Income Classification 📊

End-to-end machine-learning pipeline.
Using the 1994 UCI Adult (“Census Income”) data, we predict whether an individual earns > $50 K / year and compare four classical & ensemble algorithms.

---

🗂 Dataset

Item	Value
File	data/censusData.csv
Source	UCI Machine-Learning Repository ▸ Adult (slightly cleaned for education)
Size	32 561 rows × 15 raw features

---

🔍 Problem Definition

Aspect	Details
Learning type	Supervised
Task	Binary classification
Target	income_binary → 0 = ≤ $50 K, 1 = > $50 K
Feature groups	• Continuous (6) age, fnlwgt, education-num, capital-gain, capital-loss, hours-per-week • Categorical (9) workclass, education, marital-status, occupation, relationship, race, sex, native-country, income

---

⚙️ ML Workflow

1. Data Pre-processing

   • Drop rows with placeholder “?” entries.
   • Label-encode target; one-hot encode all categorical predictors.
   • Standardize continuous variables with StandardScaler.

2. Feature Selection

   • Use SelectKBest with mutual-information scores to keep the top 20 most informative features.

3. Model Training

   Model	Library / API	Key Hyper-params
   Logistic Regression	sklearn.linear_model.LogisticRegression	solver='lbfgs', max_iter=1000
   Random Forest	sklearn.ensemble.RandomForestClassifier	n_estimators=300, max_depth=None
   Gradient Boosting (GBDT)	sklearn.ensemble.GradientBoostingClassifier	n_estimators=400, learning_rate=0.05
   Stacking Ensemble	sklearn.ensemble.StackingClassifier	Base: LR + RF + GBDT → Meta: LR

4. Evaluation

   • Confusion matrices, accuracy, precision-recall curves, and ROC-AUC.
   • Training-vs-validation learning curves to spot over/under-fitting.
   • Basic fairness probe: check error rates across sex & race.

---

📈 Results

Model	Accuracy (%)	Notes
Logistic Regression	84.5	Fast & highly interpretable
Random Forest	86.3	Captures non-linear splits; small memory cost
Gradient Boosting	87.3	Best single model; slower to train
Stacking Ensemble	87.4	Marginal lift above GBDT

(Accuracies computed on held-out test set of 6 513 records.)

---

🔬 Fairness & Bias Checks

Preliminary disparate-impact analysis shows higher false-negative rates for women and certain racial groups. Mitigation strategies (re-sampling, re-weighting, post-processing thresholds) are discussed in notebooks/04_fairness_analysis.ipynb.

---

🛠 Tech Stack

• Python 3.11
• pandas, NumPy
• scikit-learn
• Matplotlib, Seaborn