{
  "id": 375543,
  "title": "Appreciate the host and sharing my solutions",
  "content": "<p>I am grateful to the host to prepare such amazing opportunity to learn data science especially related to TimeSeries regression.</p>\n<p>Honest speaking, I'm not good at managing TimeSeries data set so much, thus I felt it could be good challenge to try this competition when I came here for the first time.</p>\n<p>Unfortunately I was not able to overtake the target that the host set up, but I made the best effort to try it and believed my skill should be improved through the competition haha.</p>\n<p>With my gratitude, I would like to share about my approaches on the competition as follows.</p>\n<hr>\n<h5>1. LightGBM【CV = 0.0084,  PublicLB = 0.14277】</h5>\n<pre><code>- remove or clip irregular records \n- create new feature variables \"month, day, hour\" with datetime\n- Kfold cross-validation(fold=5)\n- hyper-parameters tuning with Optuna\n</code></pre>\n<p>Actually I tried making original TimeSeries-validation, but it didn't work so much, haha.<br>\nHence I applied KFold-Cross-validation to the competition instead.</p>\n<hr>\n<h5>2. XGBoost【CV = 0.1050,  PublicLB = 0.14030】</h5>\n<pre><code>- same feature engineerings with LightGBM\n- Kfold cross-validation(fold=5)\n- hyper-parameters tuning with Optuna\n</code></pre>\n<hr>\n<h5>3. CatBoost【CV = 0.0870,  PublicLB = 0.14063】</h5>\n<pre><code>- same feature engineerings with LightGBM\n- Kfold cross-validation(fold=5)\n- hyper-parameters tuning with Optuna\n</code></pre>\n<hr>\n<p>To my surprise, NN was the most effective on the competition as bellow…</p>\n<h5>4. Neural Network【CV = 0.1170,  PublicLB = 0.13185】</h5>\n<pre><code>- same feature engineerings with LightGBM\n- Kfold cross-validation(fold=5)\n- some of hyper-parameters tuning with Optuna, and I adjusted the rest of the parameters manually\n</code></pre>\n<hr>\n<h5>5. Other models</h5>\n<pre><code>- Ridge         : PublicLB =0.21813\n- RandomForest  : PublicLB =0.14068\n- ExtraTrees    : PublicLB =0.14257\n</code></pre>\n<p>As for the other models, such as Ridge, RandomForest, ExtraTrees were all not effective either.</p>\n<hr>\n<p>Through the above approaches, I decided to focus on improving Neural Network model in the end.</p>\n<h5>6. Improve NN model【CV = 0.1116,  PublicLB = 0.12056】</h5>\n<pre><code>- add parts of tsrefresh feature variables\n- remove irregular records(that I thought) more\n- add \"u10/v10\" and \"u100/v100\" \n- some of hyper-parameters tuning with Optuna\n</code></pre>\n<p>As a matter of fact, I tried the definition(wind-speed and wind-direction) that the host kindly shared with us on \"Code\" page, but my Neural network was not able to recognized them unfortunately…</p>\n<hr>\n<h5>7. Ensemble</h5>\n<pre><code>- best_NN + best_LightGBM + best_XGBoost   PublicLB = 0.13166\n- best_NN + best_ExtraTrees                PublicLB = 0.12489\n- multi NN models(1st and 2nd models)      PublicLB = 0.12048\n</code></pre>\n<hr>\n<p>In my opinion, if I had familiar with wind-power academically, I would have come up with more effective additional feature variables perhaps.<br>\nHowever I am totally beginner on the field, hence I couldn't, haha…<br>\nNeural Network model of mine could probably find its causal relationship of the data-set instead of me luckily.<br>\nI mean, if we don't have any specific knowledge related to the data-set, Neural Network could also be a good option sometimes, I am supposed.</p>\n<p>Hope that the above approaches of mine could be useful, and I will have an opportunity to participate in your another competition again in the future.</p>\n<p>Cheers,<br>\nKenta from Japan</p>",
  "comments": [
    {
      "id": 2106876,
      "author_name": "Tanguy Esteoule",
      "author_url": "",
      "post_date": "2023-01-19T12:05:15.833000",
      "content": "<p>Thank you for your detailed message Kenta !</p>\n<p>Some additional information:</p>\n<ul>\n<li>I created this challenge as part of a tutorial for students of a French engineering school (sorry for the text written in French). I didn't think that other people would participate !</li>\n<li>With my job, I have access to production data but for legal reasons I couldn't distribute them so I shamelessly used <a href=\"https://www.kaggle.com/datasets/natasha23/wind-forecasting-dataset\" target=\"_blank\">this dataset</a> by adapting it to my needs (adding failure periods, changing dates)</li>\n<li>My score is obtained by using LightGBM, calculating the wind speed/direction (at 10 and 100m), adding hour/month. I also removed the corrupted data. The only difference with you is that I add the data at the hours before and after (ff100_-3h, ff100_-2h, …, ff100_+3h), this improves the results a bit (with \"shift\" method in pandas).</li>\n<li>I would have liked to reopen the challenge but it is not possible apparently</li>\n</ul>\n<p>Tanguy</p>",
