Improvement of Neural-Network Classifiers Using Fuzzy Floating Centroids

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Traditional neural network classifiers like floating centroids neural-network (FCM) use hard decision boundaries.
• Hard boundaries lead to misclassification of noisy or boundary data points, hindering neural network training and optimization.
• Imbalanced datasets pose challenges for classifiers, often leading to a bias towards majority classes.

Purpose of the Study:

• To introduce a novel Fuzzy Floating Centroids Method (FFCM) for enhanced neural network classification.
• To address limitations of hard decision boundaries in FCM by incorporating a fuzzy strategy.
• To improve classification performance on imbalanced datasets and noisy data points.

Main Methods:

• The FFCM employs a fuzzy strategy and floating centroids to create soft decision boundaries.
• A weighted target function is integrated into FFCM to mitigate bias towards majority classes in imbalanced data.
• FFCM's performance is evaluated against ten other classification methods using average F-measure (Avg.FM) and generalization accuracy on 32 benchmark datasets.

Main Results:

• FFCM achieved optimal generalization accuracy on 17 datasets and optimal Avg.FM on 21 datasets.
• The method demonstrated superior performance compared to ten other classification algorithms across various benchmark datasets.
• FFCM effectively balanced precision and recall, outperforming competitors in estimating cement strength grade from microstructural images.

Conclusions:

• The proposed FFCM significantly enhances neural network classifier performance, particularly with noisy and imbalanced data.
• FFCM's soft boundaries and weighted target function contribute to improved optimization and classification accuracy.
• The method shows practical applicability in real-world scenarios, such as non-destructive estimation of material properties like cement strength.