DTWBoost: A Dynamic Time Warping Enhanced SVM Classifier for Time Series Data

---

📌 Abstract

This project evaluates the use of Support Vector Machine (SVM) with different kernel functions combined with Dynamic Time Warping (DTW) for classifying time series data.
We assess performance on datasets such as ECG, FordA, and HAR.

---

🔑 Keywords

• Support Vector Machine (SVM)
• Dynamic Time Warping (DTW)
• Kernel Functions
• Time Series Classification

---

🎯 Objective

Evaluate the effectiveness of various DTW-based SVM kernels on different time series datasets.

---

📚 Related Work

Past research explored DTW-SVM combinations. Kernels like Gaussian, Polynomial, and custom DTW-based ones have shown varying success.

---

⚙️ Methodology

Steps

1. Preprocessing – Normalize and balance data
2. DTW Distance Calculation
3. Kernel Application – Replace Euclidean distance with DTW in kernel
4. SVM Training – With cross-validation
5. Evaluation – Accuracy and confusion matrix

Kernels Used

• Cauchy
• Gaussian
• Inverse Multiquadric
• Laplacian
• Logarithmic
• Rational Quadratic

---

📊 Datasets

• ECG – Heartbeat signal classification
• FordA – Engine fault detection
• HAR – Human activity recognition via smartphone sensors

---

🔧 Experimental Settings

Metrics

• Accuracy
• Confusion Matrix

Sample Results (ECG)

Kernel	Hyperparameter	Time (s)
Cauchy	sigma=500000	213.1
Gaussian	sigma=1	310.7
Inverse MQ	c=0.4	346.7

Accuracy Comparison

Kernel	ECG (%)	FordA (%)	HAR (%)
Cauchy	89.0	60.5	70.0
Gaussian	76.0	45.0	57.1
Inverse MQ	85.0	60.0	80.3
Laplacian	83.0	53.5	82.7
Logarithmic	38.0	53.0	85.6
Rational Quadratic	84.0	54.0	87.0

---

💬 Insights

• Inverse Multiquadric performs best across datasets.
• DTW improves SVM accuracy but increases computation time.
• FordA is harder due to signal variance.

---

🧠 Discussion

• DTW is effective but slow; use FastDTW or parallelism for scalability.
• DTW kernels help with noisy, irregular-length time series.

---

✅ Conclusion

Combining DTW with flexible SVM kernels significantly improves time series classification.
Rational Quadratic and Inverse Multiquadric performed best overall.

---

📖 References

1. Lei & Sun, Dynamic Time Warping in Kernel Machines
2. Hofmann et al., Kernel Methods in ML
3. Zhang, List of SVM Kernels
4. Reyes-Ortiz, HAR Dataset
5. Bagnall, FordA Dataset
6. Kachuee et al., ECG Classification
7. NCBI DTW Paper

---