Duration Feature Treatment (Step 2)

Overview

Duration emerges as the strongest predictor from the EDA analysis, showing the highest correlation (r=0.519) with the target variable. This step focuses on extracting maximum value from this critical feature while addressing potential data leakage concerns.

Key EDA Findings

• Strongest correlation: Duration shows r=0.519 with target (highest among all features)
• Clear engagement pattern: Subscribers average 525 seconds vs 212 seconds for non-subscribers
• Business insight: Call duration serves as an early success indicator during campaigns
• Data quality: Duration has right-skewed distribution requiring transformation

Implementation Strategy

1. Duration Binning

Created categorical bins based on engagement patterns:

• very_short (0-120s): Quick rejections
• short (120-300s): Standard interactions
• medium (300-600s): Engaged prospects
• long (>600s): High engagement calls

2. Log Transformation

Applied np.log1p() transformation to handle right-skewed distribution and reduce impact of extreme outliers.

3. High Engagement Flag

Binary feature duration_high_engagement flags calls exceeding 300 seconds, capturing the threshold where conversion likelihood increases significantly.

Code Implementation

def apply_duration_feature_treatment(train_df: pd.DataFrame, test_df: pd.DataFrame) -> Tuple[pd.DataFrame, pd.DataFrame]:
    """Apply duration feature treatment based on EDA insights."""

    # Duration bins based on engagement patterns
    duration_bins = [0, 120, 300, 600, float('inf')]
    duration_labels = ['very_short', 'short', 'medium', 'long']

    train_df['duration_bin'] = pd.cut(train_df['duration'], bins=duration_bins, labels=duration_labels, include_lowest=True)
    test_df['duration_bin'] = pd.cut(test_df['duration'], bins=duration_bins, labels=duration_labels, include_lowest=True)

    # Log transformation for skewed distribution
    train_df['duration_log'] = np.log1p(train_df['duration'])
    test_df['duration_log'] = np.log1p(test_df['duration'])

    # High engagement binary flag
    train_df['duration_high_engagement'] = (train_df['duration'] > 300).astype(int)
    test_df['duration_high_engagement'] = (test_df['duration'] > 300).astype(int)

    return train_df, test_df

Features Created

Feature	Type	Description
duration_bin	Categorical	4 engagement levels (very_short, short, medium, long)
duration_log	Numerical	Log-transformed duration to handle skewness
duration_high_engagement	Binary	Flag for calls > 300 seconds

Data Transformation Example

Before Processing

duration	y	Description
117	0	117 seconds, no subscription
185	0	185 seconds, no subscription
111	0	111 seconds, no subscription
10	0	10 seconds, no subscription

After Processing

duration	duration_bin	duration_log	duration_high_engagement	y	Analysis
117	very_short	4.77	0	0	Short call, low engagement
185	short	5.23	0	0	Standard call, low engagement
111	very_short	4.72	0	0	Quick rejection, very low engagement
10	very_short	2.40	0	0	Immediate hang-up, no engagement

Expected Impact

• Capture non-linear patterns: Binning reveals engagement thresholds
• Handle distribution skewness: Log transformation improves model performance
• Create interpretable features: Business-meaningful engagement levels
• Preserve original signal: Keep original duration alongside engineered features

Results

MLflow Performance

Duration feature treatment results:

Single Model Performance (80/20 split): - Test AUC: 0.96844

K-Fold Cross-Validation (5 folds): - Average AUC: 0.9686

Results show consistent performance with low variance across cross-validation folds. The improvement over baseline is marginal.

Classification Metrics

Performance comparison with baseline results:

• False Positives: Increased from 3,702 to 5,374 (+1,672)
• False Negatives: Reduced from 6,187 to 4,409 (-1,778)
• Trade-off: Higher recall but more false positives

Kaggle Competition Results

Competition submission results:

• Competition Score: 0.969
• Leaderboard Position: 1170