Heuristic ternary error-correcting output codes via weight optimization and layered clustering-based approach.

Area of Science:

• Machine Learning
• Computer Science
• Data Mining

Background:

• Error-Correcting Output Codes (ECOCs) are effective ensemble methods for multiclass classification.
• ECOCs decompose complex multiclass problems into simpler binary classification tasks.
• Existing ECOC methods may face challenges with highly confusing class pairs.

Purpose of the Study:

• To introduce a novel heuristic ternary code, Weight Optimization and Layered Clustering-based ECOC (WOLC-ECOC).
• To enhance the performance and robustness of ECOC classifiers for multiclass problems.
• To address limitations in training heuristic ECOC methods on difficult binary-class problems.

Main Methods:

• WOLC-ECOC employs a layered clustering-based ECOC (LC-ECOC) approach to build strong classifiers for confusing binary problems.
• A novel optimized weighted (OW) decoding algorithm is utilized to integrate new classifiers into the ECOC framework.
• The OW decoding optimization problem is efficiently solved using the cutting plane algorithm.
• The iterative process continues until training risk converges, ensuring non-increase in risk and a small code length.

Main Results:

• WOLC-ECOC demonstrated effectiveness across 14 UCI datasets.
• The method proved successful in a real-world music genre classification task.
• The approach mitigates issues where heuristic training might stall on challenging binary-class problems.

Conclusions:

• WOLC-ECOC offers a robust and effective enhancement to ECOC methodology for multiclass classification.
• The combination of LC-ECOC and OW decoding provides a powerful strategy for improving classifier accuracy.
• The proposed method shows significant potential for applications requiring high-performance multiclass classification.