Feature selection inspired classifier ensemble reduction.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Classifier ensembles are a key area in machine learning, enhancing predictive performance over single models.
• Existing ensemble methods may include redundant classifiers, potentially hindering optimal performance and efficiency.
• Reducing ensemble size offers benefits in memory, storage, and runtime overhead.

Purpose of the Study:

• To develop a novel technique for reducing classifier ensembles.
• To enhance ensemble diversity and predictive accuracy by removing redundant classifiers.
• To improve the overall efficiency of classifier systems.

Main Methods:

• Extending feature selection concepts to classifier ensemble reduction.
• Transforming ensemble predictions into training samples, with classifiers treated as features.
• Employing global heuristic harmony search for optimal feature subset selection.

Main Results:

• The proposed technique systematically reduces classifier ensembles.
• Evaluated on high-dimensional and large datasets, the reduced ensembles showed superior performance.
• Demonstrated improved classification accuracy compared to original and randomly formed ensembles.

Conclusions:

• Classifier ensemble reduction is an effective strategy for improving performance and efficiency.
• The harmony search-based method successfully identifies and removes redundant classifiers.
• This approach offers a practical solution for developing more efficient and accurate machine learning models.