended practices?

1. Feature-Level Fusion Strategy

• When combining palm vein and fingerprint features:
• should feature selection be performed before fusion (independently per modality), or
• after feature-level fusion?
• What approach is more efficient and commonly used in MATLAB implementations?

1. Classifier Integration

• What is the most efficient way to integrate classifiers (e.g., SVM, KNN) inside the optimization loop in MATLAB?
• How can computation time be reduced during repeated fitness evaluations?

1. Evaluation Metrics

• Besides accuracy, which metrics are most suitable for palm vein and fingerprint systems (e.g., FAR, FRR, EER)?
• Are there MATLAB functions or toolboxes recommended for biometric performance evaluation?

If anyone has experience with metaheuristic feature selection, Golf Optimization, or palm vein and fingerprint recognition in MATLAB, your advice would be highly appreciated.

Thank you for your time and support.

Best regards,

Answer 1

[Ritam]

Based on the assumption that the problem statement is similar to any other optimisation problem,

1. Fitness Function Design

Weighted-sum: Simple to implement and computationally efficient, but sensitive to weight tuning and may hide good trade-offs.

Multi-objective: Provides a Pareto front for balanced solutions and better interpretability, but adds complexity and longer runtime.

In my opinion, I would say multi-objective is generally more informative for research, though weighted-sum is practical for quick implementation.

2. Binary Feature Selection Encoding

Transfer function (sigmoid/V-shaped): Easy to integrate with continuous GOA, widely used, but may lead to instability if thresholds are poorly chosen.

Binary redesign: Offers algorithmic consistency and potentially better convergence, but requires significant modification and validation.

Generally, transfer functions are commonly preferred for simplicity; binary redesign suits specialized studies.

3. Feature-Level Fusion Strategy

Before fusion (per modality): Reduces dimensionality early, avoids incompatibility, but may miss cross-modal correlations.

After fusion: Captures joint relationships, but risks high dimensionality and overfitting without strong normalization and reduction.

Although, independent selection before fusion is often more robust while after fusion can be explored for richer feature interactions, the right approach in my opinion would depend more on the problem statement that you are working on.

Classifier Integration

Direct integration: Ensures accurate fitness evaluation, but increases computation time significantly.

Approximation or surrogate models: Speeds up optimization, but may reduce accuracy and reliability.

In other words, direct integration with parallelization and caching is common, while surrogates are useful for very large datasets.

Evaluation Metrics

Accuracy only: Simple and intuitive, but ignores biometric-specific performance aspects.

FAR/FRR/EER and ROC/DET: Provide comprehensive verification insights, but require additional computation and interpretation.

Here, evaluation metrics depend more on the context of the problem statement and may call for more custom evaluation metrics (like combining both for better results).

I hope this gives you a better position to make informed decision.