{
  "id": 704284,
  "title": "WiDS Global Datathon 2026",
  "content": "<p>My solution framed the challenge as a censor-aware survival prediction problem, building calibrated and monotonic wildfire threat probabilities for 12h, 24h, 48h, and 72h from only the first five hours of fire activity.</p>\n<p>The final winning submission, <code>submission_domain_anchor_blend_95_5.csv</code>, used a conservative 95/5 blend of two survival-superlearner branches: a domain-specialist survival stack and an initial mainline survival stack. After blending, I enforced row-wise monotonicity so that each event’s predicted threat probability increased consistently across time horizons.</p>\n<p>This approach prioritized stable generalization on a very small right-censored dataset, focusing on survival-aware modeling, calibration at the metric-heavy mid horizons, robust tabular ensembling, and strict probability validity.</p>\n<p>Reproducibility materials are available here:</p>\n<p>Kaggle Model:<br>\n<a href=\"https://www.kaggle.com/models/luciayen/wids-2026-winning-survival-blend/ScikitLearn/winning-submission-package/3\" target=\"_blank\">https://www.kaggle.com/models/luciayen/wids-2026-winning-survival-blend/ScikitLearn/winning-submission-package/3</a></p>\n<p>Kaggle Reproducibility Notebook:<br>\n<a href=\"https://www.kaggle.com/code/luciayen/wids-2026-winning-survival-blend-reproducibility\" target=\"_blank\">https://www.kaggle.com/code/luciayen/wids-2026-winning-survival-blend-reproducibility</a></p>\n<p>GitHub Repository:<br>\n<a href=\"https://github.com/Yen-hub/Predicting-Time-to-Threat-for-Evacuation-Zones-Using-Survival-Analysis/tree/codex/winner-release-package\" target=\"_blank\">https://github.com/Yen-hub/Predicting-Time-to-Threat-for-Evacuation-Zones-Using-Survival-Analysis/tree/codex/winner-release-package</a></p>",
  "comments": [],
  "comment_count": 0,
  "messages": [
    {
      "id": 3466438,
      "postDate": "2026-06-04T04:48:20.447Z",
      "content": "<p>My solution framed the challenge as a censor-aware survival prediction problem, building calibrated and monotonic wildfire threat probabilities for 12h, 24h, 48h, and 72h from only the first five hours of fire activity.</p>\n<p>The final winning submission, <code>submission_domain_anchor_blend_95_5.csv</code>, used a conservative 95/5 blend of two survival-superlearner branches: a domain-specialist survival stack and an initial mainline survival stack. After blending, I enforced row-wise monotonicity so that each event’s predicted threat probability increased consistently across time horizons.</p>\n<p>This approach prioritized stable generalization on a very small right-censored dataset, focusing on survival-aware modeling, calibration at the metric-heavy mid horizons, robust tabular ensembling, and strict probability validity.</p>\n<p>Reproducibility materials are available here:</p>\n<p>Kaggle Model:<br>\n<a href=\"https://www.kaggle.com/models/luciayen/wids-2026-winning-survival-blend/ScikitLearn/winning-submission-package/3\" target=\"_blank\">https://www.kaggle.com/models/luciayen/wids-2026-winning-survival-blend/ScikitLearn/winning-submission-package/3</a></p>\n<p>Kaggle Reproducibility Notebook:<br>\n<a href=\"https://www.kaggle.com/code/luciayen/wids-2026-winning-survival-blend-reproducibility\" target=\"_blank\">https://www.kaggle.com/code/luciayen/wids-2026-winning-survival-blend-reproducibility</a></p>\n<p>GitHub Repository:<br>\n<a href=\"https://github.com/Yen-hub/Predicting-Time-to-Threat-for-Evacuation-Zones-Using-Survival-Analysis/tree/codex/winner-release-package\" target=\"_blank\">https://github.com/Yen-hub/Predicting-Time-to-Threat-for-Evacuation-Zones-Using-Survival-Analysis/tree/codex/winner-release-package</a></p>",
      "rawMarkdown": "My solution framed the challenge as a censor-aware survival prediction problem, building calibrated and monotonic wildfire threat probabilities for 12h, 24h, 48h, and 72h from only the first five hours of fire activity.\n\nThe final winning submission, `submission_domain_anchor_blend_95_5.csv`, used a conservative 95/5 blend of two survival-superlearner branches: a domain-specialist survival stack and an initial mainline survival stack. After blending, I enforced row-wise monotonicity so that each event’s predicted threat probability increased consistently across time horizons.\n\nThis approach prioritized stable generalization on a very small right-censored dataset, focusing on survival-aware modeling, calibration at the metric-heavy mid horizons, robust tabular ensembling, and strict probability validity.\n\nReproducibility materials are available here:\n\nKaggle Model:  \nhttps://www.kaggle.com/models/luciayen/wids-2026-winning-survival-blend/ScikitLearn/winning-submission-package/3\n\nKaggle Reproducibility Notebook:  \nhttps://www.kaggle.com/code/luciayen/wids-2026-winning-survival-blend-reproducibility\n\nGitHub Repository:  \nhttps://github.com/Yen-hub/Predicting-Time-to-Threat-for-Evacuation-Zones-Using-Survival-Analysis/tree/codex/winner-release-package"
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-20T08:49:10Z",
  "competition_id": 125681,
  "competition_slug": "WiDSWorldWide_GlobalDathon26",
  "team_id": 15420683,
  "solution_url": "/competitions/WiDSWorldWide_GlobalDathon26/writeups/wids-global-datathon-2026"
}