{
  "id": 453888,
  "title": "7th place solution ",
  "content": "<p>Thank you for organizing this competition.<br>\nHere is an overview of my approach to this competition.</p>\n<p><strong>FEATURUNG</strong><br>\n-Created two original features from bmi, age, heart_disease, hypertension, and smoking_status.</p>\n<p><strong>MODEL</strong><br>\n-Use XGB, CatBoost, LGBM and their ensembles (basic policy)<br>\n-Optimization of each model was performed using optuna <br>\n-Since the CatBoost single model tended to show a better metric score than the ensemble, optimization was performed using only the CatBoost single model at the final stage (resulted in overfitting) (failure) <br>\n-Model metrics not used ROC but mainly RMSE (without checking competition rules) (failure)</p>\n<p><strong>Retrospection</strong><br>\n-A lot of original featuring shoukd be prepared in the first half of the competition <br>\n-Metrics should have been aligned with competition evaluation metrics from the beginning <br>\n-There may be a difference between regressor.predict and classifier.predict_proba, I should have tried both <br>\n-Overall, I could not predict that the ensemble would give better results in private than the single model <br>\n-Therefore, I should select not simply submissions of higher public socre, but potentially best one manually.</p>\n<p>My public/private score results shown in <a href=\"https://www.kaggle.com/code/stpeteishii/stroke-my-public-private-score-comparison\" target=\"_blank\">this notebook</a>.</p>",
  "comments": [],
  "comment_count": 0,
  "messages": [
    {
      "id": 2516980,
      "postDate": "2023-11-08T05:37:07.943Z",
      "content": "<p>Thank you for organizing this competition.<br>\nHere is an overview of my approach to this competition.</p>\n<p><strong>FEATURUNG</strong><br>\n-Created two original features from bmi, age, heart_disease, hypertension, and smoking_status.</p>\n<p><strong>MODEL</strong><br>\n-Use XGB, CatBoost, LGBM and their ensembles (basic policy)<br>\n-Optimization of each model was performed using optuna <br>\n-Since the CatBoost single model tended to show a better metric score than the ensemble, optimization was performed using only the CatBoost single model at the final stage (resulted in overfitting) (failure) <br>\n-Model metrics not used ROC but mainly RMSE (without checking competition rules) (failure)</p>\n<p><strong>Retrospection</strong><br>\n-A lot of original featuring shoukd be prepared in the first half of the competition <br>\n-Metrics should have been aligned with competition evaluation metrics from the beginning <br>\n-There may be a difference between regressor.predict and classifier.predict_proba, I should have tried both <br>\n-Overall, I could not predict that the ensemble would give better results in private than the single model <br>\n-Therefore, I should select not simply submissions of higher public socre, but potentially best one manually.</p>\n<p>My public/private score results shown in <a href=\"https://www.kaggle.com/code/stpeteishii/stroke-my-public-private-score-comparison\" target=\"_blank\">this notebook</a>.</p>",
      "rawMarkdown": "Thank you for organizing this competition.\nHere is an overview of my approach to this competition.\n\n**FEATURUNG**\n-Created two original features from bmi, age, heart_disease, hypertension, and smoking_status.\n\n**MODEL**\n-Use XGB, CatBoost, LGBM and their ensembles (basic policy)\n-Optimization of each model was performed using optuna \n-Since the CatBoost single model tended to show a better metric score than the ensemble, optimization was performed using only the CatBoost single model at the final stage (resulted in overfitting) (failure) \n-Model metrics not used ROC but mainly RMSE (without checking competition rules) (failure)\n\n**Retrospection**\n-A lot of original featuring shoukd be prepared in the first half of the competition \n-Metrics should have been aligned with competition evaluation metrics from the beginning \n-There may be a difference between regressor.predict and classifier.predict_proba, I should have tried both \n-Overall, I could not predict that the ensemble would give better results in private than the single model \n-Therefore, I should select not simply submissions of higher public socre, but potentially best one manually.\n\nMy public/private score results shown in [this notebook](https://www.kaggle.com/code/stpeteishii/stroke-my-public-private-score-comparison).\n",
      "votes": 2
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-20T01:16:23Z",
  "competition_id": 63309,
  "competition_slug": "dudss-2023-stroke-prediction",
  "team_id": 11168182,
  "solution_url": "/competitions/dudss-2023-stroke-prediction/writeups/stpete-ishii-7th-place-solution"
}