Data Reliability & Noisy Input Handling in ML Models

This project analyzes how noisy and incomplete data affects the performance of machine learning models on a binary classification task. It builds a clean baseline model, then systematically corrupts the data with different levels of noise and missing values, and evaluates how model performance degrades.

Features

• Synthetic binary classification dataset (no external data needed)
• Train/test split and preprocessing
• Training of Logistic Regression and RandomForestClassifier
• Evaluation under multiple noise levels:
  • Feature noise (Gaussian noise)
  • Missing values injection
• Summary report printed to the console
• Optional visualization of performance vs. noise level

Project Structure

• src/data_prep.py
  Generates a clean synthetic dataset and saves it as data/clean_data.csv.

• src/train_model.py
  Trains baseline models on the clean dataset and saves them to models/.

• src/evaluate_noise.py
  Loads the trained models, applies different noise levels to the data and reports performance.

• data/clean_data.csv
  Synthetic dataset generated by data_prep.py.

How to Run

1. Create and activate a virtual environment (optional but recommended):

python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

2. Install dependencies:

pip install -r requirements.txt

3. Generate the synthetic dataset:

python src/data_prep.py

4. Train baseline models:

python src/train_model.py

5. Evaluate robustness under noise:

python src/evaluate_noise.py

This will print accuracy scores under different noise and missing-value levels. It will also save an optional plot noise_impact.png in the project root.

Requirements

• Python 3.9+ (should also work on 3.8+)
• Dependencies listed in requirements.txt

The code is written to be simple, readable and easily extensible for academic experiments or teaching purposes.