Performance Comparison of SVM Kernels on Fashion MNIST Dataset

 

Experiment 

Performance Comparison of SVM Kernels on Fashion MNIST Dataset

---

🎯 Objective

To implement and compare SVM classifiers with Linear, Polynomial, and RBF kernels on the Fashion MNIST dataset and analyze their performance.

---

📚 Theory

Support Vector Machines (SVM) use different kernel functions to handle different types of data:

• Linear Kernel → Straight-line separation
• Polynomial Kernel → Curved boundaries (controlled by degree)
• RBF Kernel → Highly flexible, non-linear boundaries

---

📊 Dataset

We use the Fashion MNIST dataset.

• 70,000 grayscale images (28×28 pixels)
• 10 classes (T-shirt, shoe, bag, etc.)

---

🔬 Algorithm / Procedure

1. Load dataset
2. Normalize pixel values
3. Flatten images into vectors
4. Split into training and testing sets
5. Train SVM with:
   • Linear kernel
   • Polynomial kernel
   • RBF kernel
6. Evaluate accuracy
7. Compare results

---

💻 Program

import numpy as np
from sklearn import svm, datasets
from sklearn.metrics import accuracy_score
from sklearn.model_selection import train_test_split

# Load dataset
from keras.datasets import fashion_mnist
(X_train, y_train), (X_test, y_test) = fashion_mnist.load_data()

# Reduce size (for faster training in lab)
X_train = X_train[:5000]
y_train = y_train[:5000]
X_test = X_test[:1000]
y_test = y_test[:1000]

# Normalize
X_train = X_train / 255.0
X_test = X_test / 255.0

# Flatten images
X_train = X_train.reshape(len(X_train), -1)
X_test = X_test.reshape(len(X_test), -1)

# -------- Linear SVM --------
linear_model = svm.SVC(kernel='linear')
linear_model.fit(X_train, y_train)
y_pred_linear = linear_model.predict(X_test)

# -------- Polynomial SVM --------
poly_model = svm.SVC(kernel='poly', degree=3)
poly_model.fit(X_train, y_train)
y_pred_poly = poly_model.predict(X_test)

# -------- RBF SVM --------
rbf_model = svm.SVC(kernel='rbf')
rbf_model.fit(X_train, y_train)
y_pred_rbf = rbf_model.predict(X_test)

# Accuracy
print("Linear SVM Accuracy:", accuracy_score(y_test, y_pred_linear))
print("Polynomial SVM Accuracy:", accuracy_score(y_test, y_pred_poly))
print("RBF SVM Accuracy:", accuracy_score(y_test, y_pred_rbf))

---

📈  Output 

Linear SVM Accuracy: 0.823 Polynomial SVM Accuracy: 0.814 RBF SVM Accuracy: 0.859

---

📊 Comparison Analysis

---

🔹 1. Linear Kernel

✅ Advantages:

• Fast training
• Works well for high-dimensional data
• Less risk of overfitting

❌ Disadvantages:

• Cannot capture complex patterns
• Lower accuracy for image data

---

🔹 2. Polynomial Kernel

✅ Advantages:

• Can model curved boundaries
• More flexible than linear

❌ Disadvantages:

• Sensitive to degree parameter
• Can overfit
• Slower than linear

---

🔹 3. RBF Kernel

✅ Advantages:

• Handles complex non-linear data
• High accuracy
• Most widely used

❌ Disadvantages:

• Computationally expensive
• Requires tuning (C, gamma)
• Can overfit if gamma is high

---

Key Insights

• Image data like Fashion MNIST is highly non-linear
• Linear SVM struggles to capture patterns
• RBF kernel performs best due to flexibility

---

Result

The SVM classifiers with different kernels were successfully implemented and evaluated.

• The RBF kernel achieved the highest accuracy, demonstrating its ability to model complex patterns in image data.
• The Linear kernel performed faster but with lower accuracy.
• The Polynomial kernel showed moderate performance, depending on the degree parameter.