      "votes": 3,
      "replies": [
        {
          "id": 2106956,
          "author_name": "KentaK0928",
          "author_url": "",
          "post_date": "2023-01-19T13:25:47.530000",
          "content": "<p><a href=\"https://www.kaggle.com/tanguyesteoule\" target=\"_blank\">@tanguyesteoule</a> Thank you for your reply and for the great additional tips, Tanguy San😄<br>\nYour competition is really helpful for us to learn and improve time-series data analysis.</p>\n<p>I didn't check \"shift\" of pandas, thus your additional info is pretty splendid to me as well!</p>\n<p>I will be trying the method on the weekend.</p>\n<p>Again, appreciate your competition and the beautiful insights.<br>\nHope that we will have an opportunity to enjoy a same competition in the future.</p>\n<p>Cheers, <br>\nKenta from Japan</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2096956,
      "author_name": "The Devastator",
      "author_url": "",
      "post_date": "2023-01-12T11:39:02.863000",
      "content": "<p>Thank you for the write-up! <br>\nThis is very very detailed! </p>\n<p>The Devastator.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 2106958,
          "author_name": "KentaK0928",
          "author_url": "",
          "post_date": "2023-01-19T13:27:14.263000",
          "content": "<p><a href=\"https://www.kaggle.com/thedevastator\" target=\"_blank\">@thedevastator</a> Thanks for the compliment, Devastator San too😄<br>\nI will keep making the best effort to improve my skills and to share useful tips💪</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "comment_count": 4,
  "messages": [
    {
      "id": 2082999,
      "postDate": "2023-01-02T07:25:21.173Z",
      "content": "<p>I am grateful to the host to prepare such amazing opportunity to learn data science especially related to TimeSeries regression.</p>\n<p>Honest speaking, I'm not good at managing TimeSeries data set so much, thus I felt it could be good challenge to try this competition when I came here for the first time.</p>\n<p>Unfortunately I was not able to overtake the target that the host set up, but I made the best effort to try it and believed my skill should be improved through the competition haha.</p>\n<p>With my gratitude, I would like to share about my approaches on the competition as follows.</p>\n<hr>\n<h5>1. LightGBM【CV = 0.0084,  PublicLB = 0.14277】</h5>\n<pre><code>- remove or clip irregular records \n- create new feature variables \"month, day, hour\" with datetime\n- Kfold cross-validation(fold=5)\n- hyper-parameters tuning with Optuna\n</code></pre>\n<p>Actually I tried making original TimeSeries-validation, but it didn't work so much, haha.<br>\nHence I applied KFold-Cross-validation to the competition instead.</p>\n<hr>\n<h5>2. XGBoost【CV = 0.1050,  PublicLB = 0.14030】</h5>\n<pre><code>- same feature engineerings with LightGBM\n- Kfold cross-validation(fold=5)\n- hyper-parameters tuning with Optuna\n</code></pre>\n<hr>\n<h5>3. CatBoost【CV = 0.0870,  PublicLB = 0.14063】</h5>\n<pre><code>- same feature engineerings with LightGBM\n- Kfold cross-validation(fold=5)\n- hyper-parameters tuning with Optuna\n</code></pre>\n<hr>\n<p>To my surprise, NN was the most effective on the competition as bellow…</p>\n<h5>4. Neural Network【CV = 0.1170,  PublicLB = 0.13185】</h5>\n<pre><code>- same feature engineerings with LightGBM\n- Kfold cross-validation(fold=5)\n- some of hyper-parameters tuning with Optuna, and I adjusted the rest of the parameters manually\n</code></pre>\n<hr>\n<h5>5. Other models</h5>\n<pre><code>- Ridge         : PublicLB =0.21813\n- RandomForest  : PublicLB =0.14068\n- ExtraTrees    : PublicLB =0.14257\n</code></pre>\n<p>As for the other models, such as Ridge, RandomForest, ExtraTrees were all not effective either.</p>\n<hr>\n<p>Through the above approaches, I decided to focus on improving Neural Network model in the end.