1.Acknowledgement
\nFirst and foremost，I'd like to start with a big respect to the following open source authors and my team:
\n1.one of our model from:https://www.kaggle.com/code/albansteff/refactoring-nn-pairwise-ranking-loss
\n2.some feats from:https://www.kaggle.com/code/yunsuxiaozi/cibmtr-yunbase
\n3.use the classifier to make post-processing：https://www.kaggle.com/code/kendontcare11/public-classifier-cat-xgb-lb-0-688
\n4.The target constructor provided by all the great open source workers.
\n5.And the three brothers(leaf,yueming,DeepMayNotLearn),who have been preparing the postgraduate entrance exams ,thank them for their hard work
\n (the above ranking is in no particular order)
2.Plan introduction
\n2.1 NN
\n2.1.1 Multi-target NN
\nA：target one： CV of CoxPHFitter
\nB：target two：custom_target to split the 0 1 as much as possible
def custom_target(self): \n cus_target = self.data[['efs','efs_time']]\n cus_target['y'] = self.data.efs_time.values\n mx = cus_target.loc[cus_target.efs==1,\"efs_time\"].max()\n mn = cus_target.loc[cus_target.efs==0,\"efs_time\"].min()\n cus_target.loc[cus_target.efs==0,\"y\"] = cus_target.loc[cus_target.efs==0,\"y\"] + mx - mn\n cus_target.y = cus_target.y.rank()\n cus_target.loc[cus_target.efs==0,\"y\"] += 2*len(cus_target)\n cus_target.y = cus_target.y / cus_target.y.max()\n cus_target.y = np.log( cus_target.y )\n cus_target.y -= cus_target.y.mean()\n cus_target.y *= -1.0 \n return cus_target\n
We trained two simple models （CNN，DNN） to fit the two targets
\n2.1.2 Opensource ODST
\nwe make some changes：
\n1.fill the unseen category with train[col].mode()[0] ,rather than np.nan
\nif ind.any():\n val.loc[ind, col] = train[col].mode()[0]
\n2.use weights（race_index of each fold) to ensemble each fold
2.1.3 Our ODST（pure torch version without torch lightning）
\n1.some additional feats from yunsuxiaozi‘s notebook
\n2.slights change the Hyperparameters
\n3.use weights to ensemble each fold best of our metric
\n4.use seed_torch(42+fold) to train the model of each fold
2.2 GBDT Tree
\n For the tree model part, we use two classification trees with different features to make post-processing.And use xgboost，lightgbm and catboost to train the common target mentioned in the public code.The features our tree used are easy to see in the public.
3.Overview of the Plan
\nFor more details：https://www.kaggle.com/code/kudosn/nn-tree-ensemble
\n![https://www.helloimg.com/i/2025/03/12/67d05e23c134a.png] (url to Plan Structure)
\nTips：If want to run this notebook please change \"from cib_metric import score\" with \"from metric import score\",we just find the UTILITY SCRIPTS has been deleted.
\n4.Eventually, Allow me to show great respect to the open source workers once again
\n![https://www.helloimg.com/i/2025/03/12/67d05e22c6492.png] (url to a surprise)
Congratulations on winning the gold medal, good luck on your postgraduate entrance exam!
", "votes": 0, "replies": [ { "id": 3147090, "author_name": "Luck is all you need", "author_url": "", "post_date": "2025-03-11T16:24:41.347000", "content": "Thanks bro ！！！
", "votes": 0, "replies": [] } ] } ], "comment_count": 2, "messages": [ { "id": 3146796, "postDate": "2025-03-11T08:56:37.473Z", "content": "1.Acknowledgement
\nFirst and foremost，I'd like to start with a big respect to the following open source authors and my team:
\n1.one of our model from:https://www.kaggle.com/code/albansteff/refactoring-nn-pairwise-ranking-loss
\n2.some feats from:https://www.kaggle.com/code/yunsuxiaozi/cibmtr-yunbase
\n3.use the classifier to make post-processing：https://www.kaggle.com/code/kendontcare11/public-classifier-cat-xgb-lb-0-688
\n4.The target constructor provided by all the great open source workers.
\n5.And the three brothers(leaf,yueming,DeepMayNotLearn),who have been preparing the postgraduate entrance exams ,thank them for their hard work
\n (the above ranking is in no particular order)
2.Plan introduction
\n2.1 NN
\n2.1.1 Multi-target NN
\nA：target one： CV of CoxPHFitter
\nB：target two：custom_target to split the 0 1 as much as possible
def custom_target(self): \n cus_target = self.data[['efs','efs_time']]\n cus_target['y'] = self.data.efs_time.values\n mx = cus_target.loc[cus_target.efs==1,\"efs_time\"].max()\n mn = cus_target.loc[cus_target.efs==0,\"efs_time\"].min()\n cus_target.loc[cus_target.efs==0,\"y\"] = cus_target.loc[cus_target.efs==0,\"y\"] + mx - mn\n cus_target.y = cus_target.y.rank()\n cus_target.loc[cus_target.efs==0,\"y\"] += 2*len(cus_target)\n cus_target.y = cus_target.y / cus_target.y.max()\n cus_target.y = np.log( cus_target.y )\n cus_target.y -= cus_target.y.mean()\n cus_target.y *= -1.0 \n return cus_target\n
We trained two simple models （CNN，DNN） to fit the two targets
\n2.1.2 Opensource ODST
\nwe make some changes：
\n1.fill the unseen category with train[col].mode()[0] ,rather than np.nan
\nif ind.any():\n val.loc[ind, col] = train[col].mode()[0]
\n2.use weights（race_index of each fold) to ensemble each fold
2.1.3 Our ODST（pure torch version without torch lightning）
\n1.some additional feats from yunsuxiaozi‘s notebook
\n2.slights change the Hyperparameters
\n3.use weights to ensemble each fold best of our metric
\n4.use seed_torch(42+fold) to train the model of each fold
2.2 GBDT Tree
\n For the tree model part, we use two classification trees with different features to make post-processing.And use xgboost，lightgbm and catboost to train the common target mentioned in the public code.The features our tree used are easy to see in the public.
3.Overview of the Plan
\nFor more details：https://www.kaggle.com/code/kudosn/nn-tree-ensemble
\n![https://www.helloimg.com/i/2025/03/12/67d05e23c134a.png] (url to Plan Structure)
\nTips：If want to run this notebook please change \"from cib_metric import score\" with \"from metric import score\",we just find the UTILITY SCRIPTS has been deleted.
\n4.Eventually, Allow me to show great respect to the open source workers once again
\n![https://www.helloimg.com/i/2025/03/12/67d05e22c6492.png] (url to a surprise)
Congratulations on winning the gold medal, good luck on your postgraduate entrance exam!
", "rawMarkdown": "Congratulations on winning the gold medal, good luck on your postgraduate entrance exam!", "replies": [ { "id": 3147090, "postDate": "2025-03-11T16:24:41.347Z", "content": "Thanks bro ！！！
", "rawMarkdown": "Thanks bro ！！！" } ] } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T19:50:26Z", "competition_id": 70942, "competition_slug": "equity-post-HCT-survival-predictions", "team_id": 13089230, "solution_url": "/competitions/equity-post-HCT-survival-predictions/writeups/last-dance-to-our-university-16-th-solution" }