Scientific Data Preprocessing

# Check data types
df.dtypes  # int64, float64, object, etc.

# Binary (0/1)
binary_features = [col for col in df.columns if df[col].nunique() == 2]

# Categorical (finite discrete values)
categorical_features = [col for col in df.select_dtypes(include='object').columns]

# Continuous (infinite possible values)
continuous_features = [col for col in df.select_dtypes(include=['float64', 'int64']).columns
                       if df[col].nunique() > 10]

# Check for grouping
print(f"Number of groups: {df['group_id'].nunique()}")
print(f"Avg points per group: {df.groupby('group_id').size().mean():.1f}")

# Check for time-series
df_sorted = df.sort_values(['group_id', 'timestamp'])

# Validate physical ranges
assert df['speed_mph'].max() < 200, "Speed exceeds physical limit"
assert df['distance_meters'].min() >= 0, "Negative distance impossible"

# Decision: Within-group or global processing?
def choose_processing_scope(data, feature, goal):
    """
    goal = 'relative' → within-group (e.g., "this point was intense FOR THIS MATCH")
    goal = 'absolute' → global (e.g., "this was an intense point OVERALL")
    """
    if goal == 'relative':
        return 'within_group'
    elif goal == 'absolute':
        return 'global'
    else:
        raise ValueError("Goal must be 'relative' or 'absolute'")

from scipy.interpolate import CubicSpline
import numpy as np

# ✅ CORRECT: Interpolate within each group
for group_id in df['match_id'].unique():
    mask = df['match_id'] == group_id
    group_data = df.loc[mask, 'speed_mph'].copy()

    # Get valid (non-NaN) indices
    valid_idx = group_data.notna()
    valid_positions = np.where(valid_idx)[0]
    valid_values = group_data[valid_idx].values

    if len(valid_positions) >= 4:
        cs = CubicSpline(valid_positions, valid_values)
        missing_positions = np.where(~valid_idx)[0]
        df.loc[mask & ~valid_idx, 'speed_mph'] = cs(missing_positions)

# ❌ WRONG: Cross-group interpolation
# This interpolates between match A's last point and match B's first point!
cs = CubicSpline(
    np.where(df['speed_mph'].notna())[0],  # ❌ All indices globally
    df['speed_mph'].dropna().values
)
df.loc[df['speed_mph'].isna(), 'speed_mph'] = cs(
    np.where(df['speed_mph'].isna())[0]
)

from sklearn.preprocessing import StandardScaler

# ✅ CORRECT: Standardize within each match
for match_id in df['match_id'].unique():
    mask = df['match_id'] == match_id
    scaler = StandardScaler()

    df.loc[mask, 'distance_run_std_within'] = scaler.fit_transform(
        df.loc[mask, [['distance_run']]
    )

# Interpretation: z=+2 means "2 std above average FOR THIS MATCH"

# ✅ CORRECT: Global standardization (when appropriate)
scaler = StandardScaler()
df['distance_run_std_global'] = scaler.fit_transform(df[['distance_run']])

# Interpretation: z=+2 means "2 std above average ACROSS ALL MATCHES"

# Binary variables (0/1) - KEEP AS-IS
binary_cols = ['is_ace', 'is_winner', 'is_error']
# ❌ NEVER standardize these! They have semantic meaning as 0/1

# Categorical variables - ONE-HOT ENCODE
df_encoded = pd.get_dummies(df, columns=['server', 'serve_number'], dtype=int)

# Continuous variables - STANDARDIZE (within-group or global)
continuous_cols = ['distance_run', 'rally_count', 'speed_mph']
# ✅ Apply pattern 3 or 4 based on goal

# ✅ CORRECT: Sliding window for momentum analysis
window = 10

df['win_rate_last10'] = df.groupby('match_id')['point_won'].transform(
    lambda x: x.rolling(window, min_periods=1).mean()
)

# ❌ WRONG: Cumulative features (loses temporal locality)
df['cumulative_points_won'] = df.groupby('match_id')['point_won'].cumsum()
# This just increases monotonically and correlates with point_number

def validate_data_quality(df, feature, expected_range):
    """Validate before processing"""
    # Check range
    assert df[feature].min() >= expected_range[0], f"{feature} below minimum"
    assert df[feature].max() <= expected_range[1], f"{feature} above maximum"

    # Check for anomalies
    mean = df[feature].mean()
    std = df[feature].std()

    if std > mean:
        print(f"⚠️ WARNING: {feature} has std > mean (highly skewed or errors)")

    # Check missing pattern
    missing_by_group = df.groupby('match_id')[feature].apply(lambda x: x.isna().sum())
    if missing_by_group.max() > len(df) / df['match_id'].nunique() * 0.5:
        print(f"⚠️ WARNING: {feature} has >50% missing in some groups")

# Example
validate_data_quality(df, 'speed_mph', expected_range=(50, 165))

def detect_processing_scope(df, group_col, feature_col):
    """
    Recommend within-group vs global based on variance structure
    """
    # Calculate variance components
    within_group_var = df.groupby(group_col)[feature_col].var().mean()
    global_var = df[feature_col].var()

