\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1630287%2Fd60afc32dfc0a18e8be5181338bc77a8%2Fsol.png?generation=1741271198455281&alt=media\")
Hey everyone,
\n
I've made public the notebook that earned me 6th place:
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-two-step-model
Here is the summary of what it does
\nI can publish code for individual parts of the solution as well if anyone is interested.
", "comments": [ { "id": 3143281, "author_name": "Oleg Khudyakov", "author_url": "", "post_date": "2025-03-07T04:31:48.787000", "content": "Hello, Alexandr!
\nCongratulations with gold medal!
\nI would be very grate full if You share full solution for all parts.
Hi, thanks!
\nHere you go
\nHyperparameter search for Catboost. For XGBoost and LightGBM, only the parameter names/ranges are different.
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-catboost-classifier-parameter-search
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-catboost-regressor-parameter-search
Finding weights for gradient boosting models
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-gradient-boosting-weights
Preparing OOF dataset for the linear model
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-oof-dataset
Training the linear model
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-second-step-model-training
Congrats on gold!
\nFor those keeping track at home, at least top 7 all did this model efs and model efs==1 efs_time rank separately trick.
\nI suspect it was only so critical due to synthetic data. But maybe the nature of the metric alone makes it critical. Reduce 2 dimensional label down to one dimension prediction… but then calculate score against the original two dimensions(!)
", "votes": 1, "replies": [ { "id": 3142971, "author_name": "Aleksandr Timofeev", "author_url": "", "post_date": "2025-03-06T19:45:33.913000", "content": "Thanks!
\nNo clue why this is so crucial 🙂
I suspect the synthetic data generation approach is the only reason. But not confident about it, just hypothesis.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3143740, "author_name": "Nicolae Cadin", "author_url": "", "post_date": "2025-03-07T14:41:16.670000", "content": "Congratulations on the gold! Thank you for sharing your well-written notebook.
", "votes": 0, "replies": [] } ], "comment_count": 6, "messages": [ { "id": 3142613, "postDate": "2025-03-06T14:34:33.633Z", "content": "Hey everyone,
\n
I've made public the notebook that earned me 6th place:
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-two-step-model
Here is the summary of what it does
\nI can publish code for individual parts of the solution as well if anyone is interested.
", "rawMarkdown": "Hey everyone,\n\n\nI've made public the notebook that earned me 6th place:\nhttps://www.kaggle.com/code/myprofileurl/6th-place-two-step-model\n\nHere is the summary of what it does\n\n### 1. Gradient Boosting Models \n- Use gradient boosting models to predict:\n - **p(efs == 1)**: the probability that the event occurs.\n - **E[efs_time | efs == 1]**: the expected survival time conditioned on the event occurring.\n\n### 2. Neural Network \n- Train a TabM neural network that optimizes a smoothed version of the evaluation metric (using a sigmoid instead of the indicator function).\n\n### 3. Combine the Predictions \n- Run 20-fold cross-validation with 5 random seeds for all models.\n- Train a one-layer neural network (with no non-linearities) on out-of-fold (OOF) data, with the same loss function as the TabM model. \n- **Inputs to the NN:**\n - **Risk score from the TabM model**\n - **2nd-order polynomial combinations** of **p(efs == 1)** and **E[efs_time | efs == 1]**\n\nI can publish code for individual parts of the solution as well if anyone is interested.\n", "votes": 25 }, { "id": 3143281, "postDate": "2025-03-07T04:31:48.787Z", "content": "Hello, Alexandr!
\nCongratulations with gold medal!
\nI would be very grate full if You share full solution for all parts.
Hi, thanks!
\nHere you go
\nHyperparameter search for Catboost. For XGBoost and LightGBM, only the parameter names/ranges are different.
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-catboost-classifier-parameter-search
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-catboost-regressor-parameter-search
Finding weights for gradient boosting models
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-gradient-boosting-weights
Preparing OOF dataset for the linear model
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-oof-dataset
Training the linear model
\nhttps://www.kaggle.com/code/myprofileurl/6th-place-second-step-model-training
Congrats on gold!
\nFor those keeping track at home, at least top 7 all did this model efs and model efs==1 efs_time rank separately trick.
\nI suspect it was only so critical due to synthetic data. But maybe the nature of the metric alone makes it critical. Reduce 2 dimensional label down to one dimension prediction… but then calculate score against the original two dimensions(!)
", "rawMarkdown": "Congrats on gold!\n\nFor those keeping track at home, at least top 7 all did this model efs and model efs==1 efs_time rank separately trick. \n\nI suspect it was only so critical due to synthetic data. But maybe the nature of the metric alone makes it critical. Reduce 2 dimensional label down to one dimension prediction... but then calculate score against the original two dimensions(!)", "votes": 1, "replies": [ { "id": 3142971, "postDate": "2025-03-06T19:45:33.913Z", "content": "Thanks!
\nNo clue why this is so crucial 🙂
I suspect the synthetic data generation approach is the only reason. But not confident about it, just hypothesis.
", "rawMarkdown": "I suspect the synthetic data generation approach is the only reason. But not confident about it, just hypothesis. " } ] } ] }, { "id": 3143740, "postDate": "2025-03-07T14:41:16.670Z", "content": "Congratulations on the gold! Thank you for sharing your well-written notebook.
", "rawMarkdown": "Congratulations on the gold! Thank you for sharing your well-written notebook." } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T19:48:40Z", "competition_id": 70942, "competition_slug": "equity-post-HCT-survival-predictions", "team_id": 13230408, "solution_url": "/competitions/equity-post-HCT-survival-predictions/writeups/aleksandr-timofeev-6th-place-solution" }