CharityML – Income Prediction for Donor Targeting (Supervised Machine Learning – Binary Classification)

CharityML – Income Prediction for Donor Targeting

(Supervised Machine Learning – Binary Classification)

Objective

Built a predictive model to help a non-profit identify individuals likely to donate by predicting whether annual income > $50,000 using the 1994 UCI Adult Census dataset.

Dataset

~32,561 records (after cleaning → ~26,900 usable)

Severe class imbalance (~25.6% >$50K)

14 raw features (6 numerical + 8 categorical) → expanded to ~103 features after encoding

Key Technical Achievements

Data Preparation & Cleaning

Removed duplicates, handled missing values (' ?' → NaN), standardized target labels

One-hot encoding of categorical variables

MinMax scaling of continuous features

Exploratory Data Analysis

Identified strong predictors early: capital-gain, age, education level, marital status

Visualized feature importance and class distribution

Modeling & Optimization

Compared 4 algorithms: GaussianNB, DecisionTree, AdaBoost, RandomForest

Used GridSearchCV for hyperparameter tuning

Optimized for Fβ=0.5 score (prioritizing precision – critical for donor targeting)

Final Model Performance (best configuration – full feature set)

Accuracy: 85.6%

Fβ=0.5: 72.4%

Feature Importance & Model Simplification

Top 5 most predictive features (Random Forest):

age

capital-gain

education-num

hours-per-week

marital-status_Married-civ-spouse

Model trained only on these 5 features:

Accuracy: 84.1% (–1.5%)

Fβ=0.5: 68.4% (–4%)

Training time reduced by ~45–50%

Business & Modeling Takeaways

Capital-gain is by far the strongest single signal

Marital status (especially “Married-civ-spouse”) is a surprisingly powerful proxy for higher household income

~95% of predictive power can be captured with only 5 features → excellent trade-off between performance and simplicity

Precision-focused evaluation metric (F0.5) was correctly chosen for real-world donor outreach use-case

Technologies & Skills Demonstrated

Python, pandas, scikit-learn

Data cleaning & preprocessing pipelines

One-hot encoding + feature scaling

Model comparison & hyperparameter tuning (GridSearchCV)

Feature importance analysis & dimensionality reduction

Handling class imbalance & appropriate metric selection (Fβ score)

End-to-end ML workflow in Jupyter Notebook

Clear business interpretation & trade-off analysis