| Attribute | \nDetails | \n
|---|---|
| Competition | \nWiDS Global Datathon 2026 | \n
| Team Name | \nBob the Builder(s) | \n
| Public Score | \n0.97707 | \n
| Private Score | \n0.96681 | \n
| Private Rank | \n19 | \n
| Location | \nDhaka | \n
I'm the sole member of team Bob the Builder(s). I'm a Sophomore at Bangladesh University of Engineering and Technology majoring in Computer Science and Engineering. And I like playing League of Legends : D !\nThis is my first Machine Learning competition and before this, I had no idea about Machine Learning or Deep Learning. I had to grind hard throughout the competition. My main aim was to use this as a stepping stone into the realm of ML and DL. I spent about a month or so on this competition.
\nWe built a zone-aware survival ensemble that predicts cumulative wildfire hit probabilities across 12, 24, 48, and 72-hour horizons. The core idea is a Zone-Based Routing System — fires are routed into Near (< 5 km) and Far (≥ 5 km) zones, with independent model weight arrays optimized per zone via an exhaustive OOF grid search.\nThe Level-0 ensemble combines six model families — Gradient Boosting Survival Analysis (GBSA), XGBoost AFT, Cox Proportional Hazards, Random Survival Forests, a LightGBM Timing model, and a Two-Stage Calibrator — all trained inside a 7-fold Stratified K-Fold CV loop with IPCW-weighted targets. A pre-blending step merges AFT into GBSA (82/18 split) before zone-based weight optimization. The entire pipeline (training + inference) runs in approximately 18 minutes on Kaggle's CPU kernel using Python with scikit-survival, XGBoost, LightGBM, and scikit-learn.
\nThe raw dataset contains 35 columns. Through engineering we expanded this to ~90+ features organized into several families:
\nFires behave fundamentally differently when they are at the doorstep versus miles away. We hard-coded structural zone boundaries — Near (< 5 km), Warning (5–10 km), Far (≥ 10 km), and Very Far (≥ 15 km) — and computed Relative Percentile Ranks within each zone. For example, near_speed_rank evaluates a fire's effective closing speed strictly against the historical distribution of other near-zone fires.
log(1 + area_km²) × effective_closing / (dist_km + 0.1).Different model families used curated feature subsets rather than a single universal set:
\nThe interaction between distance zone and effective closing speed was critical. A fast-closing fire in the near zone is an immediate threat; the same speed in the far zone is far less concerning. This motivated the zone-based weighting system rather than a single global model. \nNo external data was used.
\nAll models were trained inside a 7-fold Stratified K-Fold CV loop. Stratification used a composite label (event × 2 + near_zone_flag) to ensure balanced representation of close-proximity threats. Targets were IPCW-weighted binary labels via Kaplan-Meier censoring adjustment (weights clipped at 30.0).
\nsurvival:aft with sigma tuned across [1.0, 1.3, 1.6]. Predictions are pre-blended into GBSA (82% GBSA + 18% AFT) before any downstream ensembling.Rather than averaging predictions globally, independent 5-weight arrays for Near and Far zones were found via exhaustive combinatorial OOF grid search (values [0.0, 0.1, …, 0.6] per model slot) optimizing the Hybrid metric.\nFinal optimized weights:
\n| Model Slot | \nNear Zone | \nFar Zone | \n
|---|---|---|
| GBSA+AFT | \n0.5 | \n0.0 | \n
| CoxPH | \n0.5 | \n0.0 | \n
| RSF | \n0.1 | \n0.0 | \n
| Calibrator | \n0.6 | \n0.1 | \n
| Timing | \n0.15 | \n0.0 | \n
| Near zones leaned on GBSA, Cox, and Calibrator diversity; the far zone was almost entirely Calibrator-driven. | \n\n | \n |
A per-horizon LightGBM Classifier was trained on OOF base predictions alongside non-linear distance transforms. However, OOF tuning found the optimal meta weight to be 0.0 — the zone-based blend alone proved optimal and the meta-stacker did not contribute to the final predictions.
\nRaw predictions were subjected to a 7-step post-processing sequence:
\nprob_24h = 0.90 × prob_24h + 0.10 × prob_12h and prob_48h = 0.85 × prob_48h + 0.15 × prob_24h. The 72h horizon was not smoothed.np.maximum.accumulate) across the horizon axis ensured probabilities never decrease over time.The most important trick was the Zone-Based Routing System — splitting predictions into Near and Far zones with independently optimized weights. This avoided the compromise of a one-size-fits-all global blend and let specialized models dominate in their strongest regimes.\nThree things we believe set us apart:
\nFor a simpler alternative, the most important model is GBSA with Cox Partial Hazard loss. A single GBSA model trained on just ~8 key raw features could achieve roughly 99% of the full pipeline's performance:
\ndist_min_ci_0_5h (distance to target)closing_speed_abs_m_per_hradial_growth_rate_m_per_halignment_absarea_first_haarea_growth_rate_ha_per_hprojected_advance_mcentroid_speed_m_per_hAll times measured on Kaggle CPU kernel (no GPU).
