Skip to content

README.md

Latest commit

 

History

History
71 lines (59 loc) · 4.78 KB

README.md

File metadata and controls

71 lines (59 loc) · 4.78 KB

Machine-learning-projects (OLD: 2024 at my early career)

These are small machine learning projects I did early in my career to learn and explore the field. They were training projects and I kept them even though I have better, more polished work now. I want to document these older notebooks so the experiments and lessons are preserved.

Overview

  • 4 datasets used across the projects (Market basket analysis data, Telecom customer churn, Mall Customers Dataset, House prices prediction)
  • 7 algorithms I experimented with across those datasets.
  • Each project is implemented as one or more Jupyter Notebooks for preprocessing, model training, evaluation, and visualizations.
  • The repo is intentionally lightweight and experimental think of it as a learning journal.

Quick note These notebooks are from my early experiments. Some techniques maybe the most convinent and some are not I didn't know about topics like (data leakage, statistics and data distribution, Piplines and reproducable preprocessing for new input data, but later you will find more advanced techniques); the goal back then was to understand concepts rather than produce production-ready pipelines.

Projects and datasets

  • Telecom Customer Churn (classification)

    • Algorithms used: Logistic Regression, Random Forest
    • What I learned: handling imbalanced classes, working with categorical variables, model evaluation (ROC, confusion matrix), and feature importance.
    • See: Classification Task — includes a detailed README for the churn project.

  • Car Prices (regression)

    • Algorithms used: Random forest Regressor
    • What I learned: feature selection, baseline vs. complex models, evaluation with RMSE/MAE/R², and simple model interpretation.

  • Mall Customer Segmentation (clustering)

    • Algorithms used: KMeans, DBSCAN
    • What I learned: how to standardize features for clustering, selecting k (elbow, silhouette), dimensionality reduction for visualization (PCA / t-SNE), and interpreting clusters.

  • Market-basket / Groceries (association rules)

    • Algorithms used: Apriori (frequent itemset mining and association rules)
    • What I learned: transforming transactional data, mining frequent patterns, interpreting support/confidence/lift, and extracting actionable rules.

(That list totals 7 core algorithms I focused on across the 4 datasets: Logistic Regression, Random Forest Regressor and Classifier, KMeans, DBSCAN, FPGrowhs and Apriori.)

Implementation techniques, tools and patterns I used

  • Data cleaning

    • Missing-value strategies (drop, simple imputation, KNN imputation where appropriate)
    • Duplicate removal and basic sanity checks

  • Encoding categorical variables

    • Label encoding, binary mapping for two-level categoricals, and one-hot encoding for multilevel categories

  • Feature engineering and transformations

    • Creating derived features, converting types (e.g., TotalCharges → numeric), binning where useful

  • Scaling & dimensionality reduction

    • StandardScaler / MinMaxScaler, PCA and t-SNE for visualization and noise reduction

  • Imbalance handling

    • SMOTE from imbalanced-learn for classification experiments

  • Model training & evaluation

    • Train / test split, cross-validation (K-fold), GridSearchCV for simple hyperparameter tuning
    • Metrics: accuracy, precision, recall, F1, ROC-AUC for classification; RMSE, MAE, R² for regression; silhouette score for clustering

  • Model pipelines and reproducibility

    • scikit-learn Pipelines to combine preprocessing and modeling steps
    • Saving models with joblib/pickle when needed

  • Visualization and interpretation

    • matplotlib, seaborn plots for EDA and model diagnostics (confusion matrices, ROC curves, feature importances)

  • Association rules-specific

    • Transaction encoding (one-hot), Apriori algorithm and filtering rules by support/confidence/lift

  • Libraries I used commonly

    • pandas, numpy, scikit-learn, imbalanced-learn, mlxtend (for Apriori), matplotlib, seaborn, joblib/pickle, scipy

Why I kept these projects

  • They were my hands-on introduction to:
    • The end-to-end ML workflow (EDA → preprocessing → modeling → evaluation)
    • The practical quirks of real datasets (dirty types, missing values, imbalance)
    • Basic model selection and simple hyperparameter tuning
  • Even though the code is not state-of-the-art, the notebooks capture the learning process and are useful reference material.

Tips when reproducing results

  • Install the packages in the versions the notebooks were written for, or be prepared to adjust minor deprecated API calls/ Libraries - methods.
  • Some notebooks may include cells that download or reference datasets externally you can find it in the data folder.
  • If you want deterministic results, set random_state seeds in modeling and sampling calls.