Computational Intelligence Methods for Abalone Age Classification: From Traditional Neural Networks to Self-Constructing Neuro-Fuzzy Systems

This project explores the application of various computational intelligence methods in the task of abalone age classification, covering feedforward neural networks, gradient boosting trees, ensemble learning, and the SONFIN self-constructing neuro-fuzzy system. It compares the performance and interpretability of each model, focusing on analyzing the advantages of SONFIN's adaptive rule generation and the effects of ensemble learning, providing a reference for interpretable AI research.

Abalone age assessment is crucial for aquaculture management and market pricing. Traditional methods require cutting open the shell to observe growth rings, which are destructive and subjective. This project converts abalone age into an ordered classification problem with three categories: young, middle-aged, and old, which meets the needs of practical business decision-making.

Baseline Models: Feedforward Neural Networks (FNN)

• Baseline FNN: Establishes a performance baseline with basic architecture
• Advanced FNN: Optimized architecture (deep networks, regularization, etc.)

Gradient Boosting Tree Models

• XGBoost: Excellent training speed and prediction performance
• LightGBM: Lightweight with high memory efficiency
• CatBoost: Natively supports categorical features

Ensemble Learning: Stacking Strategy

Uses meta-learners to combine outputs from base models such as XGBoost and LightGBM, leveraging model complementarity to improve performance

Neuro-Fuzzy Systems

• Fixed structure: Predefined fuzzy rules and membership functions
• SONFIN: Dynamically generates rules, adaptive structure learning, Gaussian membership function learning, integrating neural self-learning and fuzzy interpretability

Performance Comparison

Evaluation Metric	Result
Best Kaggle Score	0.808
Best Single Model	XGBoost
Best Ensemble Model	Stacked Ensemble
Best Interpretable Model	SONFIN

Key Findings

1. XGBoost performs best among single models
2. Stacked ensemble achieves marginal performance improvement
3. SONFIN maintains interpretability while approaching the performance of top black-box models

Interpretability Comparison

• Neural networks/ensembles: Only feature importance analysis available
• Gradient boosting trees: Require tools like SHAP values for interpretation
• SONFIN: Generates directly understandable fuzzy rules (e.g., "Larger shell length and heavier weight → older age")

Development Tech Stack

Python, PyTorch, Scikit-learn, XGBoost/LightGBM/CatBoost, Streamlit, Pandas/NumPy, Matplotlib/Seaborn

Visualization Features

Learning curves, confusion matrices, prediction comparison charts, SONFIN membership function/rule visualization, feature importance, meta-learner weights

Streamlit Demo Application

Supports input of physical measurement values, real-time age prediction, and visualization of reasoning processes, lowering the barrier to use

Academic Value

Demonstrates core concepts of computational intelligence (neural computing, fuzzy logic, adaptive learning), suitable as a machine learning teaching case

Application Insights

1. Choose XGBoost for pure performance priority; choose SONFIN for interpretability needs
2. Stacked ensemble improves performance but requires trade-off with complexity
3. SONFIN's interpretability design is important for high-risk fields (finance, healthcare)

This project is a well-designed computational intelligence teaching and research project that implements multi-model comparative experiments, focusing on introducing the SONFIN neuro-fuzzy system to balance performance and interpretability. The code is clearly organized, the documentation is complete, covering the full process implementation, providing methodological references for related fields, and its design concept is worth learning from in practical projects.