    # Intraclass correlation
    between_group_var = global_var - within_group_var
    icc = between_group_var / global_var

    if icc > 0.5:
        return 'within_group', f"High between-group variance (ICC={icc:.2f})"
    else:
        return 'global', f"Low between-group variance (ICC={icc:.2f})"

scope, reason = detect_processing_scope(df, 'match_id', 'distance_run')
print(f"Recommended: {scope} - {reason}")

def detect_data_leakage(df, target_col, feature_cols, id_cols):
    """
    Critical checks for data leakage and AI common pitfalls
    """
    issues = []

    # 1. ID Leakage: High cardinality variables as features
    for col in feature_cols:
        if col in id_cols:
            issues.append(f"❌ FATAL: {col} is an ID - NEVER use as feature")
            continue

        # Check if looks like ID (>50% unique)
        uniqueness = df[col].nunique() / len(df)
        if uniqueness > 0.5:
            issues.append(f"⚠️ {col}: {uniqueness*100:.1f}% unique - possible ID leakage")

    # 2. Causal Inversion: Perfect correlation with target
    for col in feature_cols:
        if col == target_col:
            continue
        if df[col].dtype in ['int64', 'float64']:
            corr = abs(df[[col, target_col]].corr().iloc[0, 1])
            if corr > 0.95:
                issues.append(f"❌ FATAL: {col} correlation={corr:.3f} - likely consequence of target!")

    # 3. Meaningless Numeric: Codes treated as numbers
    for col in feature_cols:
        if df[col].dtype in ['int64', 'float64']:
            # Pattern: High values, many uniques, looks like code
            if df[col].min() > 1000 and df[col].nunique() > 100:
                issues.append(f"⚠️ {col}: Looks like code (zipcode/ID) - should be categorical")

    # 4. Time Travel: Check if standardization used global statistics
    # (Requires knowing if train/test split was done first)

    # Print report
    if issues:
        print("="*60)
        print("DATA LEAKAGE AUDIT")
        print("="*60)
        for issue in issues:
            print(issue)
        print("="*60)
    else:
        print("✅ No obvious leakage detected")

    return issues

# Example usage
issues = detect_data_leakage(
    df,
    target_col='point_won',
    feature_cols=['speed_mph', 'user_id', 'distance_run'],
    id_cols=['match_id', 'user_id']
)

from scipy.stats import skew, kurtosis
from sklearn.preprocessing import StandardScaler, RobustScaler

def smart_scaler_selection(df, col):
    """
    Choose scaler based on distribution characteristics
    """
    data = df[col].dropna()

    # Check distribution
    skewness = skew(data)
    kurt = kurtosis(data)

    print(f"{col}: skewness={skewness:.2f}, kurtosis={kurt:.2f}")

    if abs(skewness) < 0.5 and abs(kurt) < 3:
        # Roughly normal
        print("  → StandardScaler (data is roughly normal)")
        return StandardScaler(), None

    elif skewness > 1:
        # Right-skewed (long tail)
        print("  → Log transform + StandardScaler (right-skewed)")
        return StandardScaler(), 'log'

    else:
        # Heavy outliers or non-normal
        print("  → RobustScaler (heavy outliers)")
        return RobustScaler(), None

# Example usage
for col in continuous_features:
    scaler, transform = smart_scaler_selection(df, col)

    if transform == 'log':
        df[f'{col}_log'] = np.log1p(df[col])
        df[f'{col}_scaled'] = scaler.fit_transform(df[[f'{col}_log']])
    else:
        df[f'{col}_scaled'] = scaler.fit_transform(df[[col]])

import pandas as pd
from sklearn.preprocessing import StandardScaler

# 1. Load and inspect
df = pd.read_csv('tennis_data.csv')
print(f"Matches: {df['match_id'].nunique()}")
print(f"Features: {df.dtypes}")

# 2. Classify features
binary_features = ['is_ace', 'is_winner', 'is_break_point']
categorical_features = ['server', 'serve_number']
continuous_features = ['distance_run', 'speed_mph', 'rally_count']

# 3. Validate data quality
for feat in continuous_features:
    print(f"\n{feat}:")
    print(df[feat].describe())
    # Check for impossible values
    if feat == 'speed_mph':
        assert df[feat].max() < 170, "Speed exceeds world record!"

# 4. Handle missing values (within-group)
for match_id in df['match_id'].unique():
    mask = df['match_id'] == match_id
    for feat in continuous_features:
        if df.loc[mask, feat].isna().any():
            # Simple linear interpolation within match
            df.loc[mask, feat] = df.loc[mask, feat].interpolate(method='linear')

# 5. One-hot encode categorical
df = pd.get_dummies(df, columns=categorical_features, dtype=int)

# 6. Standardize continuous features WITHIN each match
for feat in continuous_features:
    df[f'{feat}_std'] = np.nan
    for match_id in df['match_id'].unique():
        mask = df['match_id'] == match_id
        scaler = StandardScaler()
        df.loc[mask, f'{feat}_std'] = scaler.fit_transform(
            df.loc[mask, [[feat]]
        )

# 7. Create sliding window features
window = 10
df['win_rate_last10'] = df.groupby('match_id')['point_won'].transform(
    lambda x: x.rolling(window, min_periods=1).mean()
)

# 8. KEEP binary features as 0/1 (don't transform!)
# binary_features are already correct

print("\n✅ Preprocessing complete!")
print(f"Final shape: {df.shape}")
print(f"Standardized features: {[f for f in df.columns if f.endswith('_std')]}")