Correlation measures the association between two variables or features. In tabular-data work, bivariate correlation compares numeric features one pair at a time, making it useful for finding duplicated or highly similar signals before modeling. section 2.2.2 section 2.4.3
Arrange pairwise correlations in a symmetric correlation matrix and inspect it as a heatmap when there are few features. For wider datasets, list only pairs above a chosen absolute threshold. Values near 0.98–0.99, including negative values approaching −1, flag potential collinearity; 0.90–0.95 can reveal features with little unique contribution. section 2.2.2 section 2.4.3
Use correlation as a feature-review step:
- Investigate highly correlated pairs to understand why they overlap, then choose which feature to retain using target association, data quality, and domain knowledge. section 2.2.2 section 2.4.3
- Distinguish pairwise correlation from variance inflation factor (VIF): VIF assesses a feature against all other features, while correlation identifies the specific related pairs. section 2.2.2
- Check linear-regression inputs for strong feature-target Pearson correlation, while recognizing that correlated predictors can make coefficient-based explanations misleading. section 4.3.1 section 4.3.1
Screening numeric features
Create a correlation matrix with data[numeric_feats].corr() and plot it with a seaborn heatmap. In a synthetic dataset augmented with near-duplicate columns, pairwise screening above 0.90 identifies feature pairs with correlations from 0.947 to 0.966. section 2.2.2 section 2.4.3
import seaborn as sns corr = data[numeric_feats].corr() sns.heatmap(corr, cmap="Blues",annot=True)
Use the listing to compute and visualize a numeric-feature correlation matrix before deciding which highly collinear features warrant review. section 2.4.3
