Dry Bean Classification Project

Overview

This project implements a multiclass classification system to predict dry bean varieties using machine learning techniques. The system classifies beans into 7 different classes: Seker, Barbunya, Bombay, Cali, Dermason, Horoz, and Sira based on their physical characteristics.

Dataset Information

Source

• Citation: KOKLU, M. and OZKAN, I.A., (2020), Multiclass Classification of Dry Beans Using Computer Vision and Machine Learning Techniques. Computers and Electronics in Agriculture, 174, 105507.
• DOI: https://doi.org/10.1016/j.compag.2020.105507
• Dataset: Murat Koklu's Dataset Page

Features (16 attributes)

• area: The area of a bean zone and the number of pixels within its boundaries
• perimeter: Bean circumference defined as the length of its border
• majorAxisLength: Distance between the ends of the longest line drawable from a bean
• minorAxisLength: Longest line drawable perpendicular to the main axis
• aspectRatio: Relationship between major and minor axis lengths
• eccentricity: Eccentricity of the ellipse having the same moments as the region
• convexArea: Number of pixels in the smallest convex polygon containing the bean area
• equivDiameter: Diameter of a circle having the same area as the bean
• extent: Ratio of pixels in bounding box to bean area
• solidity: Ratio of pixels in convex shell to those in beans
• roundness: Calculated as (4 * π * Area) / (Perimeter²)
• compactness: EquivalentDiameter / MajorAxisLength
• shapeFactor1: Area / Perimeter²
• shapeFactor2: Area / MajorAxisLength²
• shapeFactor3: Area / (MajorAxisLength × MinorAxisLength)
• shapeFactor4: Perimeter / (MajorAxisLength × MinorAxisLength)

Target Classes (7 varieties)

Seker, Barbunya, Bombay, Cali, Dermason, Horoz, Sira

Installation & Setup

# Required libraries
pip install pandas numpy scikit-learn matplotlib seaborn openpyxl

Data Preprocessing

• Dataset: 13,611 samples with 16 features
• Missing Values: None detected
• Train-Test Split: 80% training, 20% testing
• Stratification: Applied to maintain class distribution

Model Evaluation: Three Versions

Version 1: Basic Logistic Regression

Configuration:

• Default scikit-learn LogisticRegression parameters
• No feature scaling
• Standard regularization

Results

Accuracy: 0.6988615497612927

Classification Report

Class	Precision	Recall	F1-Score	Support
BARBUNYA	0.67	0.52	0.59	265
BOMBAY	1.00	0.99	1.00	104
CALI	0.73	0.78	0.75	326
DERMASON	0.81	0.87	0.84	709
HOROZ	0.60	0.55	0.58	386
SEKER	0.67	0.62	0.65	406
SIRA	0.56	0.61	0.59	527

Overall Accuracy: 0.70

Averages

Average Type	Precision	Recall	F1-Score
Macro Avg	0.72	0.71	0.71
Weighted Avg	0.70	0.70	0.70

Version 2: Tuned Logistic Regression with Feature Scaling

Improvements over Version 1:

• Feature Scaling: Applied StandardScaler for normalized features
• Hyperparameter Tuning:
  • max_iter=300 (increased iterations)
  • C=0.5 (regularization strength)
  • tol=0.001 (tolerance for stopping criteria)
  • class_weight='balanced' (handles class imbalance)

Updated Evaluation Metrics

Accuracy: 0.9162688211531399

Classification Report

Class	Precision	Recall	F1-Score	Support
BARBUNYA	0.94	0.90	0.92	265
BOMBAY	1.00	1.00	1.00	104
CALI	0.93	0.94	0.94	326
DERMASON	0.94	0.88	0.91	709
HOROZ	0.96	0.95	0.95	386
SEKER	0.92	0.95	0.94	406
SIRA	0.83	0.89	0.86	527

Overall Accuracy: 0.92

Averages

Average Type	Precision	Recall	F1-Score
Macro Avg	0.93	0.93	0.93
Weighted Avg	0.92	0.92	0.92

Key Benefits:

• Better convergence with scaled features
• Improved handling of class imbalance
• Enhanced model stability

Version 3: Optimized Configuration

Improvements over Version 2:

• Advanced Solver: solver='lbfgs' for better optimization
• Increased Iterations: max_iter=1000 for guaranteed convergence
• Feature Scaling: Maintained StandardScaler preprocessing
• Optimized Parameters: Fine-tuned for best performance

Updated Evaluation Metrics for Version 3

Accuracy: 0.9214102093279471

Classification Report

Class	Precision	Recall	F1-Score	Support
BARBUNYA	0.95	0.89	0.92	265
BOMBAY	1.00	1.00	1.00	104
CALI	0.93	0.94	0.94	326
DERMASON	0.92	0.91	0.92	709
HOROZ	0.96	0.95	0.96	386
SEKER	0.93	0.95	0.94	406
SIRA	0.85	0.88	0.86	527

Overall Accuracy: 0.92

Average Metrics

Average Type	Precision	Recall	F1-Score
Macro Avg	0.94	0.93	0.93
Weighted Avg	0.92	0.92	0.92
Performance Highlights:

• Highest accuracy achieved
• Best convergence properties
• Most stable predictions

Performance Comparison

Version	Configuration	Key Improvements	Performance Notes
V1	Basic LR	Baseline model	Initial benchmark
V2	Scaled + Tuned	Feature scaling, balanced weights	Better class handling
V3	Optimized	LBFGS solver, extended iterations	Highest accuracy

Visualizations

The project includes comprehensive data exploration:

• Class Distribution: Count plots showing bean variety frequencies
• Feature Relationships: Scatter plots of area vs perimeter, major vs minor axis
• Pairwise Analysis: Feature correlation visualization
• Confusion Matrices: Detailed classification performance for each model

Example Pair Plot

Key Findings

1. Feature Scaling Impact: Significant improvement when features are standardized
2. Class Imbalance: Some bean varieties are underrepresented in the dataset
3. Feature Importance: Area, perimeter, and axis measurements are strong predictors
4. Model Convergence: LBFGS solver with increased iterations provides best results

Usage

# Load and preprocess data
df = pd.read_excel("dry_bean_dataset.xlsx")
features = df.drop('class', axis=1)
target = df['class']

# Split data
X_train, X_test, y_train, y_test = train_test_split(
    features, target, test_size=0.2, random_state=42, stratify=target
)

# Scale features (Version 2 & 3)
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

# Train model (Version 3 - Optimized)
model = LogisticRegression(max_iter=1000, solver='lbfgs')
model.fit(X_train_scaled, y_train)

# Make predictions
predictions = model.predict(X_test_scaled)

Project Structure

Dry Bean Prediction/
├── assignment.ipynb          # Main analysis notebook
├── dry_bean_dataset.xlsx     # Dataset file
├── README.md                 # Project documentation

Future Improvements

1. Algorithm Exploration: Test Random Forest, SVM, or ensemble methods
2. Feature Engineering: Create additional derived features
3. Cross-Validation: Implement k-fold validation for robust evaluation
4. Hyperparameter Optimization: Use GridSearchCV or RandomizedSearchCV
5. Deep Learning: Explore neural network architectures

Conclusion

This project demonstrates progressive improvement in machine learning model development through iterative refinement. Version 3 achieves the best performance by combining proper feature scaling, optimized hyperparameters, and advanced solver configuration. The systematic approach shows how small adjustments can lead to significant performance gains in classification tasks.

License

This project uses the dry bean dataset under academic citation guidelines. Please cite the original authors when using this work.