</p>\n<h5>6. Improve NN model【CV = 0.1116,  PublicLB = 0.12056】</h5>\n<pre><code>- add parts of tsrefresh feature variables\n- remove irregular records(that I thought) more\n- add \"u10/v10\" and \"u100/v100\" \n- some of hyper-parameters tuning with Optuna\n</code></pre>\n<p>As a matter of fact, I tried the definition(wind-speed and wind-direction) that the host kindly shared with us on \"Code\" page, but my Neural network was not able to recognized them unfortunately…</p>\n<hr>\n<h5>7. Ensemble</h5>\n<pre><code>- best_NN + best_LightGBM + best_XGBoost   PublicLB = 0.13166\n- best_NN + best_ExtraTrees                PublicLB = 0.12489\n- multi NN models(1st and 2nd models)      PublicLB = 0.12048\n</code></pre>\n<hr>\n<p>In my opinion, if I had familiar with wind-power academically, I would have come up with more effective additional feature variables perhaps.<br>\nHowever I am totally beginner on the field, hence I couldn't, haha…<br>\nNeural Network model of mine could probably find its causal relationship of the data-set instead of me luckily.<br>\nI mean, if we don't have any specific knowledge related to the data-set, Neural Network could also be a good option sometimes, I am supposed.</p>\n<p>Hope that the above approaches of mine could be useful, and I will have an opportunity to participate in your another competition again in the future.</p>\n<p>Cheers,<br>\nKenta from Japan</p>",
      "rawMarkdown": "I am grateful to the host to prepare such amazing opportunity to learn data science especially related to TimeSeries regression.\n\nHonest speaking, I'm not good at managing TimeSeries data set so much, thus I felt it could be good challenge to try this competition when I came here for the first time.\n\nUnfortunately I was not able to overtake the target that the host set up, but I made the best effort to try it and believed my skill should be improved through the competition haha.\n\n\nWith my gratitude, I would like to share about my approaches on the competition as follows.\n\n---------------------\n##### 1. LightGBM【CV = 0.0084,  PublicLB = 0.14277】\n    - remove or clip irregular records \n    - create new feature variables \"month, day, hour\" with datetime\n    - Kfold cross-validation(fold=5)\n    - hyper-parameters tuning with Optuna\n\nActually I tried making original TimeSeries-validation, but it didn't work so much, haha.\nHence I applied KFold-Cross-validation to the competition instead.\n\n---------------------\n##### 2. XGBoost【CV = 0.1050,  PublicLB = 0.14030】\n    - same feature engineerings with LightGBM\n    - Kfold cross-validation(fold=5)\n    - hyper-parameters tuning with Optuna\n\n---------------------\n##### 3. CatBoost【CV = 0.0870,  PublicLB = 0.14063】\n    - same feature engineerings with LightGBM\n    - Kfold cross-validation(fold=5)\n    - hyper-parameters tuning with Optuna\n\n---------------------\nTo my surprise, NN was the most effective on the competition as bellow...\n\n##### 4. Neural Network【CV = 0.1170,  PublicLB = 0.13185】\n    - same feature engineerings with LightGBM\n    - Kfold cross-validation(fold=5)\n    - some of hyper-parameters tuning with Optuna, and I adjusted the rest of the parameters manually\n\n---------------------\n##### 5. Other models\n    - Ridge         : PublicLB =0.21813\n    - RandomForest  : PublicLB =0.14068\n    - ExtraTrees    : PublicLB =0.14257\nAs for the other models, such as Ridge, RandomForest, ExtraTrees were all not effective either.\n\n\n---------------------\nThrough the above approaches, I decided to focus on improving Neural Network model in the end.\n\n##### 6. Improve NN model【CV = 0.1116,  PublicLB = 0.12056】\n    - add parts of tsrefresh feature variables\n    - remove irregular records(that I thought) more\n    - add \"u10/v10\" and \"u100/v100\" \n    - some of hyper-parameters tuning with Optuna\nAs a matter of fact, I tried the definition(wind-speed and wind-direction) that the host kindly shared with us on \"Code\" page, but my Neural network was not able to recognized them unfortunately...\n\n---------------------\n##### 7. Ensemble\n    - best_NN + best_LightGBM + best_XGBoost   PublicLB = 0.13166\n    - best_NN + best_ExtraTrees                PublicLB = 0.12489\n    - multi NN models(1st and 2nd models)      PublicLB = 0.12048\n\n---------------------\n\nIn my opinion, if I had familiar with wind-power academically, I would have come up with more effective additional feature variables perhaps.\nHowever I am totally beginner on the field, hence I couldn't, haha...\nNeural Network model of mine could probably find its causal relationship of the data-set instead of me luckily.\nI mean, if we don't have any specific knowledge related to the data-set, Neural Network could also be a good option sometimes, I am supposed.\n\n\nHope that the above approaches of mine could be useful, and I will have an opportunity to participate in your another competition again in the future.\n\nCheers,\nKenta from Japan\n",
