Data Ingestion, Understanding and Cleaning

Temporal Feature Engineering

Route and Pricing Feature Engineering

EXPLATORY DATA ANALYSIS

Geographic Feature Engineering

High Risk Order Identification and Prioritisation

Examines data quality through numerical ranges, variability, record counts and missing values.

• Min/Max: Checked for invalid values, such as negative monetary amounts

• Standard deviation: Assesses variability in numerical fields

• Record counts: Verify data completeness

• Missing values: Identifies fields requiring investigation

Order Price and Commission values were retained despite wide distributions, as charges vary with vehicle and service requirements. 

Missing Driver Tip and Company Bonus values were also retained, as these are optional, event-based amounts.

Derives booking lead time, urgency categories and request/pickup timing features to capture temporal patterns potentially associated with driver-matching outcomes.

Extracts route-related information and derives pricing features to capture delivery requirements and potential factors influencing driver-matching outcomes.

Enriches delivery-location data with external URA defined geographical regions to capture spatial differences in driver-matching conditions and support more comprehensive risk predictions.

DATA TRANSFORMATION

FEATURE ENGINEERING

MACHINE LEARNING: RANDOM FOREST

Historical Gap Contribution Analysis

Orders are grouped by vehicle type and booking lead time to calculate:

• Unmatched Rate: % of orders within each profile that failed to secure a driver.

• Gap Contribution: % of all historical unmatched orders attributable to each profile.

Profiles are ranked by Gap Contribution rather than Unmatched Rate alone, as high-volume profiles may account for more matching failures despite having lower failure rates.

The six highest-contributing profiles are selected for prioritisation.

Risk-Based Order Prioritisation

The framework combines two dimensions:

• Predicted Risk: Random Forest identifies individual orders at risk of failing to secure a driver.

• Operational Impact: Historical Gap Contribution identifies profiles responsible for substantial matching failures.

Orders meeting both criteria are prioritised and ranked by predicted risk with Priority A being the highest (>20%) and Priority C being the lowest (<5%). This further distinguish the historical impact of their profiles.

Actual matching outcomes are used only to evaluate the effectiveness of the prioritisation strategy.

EDA 1: Distribution of Driver Matching Outcomes

Examines the distribution of the machine-learning target variable (secured_driver) while understanding GoGoX’s historical fulfillment performance.

Unmatched orders account for 35.94%, revealing a substantial fulfilment gap and sufficient examples for model training. The moderate class imbalance supports using stratified sampling to preserve the outcome proportions across training, validation and testing datasets.

EDA 2: Matching Difficulty and Demand by Vehicle Type

Examines whether requested vehicle type is associated with driver-matching outcomes by comparing unmatched rate (severity) and demand share (scale) across vehicle categories.

Lorry24 records the highest unmatched rate (83.85%) but minimal demand (0.53%), while Van dominates demand (59.62%) despite having the lowest unmatched rate (21.76%).

These substantial differences suggest that requested vehicle type is a potentially informative predictor of matching failures. Furthermore, the contrast between severity and scale supports prioritising order profiles by their contribution to the overall fulfilment gap, rather than unmatched rate alone.

EDA 3: Driver Matching Outcome by Booking Lead Time

Examines the relationship between booking lead time and driver-matching outcomes by comparing matched and unmatched rates across lead-time bands.

As booking lead time increases, the unmatched rate consistently decreases, suggesting that advance bookings are associated with fewer matching difficulties.

This supports including booking lead time as a potential predictor in Random Forest, while using lead-time bands to identify historically high-risk order profiles for subsequent prioritisation.

EDA 4: Unmatched Rate by Pickup Day and Period

Examines how pickup day and period jointly relate to driver-matching difficulty, with darker cells indicating higher Unmatched rates.

Overnight pickups consistently record the highest unmatched rates (58.72% to 72.71%). This suggests that matching difficulties vary substantially across pickup periods, with additional differences between days.

These findings support including pickup day and period as potential Random Forest predictors, allowing the model to capture temporal interactions associated with matching risk.

EDA 5: Pickup Zone Demand vs Unmatched Rate

Examines the relationship between pickup-zone demand and unmatched rates to identify geographical variations in driver-matching difficulty.

Only zones with ≥30 orders were included to reduce volatility from very small sample sizes and improve the reliability of comparisons. Unmatched rates vary considerably across zones, with no consistent increase as demand rises. 

This suggests that demand volume alone may not explain geographical matching difficulties, supporting the inclusion of regional features in Random Forest alongside other order characteristics.

EDA 7: Unmatched Rate by Waypoint Count

Examines whether delivery complexity, represented by waypoint count, is associated with driver-matching difficulty.

Unmatched rates fluctuate between 29.89% and 36.18% across waypoint bands, without a consistent trend. This suggests that additional waypoints are not systematically associated with greater matching difficulty.

Given the limited descriptive evidence of a clear relationship, waypoint count was not selected as a Random Forest predictor, allowing the model to focus on other relevant order characteristics.

EDA 6: Dropoff Zone Demand vs Unmatched Rate

Examines the relationship between dropoff-zone demand and unmatched rates to identify geographical differences in driver-matching difficulty.

Unmatched rates vary considerably among lower-demand zones, while higher-demand zones do not consistently experience greater matching difficulties. This suggests that demand volume alone may not explain destination-related fulfilment gaps.

These findings support including dropoff region as a potential Random Forest predictor, allowing the model to examine geographical patterns alongside booking urgency, vehicle requirements and pickup characteristics.

Model Test and Training

Model Development and Data Partitioning

The dataset was divided into 60% training, 20% validation and 20% testing using two Table Partitioners.

• Training (60%): Trains the Random Forest using 200 Gini-based decision trees to identify matching-risk patterns.

• Validation (20%): Evaluates predictions and compares classification thresholds to select a suitable balance of Recall and Precision.

• Testing (20%): Independently evaluates the selected model and threshold on unseen orders.

Stratified sampling preserves the proportion of Matched and Unmatched orders, while fixed random seeds ensure reproducibility.

Classification Threshold Selection

The default threshold achieved low unmatched recall (40.8%) and F1-score (51.8%), indicating that many actual driver-matching failures were not detected.

Hence, alternative thresholds were manually evaluated to balance failure detection (recall) and unnecessary interventions (precision).

Threshold 0.21 was selected, achieving 63.3% recall, 62.1% precision and 62.7% F1-score. This allowed for a better balance failure detection against false alarms for GoGoX.

ROC-AUC was assessed separately to evaluate the model's overall ability to distinguish matched from unmatched orders, independent of the selected classification threshold.