\n| Stage | \nTime | \n
|---|---|
| Full pipeline (training + inference) | \n~18 minutes | \n
| Attribute | \nDetails | \n
|---|---|
| Competition | \nWiDS Global Datathon 2026 | \n
| Team Name | \nBob the Builder(s) | \n
| Public Score | \n0.97707 | \n
| Private Score | \n0.96681 | \n
| Private Rank | \n19 | \n
| Location | \nDhaka | \n
I'm the sole member of team Bob the Builder(s). I'm a Sophomore at Bangladesh University of Engineering and Technology majoring in Computer Science and Engineering. And I like playing League of Legends : D !\nThis is my first Machine Learning competition and before this, I had no idea about Machine Learning or Deep Learning. I had to grind hard throughout the competition. My main aim was to use this as a stepping stone into the realm of ML and DL. I spent about a month or so on this competition.
\nWe built a zone-aware survival ensemble that predicts cumulative wildfire hit probabilities across 12, 24, 48, and 72-hour horizons. The core idea is a Zone-Based Routing System — fires are routed into Near (< 5 km) and Far (≥ 5 km) zones, with independent model weight arrays optimized per zone via an exhaustive OOF grid search.\nThe Level-0 ensemble combines six model families — Gradient Boosting Survival Analysis (GBSA), XGBoost AFT, Cox Proportional Hazards, Random Survival Forests, a LightGBM Timing model, and a Two-Stage Calibrator — all trained inside a 7-fold Stratified K-Fold CV loop with IPCW-weighted targets. A pre-blending step merges AFT into GBSA (82/18 split) before zone-based weight optimization. The entire pipeline (training + inference) runs in approximately 18 minutes on Kaggle's CPU kernel using Python with scikit-survival, XGBoost, LightGBM, and scikit-learn.
\nThe raw dataset contains 35 columns. Through engineering we expanded this to ~90+ features organized into several families:
\nFires behave fundamentally differently when they are at the doorstep versus miles away. We hard-coded structural zone boundaries — Near (< 5 km), Warning (5–10 km), Far (≥ 10 km), and Very Far (≥ 15 km) — and computed Relative Percentile Ranks within each zone. For example, near_speed_rank evaluates a fire's effective closing speed strictly against the historical distribution of other near-zone fires.
log(1 + area_km²) × effective_closing / (dist_km + 0.1).Different model families used curated feature subsets rather than a single universal set:
\nThe interaction between distance zone and effective closing speed was critical. A fast-closing fire in the near zone is an immediate threat; the same speed in the far zone is far less concerning. This motivated the zone-based weighting system rather than a single global model. \nNo external data was used.
\nAll models were trained inside a 7-fold Stratified K-Fold CV loop. Stratification used a composite label (event × 2 + near_zone_flag) to ensure balanced representation of close-proximity threats. Targets were IPCW-weighted binary labels via Kaplan-Meier censoring adjustment (weights clipped at 30.0).
\nsurvival:aft with sigma tuned across [1.0, 1.3, 1.6]. Predictions are pre-blended into GBSA (82% GBSA + 18% AFT) before any downstream ensembling.Rather than averaging predictions globally, independent 5-weight arrays for Near and Far zones were found via exhaustive combinatorial OOF grid search (values [0.0, 0.1, …, 0.6] per model slot) optimizing the Hybrid metric.\nFinal optimized weights:
\n| Model Slot | \nNear Zone | \nFar Zone | \n
|---|---|---|
| GBSA+AFT | \n0.5 | \n0.0 | \n
| CoxPH | \n0.5 | \n0.0 | \n
| RSF | \n0.1 | \n0.0 | \n
| Calibrator | \n0.6 | \n0.1 | \n
| Timing | \n0.15 | \n0.0 | \n
| Near zones leaned on GBSA, Cox, and Calibrator diversity; the far zone was almost entirely Calibrator-driven. | \n\n | \n |
A per-horizon LightGBM Classifier was trained on OOF base predictions alongside non-linear distance transforms. However, OOF tuning found the optimal meta weight to be 0.0 — the zone-based blend alone proved optimal and the meta-stacker did not contribute to the final predictions.
\nRaw predictions were subjected to a 7-step post-processing sequence:
\nprob_24h = 0.90 × prob_24h + 0.10 × prob_12h and prob_48h = 0.85 × prob_48h + 0.15 × prob_24h. The 72h horizon was not smoothed.np.maximum.accumulate) across the horizon axis ensured probabilities never decrease over time.The most important trick was the Zone-Based Routing System — splitting predictions into Near and Far zones with independently optimized weights. This avoided the compromise of a one-size-fits-all global blend and let specialized models dominate in their strongest regimes.\nThree things we believe set us apart:
\nFor a simpler alternative, the most important model is GBSA with Cox Partial Hazard loss. A single GBSA model trained on just ~8 key raw features could achieve roughly 99% of the full pipeline's performance:
\ndist_min_ci_0_5h (distance to target)closing_speed_abs_m_per_hradial_growth_rate_m_per_halignment_absarea_first_haarea_growth_rate_ha_per_hprojected_advance_mcentroid_speed_m_per_hAll times measured on Kaggle CPU kernel (no GPU).
\n| Stage | \nTime | \n
|---|---|
| Full pipeline (training + inference) | \n~18 minutes | \n
| Attribute | \nDetails | \n
|---|---|
| Competition | \nWiDS Global Datathon 2026 | \n
| Team Name | \nBob the Builder(s) | \n
| Public Score | \n0.97707 | \n
| Private Score | \n0.96681 | \n
| Private Rank | \n19 | \n
| Location | \nDhaka | \n