ML Assignment 2 - Classification Model Comparison

a) Problem Statement

The objective is to build and compare six classification models on a real-world dataset, evaluate them using six required metrics, and deploy an interactive Streamlit app where users can upload test data and inspect model performance.

b) Dataset Description

• Dataset Name: Adult Income Dataset
• Source: UCI Machine Learning Repository
• Link: https://archive.ics.uci.edu/ml/datasets/adult
• Task Type: Binary classification (income <=50K vs >50K)
• Instances: 48,842 rows before cleaning (after removing missing values represented as ?, remaining rows are used for training/testing)
• Features: 14 input features (mix of numerical and categorical), satisfies assignment requirement of at least 12 features
• Target Column: income

Project Structure

project-folder/
|-- app.py
|-- requirements.txt
|-- README.md
|-- test_data.csv
|-- data/
|   |-- adult_train.csv
|   |-- adult_test.csv
|-- model/
|   |-- train_models.py
|   |-- model_metrics.csv
|   |-- logistic_regression.pkl
|   |-- decision_tree.pkl
|   |-- knn.pkl
|   |-- naive_bayes.pkl
|   |-- random_forest.pkl
|   |-- xgboost.pkl

c) Models Used With Comparison Table

All models were trained using an 80:20 stratified train-test split with random_state=42.

ML Model Name	Accuracy	AUC	Precision	Recall	F1	MCC
Logistic Regression	0.8450	0.9020	0.7344	0.5870	0.6525	0.5601
Decision Tree	0.8100	0.7470	0.6154	0.6222	0.6188	0.4922
kNN	0.8387	0.8852	0.7007	0.6097	0.6520	0.5500
Naive Bayes	0.6193	0.8389	0.3877	0.9242	0.5462	0.3891
Random Forest	0.8569	0.9134	0.7908	0.5749	0.6658	0.5895
XGBoost	0.8690	0.9259	0.7852	0.6490	0.7106	0.6317

d) Observations Table

ML Model Name	Observation about model performance
Logistic Regression	Strong baseline with good AUC and balanced precision-recall; performs consistently on this dataset.
Decision Tree	Easy to interpret, but lower AUC and MCC compared to ensemble methods; likely overfits relative to simpler linear model.
kNN	Competitive with Logistic Regression in F1, but slightly weaker than ensemble methods in overall discrimination (AUC).
Naive Bayes	Very high recall but poor precision and low accuracy; predicts many positives, causing more false positives.
Random Forest	Better accuracy, F1, and MCC than single-tree and linear baseline; robust performance due to ensemble averaging.
XGBoost	Best overall metrics in this run (Accuracy, AUC, F1, MCC), showing strongest classification quality across metrics.

Streamlit App Features Implemented

• CSV upload option for test dataset
• Model selection dropdown (all 6 models)
• Display of required metrics: Accuracy, AUC, Precision, Recall, F1, MCC
• Confusion matrix and classification report display
• Quick download button for test_data.csv

How to Run Locally

pip install -r requirements.txt
python model/train_models.py
streamlit run app.py

GitHub Repository Link: https://github.com/gdesai23/mlassignment2 Live Streamlit App Link: https://mlassignment2-2025aa05217.streamlit.app/

Notes

• Trained model files (.pkl) are included for app inference.
• Core training logic is available in model/train_models.py.
• Use your own BITS Lab screenshot and your own repository/app links before final submission.