First of all, Thank you so much @usharengaraju for hosting this amazing competition & to all the fellow Kagglers for sharing their knowledge via notebooks & discussions. I was able to learn a lot from them. I have participated in a few competitions, but this is the first time I have placed so high on the leaderboard. Also, this is my first post so Kindly upvote the post if you find it helpful.
\nI have started off with exploratory data analysis, where I found out that the data does not have null values.
The contry_of_res variable had 61 countries and some countries only had a few data records, so I created a new feature continent which maps the respective countries to their continents and dropped the contry_of_res variable.
Variable age_desc had only one category, so I removed that variable.
Variables relation and ethnicity had a category ?. I have renamed that category as unknown.
Variables relation and ethnicity also had some categories with fewer data records, so I have binned them into the Others category.
I have trained all the models using 10 fold stratified cross-validation. Normalized the numerical variables using StandardScaler, ordinal encoded the categorical variables using OrdinalEncoder
\nBaseline model - Scikit-learn's Logistic Regression model with default parameters. Public LB: 0.82655, Private LB: 0.98051. This model would still have been the 8th place solution.
XGBoost model with hyperparameter tuned using Optuna. I have used scale_pos_weight as the data was imbalanced. Public LB: 0.93939, Private LB: 0.98268. This was my best submitted solution.
counter = Counter(Y_train)\nscale_pos_weight = counter.(0) / counter.(1)\n\nXGB_Params = {\n : 201, : 15,\n : 0.0038, : 0.7742,\n : 6.654, : 0.3,\n : 0.5, : 7\n}\n\nmodel = XGBClassifier(\n =,\n =scale_pos_weight,\n =,\n =,\n =,\n =,\n =42,\n =-1,\n **XGB_Params\n)\nis_unbalance as the data was imbalanced. Public LB: 0.92477, Private LB: 0.96103. LGBM_Params = {\n : 381, : 0.0451, : 2680,\n : 7, : 200, : 5,\n : 40, : 0.5696,\n : 0.8, : 1, : 0.25\n}\n\nmodel = LGBMClassifier(\n =,\n =,\n =,\n =42,\n =-1,\n **LGBM_Params\n)\nclass_weight as the data was imbalanced. Public LB: 0.93791, Private LB: 0.98268. This score was exactly same as my best submitted XGBoost model.counter = Counter(Y_train)\n\nparam_grid = {\n : 958, : ,\n : 3\n}\n\nclass_weight = {\n 0: n_samples / (counter.(0) * len(counter)),\n 1: n_samples / (counter.(1) * len(counter))\n}\n\nmodel = RandomForestClassifier(\n =class_weight,\n =0,\n =-1,\n **param_grid\n)\nis_unbalance as the data was imbalanced. Public LB: 0.93748 Private LB: 0.97835. counter = Counter(Y_train)\nscale_pos_weight = counter.(0) / counter.(1)\n\nCatBoost_Params = {\n : 186, : 0.00032,\n : 0.0801, : 1,\n : , : 4.3617\n}\n\nmodel = CatBoostClassifier(\n *tBoost_Params,\n =scale_pos_weight,\n =,\n =,\n =42,\n =0\n)\nHope this was helpful.
", "comments": [ { "id": 1718317, "author_name": "dineshrajan", "author_url": "", "post_date": "2022-03-10T18:08:14.577000", "content": "Good job !
", "votes": 1, "replies": [ { "id": 1718341, "author_name": "Kishore S", "author_url": "", "post_date": "2022-03-10T18:30:49.860000", "content": "Thank you @dineshrajan.
", "votes": 0, "replies": [] } ] }, { "id": 2012905, "author_name": "Kishore S", "author_url": "", "post_date": "2022-11-01T14:15:43.543000", "content": "Link to my solution notebooks: https://www.kaggle.com/competitions/autism-prediction/discussion/312178
", "votes": 0, "replies": [] }, { "id": 1718338, "author_name": "Kishore S", "author_url": "", "post_date": "2022-03-10T18:29:27.687000", "content": "I have made my 8th place solution and other high-scoring notebooks public. You can find them in the competition codes section.
", "votes": 0, "replies": [] } ], "comment_count": 4, "messages": [ { "id": 1718236, "postDate": "2022-03-10T16:23:46.843Z", "content": "First of all, Thank you so much @usharengaraju for hosting this amazing competition & to all the fellow Kagglers for sharing their knowledge via notebooks & discussions. I was able to learn a lot from them. I have participated in a few competitions, but this is the first time I have placed so high on the leaderboard. Also, this is my first post so Kindly upvote the post if you find it helpful.