      "votes": 5
    },
    {
      "id": 2106876,
      "postDate": "2023-01-19T12:05:15.833Z",
      "content": "<p>Thank you for your detailed message Kenta !</p>\n<p>Some additional information:</p>\n<ul>\n<li>I created this challenge as part of a tutorial for students of a French engineering school (sorry for the text written in French). I didn't think that other people would participate !</li>\n<li>With my job, I have access to production data but for legal reasons I couldn't distribute them so I shamelessly used <a href=\"https://www.kaggle.com/datasets/natasha23/wind-forecasting-dataset\" target=\"_blank\">this dataset</a> by adapting it to my needs (adding failure periods, changing dates)</li>\n<li>My score is obtained by using LightGBM, calculating the wind speed/direction (at 10 and 100m), adding hour/month. I also removed the corrupted data. The only difference with you is that I add the data at the hours before and after (ff100_-3h, ff100_-2h, …, ff100_+3h), this improves the results a bit (with \"shift\" method in pandas).</li>\n<li>I would have liked to reopen the challenge but it is not possible apparently</li>\n</ul>\n<p>Tanguy</p>",
      "rawMarkdown": "Thank you for your detailed message Kenta !\n\nSome additional information:\n- I created this challenge as part of a tutorial for students of a French engineering school (sorry for the text written in French). I didn't think that other people would participate !\n- With my job, I have access to production data but for legal reasons I couldn't distribute them so I shamelessly used [this dataset](https://www.kaggle.com/datasets/natasha23/wind-forecasting-dataset) by adapting it to my needs (adding failure periods, changing dates)\n- My score is obtained by using LightGBM, calculating the wind speed/direction (at 10 and 100m), adding hour/month. I also removed the corrupted data. The only difference with you is that I add the data at the hours before and after (ff100_-3h, ff100_-2h, ..., ff100_+3h), this improves the results a bit (with \"shift\" method in pandas).\n- I would have liked to reopen the challenge but it is not possible apparently\n\nTanguy",
      "votes": 3,
      "replies": [
        {
          "id": 2106956,
          "postDate": "2023-01-19T13:25:47.530Z",
          "content": "<p><a href=\"https://www.kaggle.com/tanguyesteoule\" target=\"_blank\">@tanguyesteoule</a> Thank you for your reply and for the great additional tips, Tanguy San😄<br>\nYour competition is really helpful for us to learn and improve time-series data analysis.</p>\n<p>I didn't check \"shift\" of pandas, thus your additional info is pretty splendid to me as well!</p>\n<p>I will be trying the method on the weekend.</p>\n<p>Again, appreciate your competition and the beautiful insights.<br>\nHope that we will have an opportunity to enjoy a same competition in the future.</p>\n<p>Cheers, <br>\nKenta from Japan</p>",
          "rawMarkdown": "@tanguyesteoule Thank you for your reply and for the great additional tips, Tanguy San😄\nYour competition is really helpful for us to learn and improve time-series data analysis.\n\nI didn't check \"shift\" of pandas, thus your additional info is pretty splendid to me as well!\n\nI will be trying the method on the weekend.\n\nAgain, appreciate your competition and the beautiful insights.\nHope that we will have an opportunity to enjoy a same competition in the future.\n\nCheers, \nKenta from Japan"
        }
      ]
    },
    {
      "id": 2096956,
      "postDate": "2023-01-12T11:39:02.863Z",
      "content": "<p>Thank you for the write-up! <br>\nThis is very very detailed! </p>\n<p>The Devastator.</p>",
      "rawMarkdown": "Thank you for the write-up! \nThis is very very detailed! \n\nThe Devastator.\n",
      "votes": 1,
      "replies": [
        {
          "id": 2106958,
          "postDate": "2023-01-19T13:27:14.263Z",
          "content": "<p><a href=\"https://www.kaggle.com/thedevastator\" target=\"_blank\">@thedevastator</a> Thanks for the compliment, Devastator San too😄<br>\nI will keep making the best effort to improve my skills and to share useful tips💪</p>",
          "rawMarkdown": "@thedevastator Thanks for the compliment, Devastator San too😄\nI will keep making the best effort to improve my skills and to share useful tips💪"
        }
      ]
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-20T02:07:39Z",
  "competition_id": 37930,
  "competition_slug": "ensimag-wpf",
  "team_id": 9628939,
  "solution_url": "/competitions/ensimag-wpf/writeups/kentak0928-appreciate-the-host-and-sharing-my-solu"
}