A research-style project exploring how machine learning systems behave under real-world imperfect data conditions.
Real-world data is rarely clean or complete.
This project focuses on understanding how machine learning models behave when:
- data is noisy
- data is missing
- important features are removed
- conditions change gradually
The goal is not just to optimize accuracy, but to study how systems respond when things go wrong.
- What happens when data breaks?
- What happens when data is missing?
- Which features actually matter?
- Final insights
- Do different models break differently?
- What happens when important features disappear?
- When models become overconfident
- How robust is a model to increasing noise?
-
Noise Injection
Introduced randomness to observe impact on accuracy -
Missing Data Simulation
Tested how incomplete inputs affect model performance -
Feature Importance Analysis
Identified which features influence predictions most -
Model Comparison
Compared how different models respond to imperfect data -
Feature Removal Sensitivity
Removed key features to observe system degradation -
Robustness Curve
Measured how accuracy changes as noise increases
- Not all imperfections affect systems equally
- Removing critical features causes sharper failure than random noise
- Models can appear stable while becoming internally unreliable
- Performance degradation is gradual, not always immediate
- Python
- NumPy, Pandas
- Scikit-learn
- Matplotlib
To move beyond “building models” and toward understanding how systems behave when things go wrong.