\nI have started off with exploratory data analysis, where I found out that the data does not have null values.
The contry_of_res variable had 61 countries and some countries only had a few data records, so I created a new feature continent which maps the respective countries to their continents and dropped the contry_of_res variable.
Variable age_desc had only one category, so I removed that variable.
Variables relation and ethnicity had a category ?. I have renamed that category as unknown.
Variables relation and ethnicity also had some categories with fewer data records, so I have binned them into the Others category.
I have trained all the models using 10 fold stratified cross-validation. Normalized the numerical variables using StandardScaler, ordinal encoded the categorical variables using OrdinalEncoder
\nBaseline model - Scikit-learn's Logistic Regression model with default parameters. Public LB: 0.82655, Private LB: 0.98051. This model would still have been the 8th place solution.
XGBoost model with hyperparameter tuned using Optuna. I have used scale_pos_weight as the data was imbalanced. Public LB: 0.93939, Private LB: 0.98268. This was my best submitted solution.
counter = Counter(Y_train)\nscale_pos_weight = counter.(0) / counter.(1)\n\nXGB_Params = {\n : 201, : 15,\n : 0.0038, : 0.7742,\n : 6.654, : 0.3,\n : 0.5, : 7\n}\n\nmodel = XGBClassifier(\n =,\n =scale_pos_weight,\n =,\n =,\n =,\n =,\n =42,\n =-1,\n **XGB_Params\n)\nis_unbalance as the data was imbalanced. Public LB: 0.92477, Private LB: 0.96103. LGBM_Params = {\n : 381, : 0.0451, : 2680,\n : 7, : 200, : 5,\n : 40, : 0.5696,\n : 0.8, : 1, : 0.25\n}\n\nmodel = LGBMClassifier(\n =,\n =,\n =,\n =42,\n =-1,\n **LGBM_Params\n)\nclass_weight as the data was imbalanced. Public LB: 0.93791, Private LB: 0.98268. This score was exactly same as my best submitted XGBoost model.counter = Counter(Y_train)\n\nparam_grid = {\n : 958, : ,\n : 3\n}\n\nclass_weight = {\n 0: n_samples / (counter.(0) * len(counter)),\n 1: n_samples / (counter.(1) * len(counter))\n}\n\nmodel = RandomForestClassifier(\n =class_weight,\n =0,\n =-1,\n **param_grid\n)\nis_unbalance as the data was imbalanced. Public LB: 0.93748 Private LB: 0.97835. counter = Counter(Y_train)\nscale_pos_weight = counter.(0) / counter.(1)\n\nCatBoost_Params = {\n : 186, : 0.00032,\n : 0.0801, : 1,\n : , : 4.3617\n}\n\nmodel = CatBoostClassifier(\n *tBoost_Params,\n =scale_pos_weight,\n =,\n =,\n =42,\n =0\n)\nHope this was helpful.
", "rawMarkdown": "First of all, **Thank you so much @usharengaraju** for hosting this amazing competition & to all the fellow Kagglers for sharing their knowledge via notebooks & discussions. I was able to learn a lot from them. I have participated in a few competitions, but this is the first time I have placed so high on the leaderboard. Also, this is my first post so **Kindly upvote the post if you find it helpful.**\n\n## 1. Exploratory Data Analysis & Feature Engineering:\n\n1. I have started off with exploratory data analysis, where I found out that the data does not have null values.\n\n2. The `contry_of_res` variable had 61 countries and some countries only had a few data records, so I created a new feature `continent` which maps the respective countries to their continents and dropped the `contry_of_res` variable.\n\n3. Variable `age_desc` had only one category, so I removed that variable.\n\n4. Variables `relation` and `ethnicity` had a category `?`. I have renamed that category as `unknown`.\n\n5. Variables `relation` and `ethnicity` also had some categories with fewer data records, so I have binned them into the `Others` category.\n\n## 2. Model Building\n\n### Strategy:\n\nI have trained all the models using 10 fold stratified cross-validation. Normalized the numerical variables using StandardScaler, ordinal encoded the categorical variables using OrdinalEncoder\n\n### Models:\n\n- Baseline model - Scikit-learn's Logistic Regression model with default parameters. **Public LB: 0.82655, Private LB: 0.98051**. This model would still have been the 8th place solution.