Correlation

r = +1.0: perfect positive linear relationship
           As X increases, Y increases proportionally

r = 0.0:  no linear relationship
           X tells you nothing about Y (linearly)

r = -1.0: perfect negative linear relationship
           As X increases, Y decreases proportionally

r ∈ (-1, 1) in practice — exactly ±1 only in deterministic relationships

r = Σ[(xᵢ - x̄)(yᵢ - ȳ)] / [n × σₓ × σᵧ]

Assumptions:
  Both variables are continuous
  Relationship is approximately linear
  No severe outliers (outliers distort r)
  Both roughly normally distributed (for significance testing)

import numpy as np
from scipy import stats
import pandas as pd

x = np.array([1, 2, 3, 4, 5])
y = np.array([2, 4, 5, 4, 5])

# Pearson correlation
r, p_value = stats.pearsonr(x, y)
print(f"Pearson r = {r:.3f}, p = {p_value:.4f}")

# NumPy
r_matrix = np.corrcoef(x, y)  # 2×2 correlation matrix
r = r_matrix[0, 1]

# DataFrame
df = pd.DataFrame({"x": x, "y": y, "z": [5, 3, 2, 1, 4]})
print(df.corr(method="pearson"))  # correlation matrix

Spearman ρ = Pearson correlation of the RANKS of x and y

Convert values to ranks, then compute Pearson on the ranks.
Captures: monotonic relationships (as X increases, Y tends to increase)
Misses: non-monotonic relationships (U-shaped, etc.)

# Spearman — when data is ordinal, or when Pearson assumptions are violated
rho, p_value = stats.spearmanr(x, y)
print(f"Spearman ρ = {rho:.3f}, p = {p_value:.4f}")

# Example where Spearman > Pearson (non-linear but monotonic)
x_nonlinear = np.array([1, 2, 3, 4, 5])
y_exponential = np.array([1, 4, 9, 16, 25])  # x²

r_pearson, _ = stats.pearsonr(x_nonlinear, y_exponential)
r_spearman, _ = stats.spearmanr(x_nonlinear, y_exponential)
print(f"Pearson: {r_pearson:.3f}")   # 0.975 — high but not 1.0 (non-linear)
print(f"Spearman: {r_spearman:.3f}") # 1.000 — perfect monotonic relationship

|r|         | Interpretation
------------|-----------------------------
0.00–0.10   | Negligible (no relationship)
0.10–0.30   | Weak
0.30–0.50   | Moderate
0.50–0.70   | Strong
0.70–0.90   | Very strong
0.90–1.00   | Nearly perfect

Context matters:
  r=0.3 in social science: meaningful
  r=0.3 in physics: very weak (physical laws are near-perfect)
  r=0.3 in medical prediction: potentially useful for a feature

# Find highly correlated features (potential multicollinearity)
def get_high_correlation_pairs(
    df: pd.DataFrame,
    threshold: float = 0.9,
) -> list[tuple]:
    corr_matrix = df.corr(method="pearson").abs()
    
    # Upper triangle only (avoid duplicates)
    pairs = []
    cols = corr_matrix.columns
    for i in range(len(cols)):
        for j in range(i + 1, len(cols)):
            if corr_matrix.iloc[i, j] >= threshold:
                pairs.append((cols[i], cols[j], corr_matrix.iloc[i, j]))
    
    return sorted(pairs, key=lambda x: -x[2])


# Filter features by correlation with target
def select_features_by_correlation(
    X: pd.DataFrame,
    y: pd.Series,
    min_correlation: float = 0.1,
    method: str = "spearman",
) -> list[str]:
    correlations = X.apply(lambda col: col.corr(y, method=method)).abs()
    return correlations[correlations >= min_correlation].index.tolist()


# Correlation heatmap (for exploration)
import matplotlib.pyplot as plt
import seaborn as sns

def plot_correlation_matrix(df: pd.DataFrame) -> None:
    plt.figure(figsize=(10, 8))
    sns.heatmap(
        df.corr(),
        annot=True, fmt=".2f",
        cmap="RdBu_r", center=0,
        vmin=-1, vmax=1,
    )
    plt.title("Feature Correlation Matrix")
    plt.tight_layout()

High correlation between input features → multicollinearity

Effect on linear models:
  Coefficients become unstable (high variance)
  Hard to interpret which feature drives the prediction

Effect on tree models:
  Less severe — trees can handle correlated features
  But feature importance becomes unreliable

Detection: variance inflation factor (VIF)
  VIF = 1/(1 - R²) where R² is from regressing that feature on all others
  VIF > 5: concerning; VIF > 10: serious multicollinearity

Fix:
  Remove one of the highly correlated features
  PCA to orthogonalise features
  Ridge regression (L2 regularisation handles multicollinearity well)