Build ML Projects That Get You Hired

• 1
  EDA — visualise churn rate, plot distributions, identify missing values and outliers
• 2
  Feature Engineering — encode categoricals, create tenure buckets, scale numerics
• 3
  Baseline Model — fit Logistic Regression, evaluate with accuracy, precision, recall, F1
• 4
  Handle Imbalance — apply SMOTE oversampling or class_weight='balanced', compare results
• 5
  Improve with RF — train Random Forest, tune n_estimators and max_depth via GridSearchCV
• 6
  Summarise — plot ROC curve, feature importances, write up business impact of each churning segment

churn-prediction/
├── data/           # raw + processed CSVs
├── notebooks/
│   ├── 01_eda.ipynb
│   ├── 02_features.ipynb
│   └── 03_models.ipynb
├── src/
│   ├── features.py
│   └── model.py
├── reports/        # charts, ROC curves
└── README.md

• 1
  EDA — plot SalePrice distribution, identify skewness, correlation heatmap
• 2
  Missing Values — impute numerics with median, categoricals with mode or "None"
• 3
  Feature Engineering — create TotalSF, HouseAge, RemodelAge; one-hot encode categoricals
• 4
  Baseline — Linear Regression, evaluate RMSE on log(SalePrice)
• 5
  Regularisation — Ridge and Lasso with alpha tuning; compare feature selection
• 6
  XGBoost — tune n_estimators, learning_rate, max_depth; plot feature importances

house-prices/
├── data/           # train.csv, test.csv
├── notebooks/
│   ├── 01_eda.ipynb
│   ├── 02_preprocessing.ipynb
│   └── 03_modelling.ipynb
├── src/
│   ├── preprocess.py
│   └── train.py
└── README.md

• 1
  EDA — visualise class imbalance, transaction amount/time distributions
• 2
  Baseline — fit Logistic Regression, show why accuracy is misleading (99.8% by predicting all legit)
• 3
  Handle Imbalance — compare SMOTE oversampling vs undersampling vs class_weight
• 4
  Ensemble Models — train Random Forest and XGBoost; compare AUC-PR
• 5
  Threshold Tuning — plot precision-recall curve, choose threshold based on business cost
• 6
  Explainability — SHAP values to show which PCA components drive fraud predictions

• 1
  EDA — attrition rate by department, job role, overtime, salary band
• 2
  Preprocessing — encode ordinal and nominal features, drop constant columns (EmployeeCount, Over18)
• 3
  Train XGBoost — with scale_pos_weight to handle imbalance, 5-fold stratified CV
• 4
  Tune Hyperparameters — GridSearchCV on max_depth, n_estimators, learning_rate, subsample
• 5
  SHAP Analysis — generate summary plot, bar plot, and waterfall for individual predictions
• 6
  Report — write 5 data-driven HR recommendations with SHAP evidence

• 1
  EDA — plot weekly sales trends, seasonality, holiday effects, store variance
• 2
  Lag Features — create lag_1w, lag_4w, lag_52w (same week last year)
• 3
  Rolling Features — 4-week and 12-week rolling mean and std, trend indicators
• 4
  Date Features — week_of_year, month, is_holiday, days_to_holiday
• 5
  Model — XGBoost with TimeSeriesSplit CV, tune with Optuna or manual grid
• 6
  Compare — benchmark XGBoost vs LightGBM vs simple moving average baseline

• 1
  Preprocessing — lowercase, remove punctuation, stopwords, stemming/lemmatisation
• 2
  Vectorise — CountVectorizer vs TF-IDF; explain the difference with examples
• 3
  Train — MultinomialNB as baseline; compare with LinearSVC and Logistic Regression
• 4
  Evaluate — focus on recall for spam (you don't want spam in inbox) vs precision tradeoff
• 5
  Pipeline — wrap TF-IDF + classifier in sklearn Pipeline; show how to predict new messages

• 1
  Feature Engineering — extract Title, FamilySize, IsAlone, CabinDeck, FarePerPerson
• 2
  Imputation — median Age by Title group, "Missing" cabin category
• 3
  Base Models — Logistic Regression, Random Forest, GradientBoostingClassifier
• 4
  Stacking — use sklearn's StackingClassifier with LR meta-learner; evaluate with stratified CV
• 5
  Ablation Study — measure accuracy impact of each engineered feature added/removed

• 1
  EDA — distribution of delinquencies, debt ratio, income; identify outliers and heavy tails
• 2
  Imputation — median for monthly income, 0 for NumberOfDependents missing
• 3
  Feature Selection — mutual information, XGBoost built-in importance, drop multicollinear features
• 4
  Train XGBoost — stratified k-fold, early stopping, scale_pos_weight for imbalance
• 5
  Calibration — apply Platt scaling (CalibratedClassifierCV) and compare reliability diagrams
• 6
  Business Threshold — sweep thresholds, compute expected loss = P(default) × loan_amount for each

• 1
  EDA — salary distributions by role, seniority, country; outlier inspection
• 2
  Encoding — ordinal for experience_level, target encoding for job_title and company_location
• 3
  Baseline — Linear Regression with one-hot encoding; set RMSE benchmark
• 4
  XGBoost — train with 5-fold CV, tune learning_rate and max_depth, evaluate MAE
• 5
  Visualise — Plotly bar chart of median salary by title; feature importance chart
• 6
  Predict Tool — simple input() loop or Streamlit widget to predict salary from user inputs

• 1
  Load Efficiently — parse with chunked read_csv, downcast dtypes, resample to hourly
• 2
  EDA — daily/weekly/seasonal decomposition, missing value patterns, outlier removal
• 3
  Feature Engineering — hour, dayofweek, month, is_weekend, lag_1h, lag_24h, lag_168h, rolling stats
• 4
  Fourier Features — sin/cos transforms of hour and day_of_year to encode cyclical patterns
• 5
  Models — LightGBM and XGBoost with TimeSeriesSplit; compare MAPE per hour-of-day
• 6
  Stack — average predictions of best LightGBM + XGBoost; show ensemble outperforms both

energy-forecasting/
├── data/           # raw UCI data + hourly resampled
├── notebooks/
│   ├── 01_eda.ipynb
│   ├── 02_features.ipynb
│   ├── 03_lgbm.ipynb
│   └── 04_ensemble.ipynb
├── src/
│   ├── features.py   # all feature engineering
│   ├── train.py      # train & evaluate pipeline
│   └── ensemble.py   # stacking logic
└── README.md