First of all, my teammate @siddhant1104 and I would like to thank the organizers for hosting this competition and congratulate all the winners. It was such a hardship to decide focus on improving whether the LB score or the CV score because when the CV score was increased, our LB scores mostly got worse. Our solution is weighted ensemble of two separate approaches.
\ndri_score and conditioning_intensity (tuned with optuna)donor_age and age_at_hct features to extract somesin_year andcos_year)dri_score_mapping = {\n : 0.6850752146154907,\n : 0.18589473149703015,\n : 0.5683465067841215,\n : 0.7708720693082163,\n : 0.9586424711654987,\n : 0.6831791561653417,\n : 0.7166435651957048,\n : 0.8821201547093761,\n : 0.49866306284678735,\n : 0.9411056819278409,\n : 0.1890854786067684,\n : 0.5377767827330516\n}\n\nconditioning_intensity_mapping = {\n : 0.6026915942898587,\n : 0.02153075067313332,\n : 0.8995437792670338,\n : 0.9211477712757186,\n : 0.7388173559148422,\n : 0.22412092165882558,\n : 0.32745500022610163\n}\nsin_year andcos_year)Congratulations for having gold medal @siddhant1104 and @sercanyesiloz 🥳
", "votes": 2, "replies": [ { "id": 3143532, "author_name": "Sercan Yeşilöz", "author_url": "", "post_date": "2025-03-07T10:06:19.820000", "content": "Thank you!
", "votes": 0, "replies": [] } ] }, { "id": 3144105, "author_name": "Robert Hatch", "author_url": "", "post_date": "2025-03-08T00:13:07.293000", "content": "\n\nIt was such a hardship to decide focus on improving whether the LB score or the CV score because when the CV score was increased, our LB scores mostly got worse
\n
What ended up working and getting you 11th? I'm guessing from my own experience that the higher CV submission is what ended up doing really well?
", "votes": 0, "replies": [] }, { "id": 3142460, "author_name": "Andreas Bisiadis", "author_url": "", "post_date": "2025-03-06T12:11:03.697000", "content": "@sercanyesiloz How did you select the race group sample weights?
\n raceweight = {: / ,\n : / ,\n : / ,\n : / ,\n : / ,\n : / \n }\nThese weights are actually the average of the two race group weights that received the highest LB scores.
", "votes": 0, "replies": [] } ] } ], "comment_count": 5, "messages": [ { "id": 3142281, "postDate": "2025-03-06T09:08:02.037Z", "content": "First of all, my teammate @siddhant1104 and I would like to thank the organizers for hosting this competition and congratulate all the winners. It was such a hardship to decide focus on improving whether the LB score or the CV score because when the CV score was increased, our LB scores mostly got worse. Our solution is weighted ensemble of two separate approaches.
\ndri_score and conditioning_intensity (tuned with optuna)donor_age and age_at_hct features to extract somesin_year andcos_year)dri_score_mapping = {\n : 0.6850752146154907,\n : 0.18589473149703015,\n : 0.5683465067841215,\n : 0.7708720693082163,\n : 0.9586424711654987,\n : 0.6831791561653417,\n : 0.7166435651957048,\n : 0.8821201547093761,\n : 0.49866306284678735,\n : 0.9411056819278409,\n : 0.1890854786067684,\n : 0.5377767827330516\n}\n\nconditioning_intensity_mapping = {\n : 0.6026915942898587,\n : 0.02153075067313332,\n : 0.8995437792670338,\n : 0.9211477712757186,\n : 0.7388173559148422,\n : 0.22412092165882558,\n : 0.32745500022610163\n}\nsin_year andcos_year)dri_score and conditioning_intensity (tuned with optuna)\n- Crossing donor_age and age_at_hct features to extract some\n- Cyclic year features (sin_year andcos_year)\n- Kaplan Meier, Nelson Aalen and Cox target transformations\n- 4 LightGBM and 6 CatBoost models\n- GBDT Cross-Validation Setup -> Stratified 10 Folds by **race groups**\n- Neural Networks Cross-Validation Setup -> Stratified 5 Folds by **race groups** and **age_at_hct==0.44**\n- Predictions were scaled before blending\n\n### Ordinal Encoding\n```\ndri_score_mapping = {\n \"High\": 0.6850752146154907,\n \"High - TED AML case cos_year)\n- 3 custom target transformations using Kaplan Meier and Nelson Aalen methods\n- Classification and Regression models\n- Models trained with entire train set, didn't use any cross-validation method\n- Predictions were scaled before blending",
"votes": 19
},
{
"id": 3143256,
"postDate": "2025-03-07T03:55:00.413Z",
"content": "Congratulations for having gold medal @siddhant1104 and @sercanyesiloz 🥳
", "rawMarkdown": "Congratulations for having gold medal @siddhant1104 and @sercanyesiloz 🥳", "votes": 2, "replies": [ { "id": 3143532, "postDate": "2025-03-07T10:06:19.820Z", "content": "Thank you!
", "rawMarkdown": "Thank you!" } ] }, { "id": 3144105, "postDate": "2025-03-08T00:13:07.293Z", "content": "\n\nIt was such a hardship to decide focus on improving whether the LB score or the CV score because when the CV score was increased, our LB scores mostly got worse
\n
What ended up working and getting you 11th? I'm guessing from my own experience that the higher CV submission is what ended up doing really well?
", "rawMarkdown": "> It was such a hardship to decide focus on improving whether the LB score or the CV score because when the CV score was increased, our LB scores mostly got worse\n\nWhat ended up working and getting you 11th? I'm guessing from my own experience that the higher CV submission is what ended up doing really well?" }, { "id": 3142460, "postDate": "2025-03-06T12:11:03.697Z", "content": "@sercanyesiloz How did you select the race group sample weights?
\n raceweight = {: / ,\n : / ,\n : / ,\n : / ,\n : / ,\n : / \n }\nThese weights are actually the average of the two race group weights that received the highest LB scores.
", "rawMarkdown": "These weights are actually the average of the two race group weights that received the highest LB scores." } ] } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T19:47:53Z", "competition_id": 70942, "competition_slug": "equity-post-HCT-survival-predictions", "team_id": 13063402, "solution_url": "/competitions/equity-post-HCT-survival-predictions/writeups/sercan-siddhant-11th-place-solution" }