{
  "id": 380169,
  "title": "Last tips that helps me beat the benchmark recall using GBDT",
  "url": "/competitions/otto-recommender-system/discussion/380169",
  "author_name": "",
  "post_date": "2023-01-22T11:59:42.764741900Z",
  "votes": 16,
  "comment_count": 2,
  "views": 0,
  "content": "<p>In this topic, I will share some ideas that helped me improve **my orders recall ** in the Otto Recommender System competition on Kaggle from 0.62 to 0.65. These ideas include :</p>\n<ol>\n<li>Creating <strong>similarity scores variables</strong> between users and candidates, as well as between candidates and the last item seen by the user.</li>\n<li>I also found that <strong>increasing the learning rate</strong> helped to boost my score,  it gave me the little boost I needed to reach 0.6521), it also helps me wait less during the validation and inference haha.</li>\n</ol>\n<p>Additionally, If your ranker can\"t beat the benchmark I suggest trying to mix the predictions of your ranker with the heuristic approach, as co-visitation matrices can be powerful in some cases<br>\nThanks to <a href=\"https://www.kaggle.com/cdeotte\" target=\"_blank\">@cdeotte</a> for providing the validation benchmark predictions.<br>\nHope it'll help</p>",
  "messages": [
    {
      "id": "2110756",
      "postDate": "01/22/2023 11:59:42",
      "content": "<p>In this topic, I will share some ideas that helped me improve **my orders recall ** in the Otto Recommender System competition on Kaggle from 0.62 to 0.65. These ideas include :</p>\n<ol>\n<li>Creating <strong>similarity scores variables</strong> between users and candidates, as well as between candidates and the last item seen by the user.</li>\n<li>I also found that <strong>increasing the learning rate</strong> helped to boost my score,  it gave me the little boost I needed to reach 0.6521), it also helps me wait less during the validation and inference haha.</li>\n</ol>\n<p>Additionally, If your ranker can\"t beat the benchmark I suggest trying to mix the predictions of your ranker with the heuristic approach, as co-visitation matrices can be powerful in some cases<br>\nThanks to <a href=\"https://www.kaggle.com/cdeotte\" target=\"_blank\">@cdeotte</a> for providing the validation benchmark predictions.<br>\nHope it'll help</p>",
      "rawMarkdown": "In this topic, I will share some ideas that helped me improve **my orders recall ** in the Otto Recommender System competition on Kaggle from 0.62 to 0.65. These ideas include :\n\n1.   Creating **similarity scores variables** between users and candidates, as well as between candidates and the last item seen by the user.\n\n\n2. I also found that **increasing the learning rate** helped to boost my score,  it gave me the little boost I needed to reach 0.6521), it also helps me wait less during the validation and inference haha.\n\n \n\n\n Additionally, If your ranker can\"t beat the benchmark I suggest trying to mix the predictions of your ranker with the heuristic approach, as co-visitation matrices can be powerful in some cases\n\nThanks to @cdeotte for providing the validation benchmark predictions.\nHope it'll help",
      "votes": null
    },
    {
      "id": "2111564",
      "postDate": "01/23/2023 01:31:24",
      "content": "<p>Thank you for sharing, very interesting.</p>\n<p>The Devastator.</p>",
      "rawMarkdown": "Thank you for sharing, very interesting.\n\nThe Devastator.",
      "votes": null
    },
    {
      "id": "2112606",
      "postDate": "01/23/2023 17:51:49",
      "content": "<p>Try thinking about information. There must be some information about session/aids, that is available to heuristic, but not available to GBDT Ranker (the other possibility is that you could not find right parameters for the GBDT model, but I guess this is usually not the case).</p>",
      "rawMarkdown": "Try thinking about information. There must be some information about session/aids, that is available to heuristic, but not available to GBDT Ranker (the other possibility is that you could not find right parameters for the GBDT model, but I guess this is usually not the case).",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2111564,
      "author_name": "thedevastator",
      "author_url": "",
      "post_date": "01/23/2023 01:31:24",
      "content": "<p>Thank you for sharing, very interesting.</p>\n<p>The Devastator.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2112606,
      "author_name": "artemfedorov",
      "author_url": "",
      "post_date": "01/23/2023 17:51:49",
      "content": "<p>Try thinking about information. There must be some information about session/aids, that is available to heuristic, but not available to GBDT Ranker (the other possibility is that you could not find right parameters for the GBDT model, but I guess this is usually not the case).</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2110756": "In this topic, I will share some ideas that helped me improve **my orders recall ** in the Otto Recommender System competition on Kaggle from 0.62 to 0.65. These ideas include :\n\n1.   Creating **similarity scores variables** between users and candidates, as well as between candidates and the last item seen by the user.\n\n\n2. I also found that **increasing the learning rate** helped to boost my score,  it gave me the little boost I needed to reach 0.6521), it also helps me wait less during the validation and inference haha.\n\n \n\n\n Additionally, If your ranker can\"t beat the benchmark I suggest trying to mix the predictions of your ranker with the heuristic approach, as co-visitation matrices can be powerful in some cases\n\nThanks to @cdeotte for providing the validation benchmark predictions.\nHope it'll help",
    "2111564": "Thank you for sharing, very interesting.\n\nThe Devastator.",
    "2112606": "Try thinking about information. There must be some information about session/aids, that is available to heuristic, but not available to GBDT Ranker (the other possibility is that you could not find right parameters for the GBDT model, but I guess this is usually not the case)."
  },
  "source": "meta"
}