AWS Certified Machine Learning – SpecialtyExploratory Data AnalysisHard
A machine learning engineer is tasked with building a model to predict equipment failure based on sensor readings. During EDA, they notice that several sensor readings ('temperature', 'vibration', 'pressure') are highly correlated with each other, exhibiting multicollinearity. This can lead to unstable model coefficients and reduced interpretability. What is the most effective data cleaning/feature engineering technique to address this multicollinearity while retaining as much information as possible from the original features?
- APerform Principal Component Analysis (PCA) to create uncorrelated features.
- BIncrease the amount of data to dilute the effect of multicollinearity.
- CRemove one of the highly correlated features, such as 'temperature'.
- DApply Min-Max scaling to all three features.
Show answer & explanationAnswer & explanation
Correct answer: A. Perform Principal Component Analysis (PCA) to create uncorrelated features.
PCA is a dimensionality reduction technique that transforms a set of correlated variables into a set of linearly uncorrelated variables called principal components. This effectively addresses multicollinearity by creating new features that capture the variance of the original correlated features without redundancy, thus retaining information.
Why the other options are wrong
- B. Incorrect. Increasing data volume does not inherently resolve multicollinearity, which is a structural property of the relationships between features, not a sample size issue.
- C. Incorrect. Removing a highly correlated feature is a valid strategy but results in a loss of potentially useful information present in the discarded feature.
- D. Incorrect. Min-Max scaling normalizes the range of features but does not address or reduce multicollinearity.
Principal Component Analysis (PCA)
A statistical procedure that uses an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of linearly uncorrelated variables called principal components.
- Reduces dimensionality by creating new, uncorrelated features.
- Effective for addressing multicollinearity.
- Retains most of the variance from original features.
Memory trick: Correlated features cause a fuss, PCA brings order to us.