\n\n- XGBoost model with hyperparameter tuned using Optuna. I have used `scale_pos_weight` as the data was imbalanced. **Public LB: 0.93939, Private LB: 0.98268**. This was my best submitted solution.\n\n```\ncounter = Counter(Y_train)\nscale_pos_weight = counter.get(0) / counter.get(1)\n\nXGB_Params = {\n \"n_estimators\": 201, \"max_depth\": 15,\n \"learning_rate\": 0.0038, \"reg_lambda\": 0.7742,\n \"reg_alpha\": 6.654, \"colsample_bytree\": 0.3,\n \"subsample\": 0.5, \"min_child_weight\": 7\n}\n\nmodel = XGBClassifier(\n use_label_encoder=False,\n scale_pos_weight=scale_pos_weight,\n tree_method=\"gpu_hist\",\n predictor=\"gpu_predictor\",\n objective=\"binary:logistic\",\n eval_metric=\"auc\",\n random_state=42,\n n_jobs=-1,\n **XGB_Params\n)\n\n```\n\n- LightGBM model with hyperparameter tuned using Optuna. I have used `is_unbalance` as the data was imbalanced. **Public LB: 0.92477, Private LB: 0.96103**. \n\n```\nLGBM_Params = {\n \"n_estimators\": 381, \"learning_rate\": 0.0451, \"num_leaves\": 2680,\n \"max_depth\": 7, \"min_child_samples\": 200, \"reg_alpha\": 5,\n \"reg_lambda\": 40, \"min_split_gain\": 0.5696,\n \"subsample\": 0.8, \"subsample_freq\": 1, \"colsample_bytree\": 0.25\n}\n \nmodel = LGBMClassifier(\n objective=\"binary\",\n device=\"gpu\",\n is_unbalance=True,\n random_state=42,\n n_jobs=-1,\n **LGBM_Params\n)\n```\n\n- Random Forest model with hyperparameter tuned using Optuna. I have used `class_weight` as the data was imbalanced. **Public LB: 0.93791, Private LB: 0.98268**. This score was exactly same as my best submitted XGBoost model.\n\n```\ncounter = Counter(Y_train)\n \nparam_grid = {\n \"n_estimators\": 958, \"criterion\": \"entropy\",\n \"max_depth\": 3\n}\n\nclass_weight = {\n 0: n_samples / (counter.get(0) * len(counter)),\n 1: n_samples / (counter.get(1) * len(counter))\n}\n\nmodel = RandomForestClassifier(\n class_weight=class_weight,\n random_state=0,\n n_jobs=-1,\n **param_grid\n)\n```\n\n- Finally the CatBoost model with hyperparameter tuned using Optuna. I have used `is_unbalance` as the data was imbalanced. **Public LB: 0.93748 Private LB: 0.97835**. \n\n```\ncounter = Counter(Y_train)\nscale_pos_weight = counter.get(0) / counter.get(1)\n \nCatBoost_Params = {\n \"n_estimators\": 186, \"learning_rate\": 0.00032,\n \"colsample_bylevel\": 0.0801, \"depth\": 1,\n \"bootstrap_type\": \"Bayesian\", \"bagging_temperature\": 4.3617\n}\n \nmodel = CatBoostClassifier(\n **CatBoost_Params,\n scale_pos_weight=scale_pos_weight,\n objective=\"Logloss\",\n eval_metric=\"AUC\",\n random_state=42,\n verbose=0\n)\n```\n\nHope this was helpful.", "votes": 6 }, { "id": 1718317, "postDate": "2022-03-10T18:08:14.577Z", "content": "Good job !
", "rawMarkdown": "Good job !", "votes": 1, "replies": [ { "id": 1718341, "postDate": "2022-03-10T18:30:49.860Z", "content": "Thank you @dineshrajan.
", "rawMarkdown": "Thank you @dineshrajan." } ] }, { "id": 2012905, "postDate": "2022-11-01T14:15:43.543Z", "content": "Link to my solution notebooks: https://www.kaggle.com/competitions/autism-prediction/discussion/312178
", "rawMarkdown": "Link to my solution notebooks: https://www.kaggle.com/competitions/autism-prediction/discussion/312178" }, { "id": 1718338, "postDate": "2022-03-10T18:29:27.687Z", "content": "I have made my 8th place solution and other high-scoring notebooks public. You can find them in the competition codes section.
", "rawMarkdown": "I have made my 8th place solution and other high-scoring notebooks public. You can find them in the competition codes section." } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T08:19:18Z", "competition_id": 34093, "competition_slug": "autism-prediction", "team_id": 8156970, "solution_url": "/competitions/autism-prediction/writeups/kishore-s-8th-place-solution-summary" }