📊 Machine Learning Project Portfolio

This repository contains a diverse set of machine learning projects exploring regression, classification, clustering, and inductive logic programming (ILP) tasks. These mini-projects apply multiple learning algorithms to structured real-world datasets and analyze their performance using relevant metrics.

🔢 Summary of Experiments

Task Type	Dataset Description	Models/Techniques Used
Clustering	Lifestyle & obesity prediction (2,078 rows)	K-Means, Hierarchical Clustering, PCA
Regression (Small)	Garment productivity prediction (1,197 rows)	SVR, MLP, Decision Tree
Regression (Large)	IT incident resolution prediction (141,712 rows)	SVR, MLP, Decision Tree
Classification	Income level prediction (>50K) (48,842 rows)	LinearSVC, MLPClassifier, Decision Tree
ILP (Symbolic)	Income rule learning via Inductive Logic Programming (500 instances)	Aleph (Prolog-based ILP engine)

📅 Subfolders

• clustering – Obesity risk analysis using K-Means and Hierarchical Clustering
• regression_small – Predicting garment productivity
• regression_large – Predicting IT incident resolution times
• classification – Income classification with various ML models
• ilp – Rule-based classification using Aleph ILP engine

🔍 Discussion of Results and Critical Assessment

🔄 Regression

• IT Incident Resolution: MLP achieved the best performance (MAE = 0.823, R² = 0.755), capturing complex patterns in larger datasets. Decision Trees provided moderate performance and transparency, while SVR struggled with linear assumptions.

• Garment Productivity: On the smaller, noisier dataset, Decision Tree achieved the highest test R² (0.39). This shows that simpler models often perform better on small datasets, especially when complex models may overfit.

🧠 Classification & ILP

• Income Prediction: MLPClassifier performed best in terms of accuracy and ROC-AUC. LinearSVC showed higher recall, valuable in contexts where missing true positives is costly. Decision Trees balanced accuracy and interpretability.

• ILP: Performance improved from 80% to 98% when background knowledge and parameters were fine-tuned. The trade-off observed was between rule generalization and instance-specific overfitting, showcasing ILP’s power in symbolic, structured domains.

🔎 Clustering

• Obesity Risk Clustering: K-Means outperformed Hierarchical Clustering based on Silhouette Score and Davies-Bouldin Index. Clusters were well-separated and interpretable, confirming that clustering algorithm selection must consider data assumptions.

💪 Skills Demonstrated

• Data preprocessing and cleaning for real-world datasets
• Model selection and hyperparameter tuning
• Evaluation using cross-validation and domain-relevant metrics
• Interpretation of results and comparison across models
• Use of both symbolic (ILP) and statistical ML approaches

🛠️ Tools & Libraries

• Python, Jupyter Notebooks
• scikit-learn, pandas, numpy, matplotlib, seaborn
• SWI-Prolog and Aleph ILP engine
• Git & GitHub for version control and documentation

👩‍💻 Author

Shivasmi Sharma
Bioinformatics Analyst | MSc Data Science
📧 Email: shivasmisharma646@gmail.com
🔗 LinkedIn: linkedin.com/in/shivasmi-sharma
🐙 GitHub: github.com/shivasmi07