IntelexVsCuda

Benchmarking Intel® Extension for Scikit-learn (scikit-learn-intelex) vs. CuPy-accelerated methods for machine learning workflows.

Overview

This project benchmarks and compares the performance of various machine learning algorithms using Intel's scikit-learn-intelex and NVIDIA's CuPy acceleration across multiple datasets. The goal is to determine which optimization—CPU-based (Intelex) or GPU-based (CuPy)—performs better for specific algorithms and dataset types, as part of a college research project.

• scikit-learn-intelex: Optimizes select scikit-learn estimators using low-level CPU instructions and multi-threading for faster performance.
• CuPy: GPU array library with a NumPy-compatible API; enables GPU acceleration for array computations.

Setup

Install the required dependencies:

pip install scikit-learn scikit-learn-intelex cupy-cuda xgboost

If you encounter missing dependencies, also install:

pip install pandas matplotlib numpy

Note: Some datasets or algorithms may require additional libraries (e.g., warnings for suppression).

Experiment Methodology

1. Preprocessing: Each dataset was cleaned and preprocessed to fit the requirements of the algorithms and libraries.
2. Model Selection: The following ML models were tested:
   • LightGBM
   • XGBoost
   • Logistic Regression
   • Random Forest
   • AdaBoost
   • Multi-Layer Perceptron (MLP)
3. Implementation Variants:
   • BareBones: Standard scikit-learn implementation.
   • Intelex: scikit-learn accelerated with scikit-learn-intelex.
   • CuPy: Data arrays and supported models accelerated with CuPy (where possible).
4. Benchmarking: Each model was run on all datasets, measuring wall-clock training time.

How the experiments were conducted

Hardware used:

• CPU: Intel Core i5 (12th Gen)
• GPU: NVIDIA GeForce GTX 1660 Super
• Storage: SSD
• RAM: 32 GB

Datasets

Name	Size	Features	Notes
Wiretap	7 GB	115 numeric	Largest
Loan Prediction	20 MB	12 categorical	Small
Student Performance	41 KB	10 numeric	Smallest
Diabetes	636 KB	10 numeric	Small
Cyberbullying Detection	3.67 MB	7 text/cat.	Medium, text-heavy

Results

Key Observations

1. Correlation Between Dataset Size and Speed

• Small datasets (Student Performance, Diabetes): Most models execute in under 1 second (except MLP).
• Medium datasets (Loan Prediction, Cyberbullying): Performance varies widely by model and implementation. Some models (MLP) are outliers in runtime.
• Large datasets (Wiretap): Training can take several minutes, especially with Random Forest, AdaBoost, and MLP.

2. Model Performance Trends

• LightGBM: Fastest across all datasets.
• XGBoost: Consistently fast, second only to LightGBM.
• Logistic Regression: Efficient on small data, but can spike in runtime for certain medium datasets.
• Random Forest & AdaBoost: Slow on large datasets; can be impractical for real-time use.
• MLP: Slowest, especially on large/medium datasets.

3. Impact of Acceleration (Intelex, CuPy)

• Intelex & CuPy: Both can provide major speed-ups, but the effect depends on the dataset and model.
• For small datasets: Acceleration is noticeable, but even non-accelerated models are fast.
• For large datasets: Acceleration is crucial for practical runtimes, but not all models benefit equally.
• Note: Some models or array conversions can cause CuPy to worsen performance (e.g., Logistic Regression on Loan Prediction).

4. General Takeaways

• LightGBM is the most robust and efficient overall.
• CuPy and Intelex accelerations are not universally beneficial—always profile for your use case.
• Dataset size and model choice both impact execution time, but model/implementation often matters more than raw size.

---

Contributing

Feel free to fork this repository, create a new branch, and submit a pull request with your changes. Please make sure to document it well.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

Diabetes Dataset (Version V1). (n.d.). [Dataset]. kaggle. https://www.kaggle.com/datasets/asinow/diabetes-dataset/data

Maree, A. (2025). Student Performance Prediction [Dataset]. In Kaggle (Version V2). Keggal. https://www.kaggle.com/datasets/amrmaree/student-performance-prediction

Mirsky, Yisroel, Doitshman, Tomer, Elovici, Yuval, & Shabtai, Asaf. (2018). Kitsune: An Ensemble of Autoencoders for Online Network Intrusion Detection. In The Network and Distributed System Security Symposium (NDSS) 2018.

CyberBullying Detection Dataset. (n.d.). [Dataset]. In Kaggle (Version V2). Keggal. https://www.kaggle.com/datasets/sayankr007/cyber-bullying-data-for-multi-label-classification?select=final_hateXplain.csv

Star, Ethical. (2025). Loan Prediction. Kaggle.com.. https://www.kaggle.com/datasets/ethicalstar/loan-prediction/code