{
  "id": 546909,
  "title": "Choosing a model and understanding why it works",
  "url": "/competitions/jane-street-real-time-market-data-forecasting/discussion/546909",
  "author_name": "",
  "post_date": "2024-11-18T19:06:14.185829800Z",
  "votes": null,
  "comment_count": 8,
  "views": 0,
  "content": "<p>I just wanted to step back and ask a fundamental question about what we are supposed to do in a Kaggle contest.</p>\n<p>There are a set of known models that can be applied to a contest like this one. We may try all these models and find the best one. </p>\n<p>Still, can we say that we know why this model works best? Isn't it that we just picked a model that gives the best scores?</p>\n<p>In the past contests, I saw winners used a model that I didn't even know it existed. Then, is it that we need to rummage papers on ML or DL models that deal with similar datasets?</p>\n<p>I wish to know if it is possible to learn the patterns of the given data and invent a new model that best fits the data. Are we supposed to invent a new model or tweak existing models? Is it what all contests are about?</p>",
  "messages": [
    {
      "id": "3049149",
      "postDate": "11/18/2024 19:06:14",
      "content": "<p>I just wanted to step back and ask a fundamental question about what we are supposed to do in a Kaggle contest.</p>\n<p>There are a set of known models that can be applied to a contest like this one. We may try all these models and find the best one. </p>\n<p>Still, can we say that we know why this model works best? Isn't it that we just picked a model that gives the best scores?</p>\n<p>In the past contests, I saw winners used a model that I didn't even know it existed. Then, is it that we need to rummage papers on ML or DL models that deal with similar datasets?</p>\n<p>I wish to know if it is possible to learn the patterns of the given data and invent a new model that best fits the data. Are we supposed to invent a new model or tweak existing models? Is it what all contests are about?</p>",
      "rawMarkdown": "I just wanted to step back and ask a fundamental question about what we are supposed to do in a Kaggle contest.\n\nThere are a set of known models that can be applied to a contest like this one. We may try all these models and find the best one. \n\nStill, can we say that we know why this model works best? Isn't it that we just picked a model that gives the best scores?\n\nIn the past contests, I saw winners used a model that I didn't even know it existed. Then, is it that we need to rummage papers on ML or DL models that deal with similar datasets?\n\nI wish to know if it is possible to learn the patterns of the given data and invent a new model that best fits the data. Are we supposed to invent a new model or tweak existing models? Is it what all contests are about?",
      "votes": null
    },
    {
      "id": "3049166",
      "postDate": "11/18/2024 19:26:31",
      "content": "<p>\"invent a new model\" -- to what extent do you consider a model is a \"new model\"?</p>",
      "rawMarkdown": "\"invent a new model\" -- to what extent do you consider a model is a \"new model\"?",
      "votes": null
    },
    {
      "id": "3049183",
      "postDate": "11/18/2024 19:58:01",
      "content": "<p>Hard to define one, but let's say one needing a name to distinguish from other models.</p>",
      "rawMarkdown": "Hard to define one, but let's say one needing a name to distinguish from other models.",
      "votes": null
    },
    {
      "id": "3049307",
      "postDate": "11/19/2024 00:09:08",
      "content": "<p>For example, like when transformer was invented in the \"Attention is all you need\" paper? It would be too ambitious :)</p>",
      "rawMarkdown": "For example, like when transformer was invented in the \"Attention is all you need\" paper? It would be too ambitious :)",
      "votes": null
    },
    {
      "id": "3049322",
      "postDate": "11/19/2024 01:05:55",
      "content": "<p>Yes, it would be too ambitious. Almost zero possibility. So, now I am wondering, the purpose of contestees is to apply existing models? Is this contest all about?</p>",
      "rawMarkdown": "Yes, it would be too ambitious. Almost zero possibility. So, now I am wondering, the purpose of contestees is to apply existing models? Is this contest all about?",
      "votes": null
    },
    {
      "id": "3049499",
      "postDate": "11/19/2024 07:09:53",
      "content": "<p>Try for the start to apply the existing ones on this data and you'll meet all the additional problems so you'll have the answer if it's just about the type of the model. The short answer is no, because we have to deal with a lot more and the type of the model is just one of the questions.</p>",
      "rawMarkdown": "Try for the start to apply the existing ones on this data and you'll meet all the additional problems so you'll have the answer if it's just about the type of the model. The short answer is no, because we have to deal with a lot more and the type of the model is just one of the questions.",
      "votes": null
    },
    {
      "id": "3049829",
      "postDate": "11/19/2024 14:51:14",
      "content": "<p>I'm quite sure \"inventing new models\" will be very rare in competitions such as these. To answer your question, no, we're not \"supposed to\" do that, but hey, if you have a big idea, are able to implement it, and it works, by all means, congratulations. I'm far from first positions here but I've made some creative solutions I'm hoping will make it there, and I do expect some creative geniuses in the top positions (maybe not \"new model\" level but still), but I also expect most top solutions to be a very good application of known models, with tweaks.</p>",
      "rawMarkdown": "I'm quite sure \"inventing new models\" will be very rare in competitions such as these. To answer your question, no, we're not \"supposed to\" do that, but hey, if you have a big idea, are able to implement it, and it works, by all means, congratulations. I'm far from first positions here but I've made some creative solutions I'm hoping will make it there, and I do expect some creative geniuses in the top positions (maybe not \"new model\" level but still), but I also expect most top solutions to be a very good application of known models, with tweaks.",
      "votes": null
    },
    {
      "id": "3050184",
      "postDate": "11/19/2024 22:44:28",
      "content": "<p>I appreciate your comment.</p>",
      "rawMarkdown": "I appreciate your comment.",
      "votes": null
    },
    {
      "id": "3050185",
      "postDate": "11/19/2024 22:44:42",
      "content": "<p>Thank you for sharing your opinion.</p>",
      "rawMarkdown": "Thank you for sharing your opinion.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3049166,
      "author_name": "shiyili",
      "author_url": "",
      "post_date": "11/18/2024 19:26:31",
      "content": "<p>\"invent a new model\" -- to what extent do you consider a model is a \"new model\"?</p>",
      "votes": null,
      "replies": [
        {
          "id": 3049183,
          "author_name": "cgl029",
          "author_url": "",
          "post_date": "11/18/2024 19:58:01",
          "content": "<p>Hard to define one, but let's say one needing a name to distinguish from other models.</p>",
          "votes": null,
          "replies": [
            {
              "id": 3049307,
              "author_name": "shiyili",
              "author_url": "",
              "post_date": "11/19/2024 00:09:08",
              "content": "<p>For example, like when transformer was invented in the \"Attention is all you need\" paper? It would be too ambitious :)</p>",
              "votes": null,
              "replies": [
                {
                  "id": 3049322,
                  "author_name": "cgl029",
                  "author_url": "",
                  "post_date": "11/19/2024 01:05:55",
                  "content": "<p>Yes, it would be too ambitious. Almost zero possibility. So, now I am wondering, the purpose of contestees is to apply existing models? Is this contest all about?</p>",
                  "votes": null,
                  "replies": [
                    {
                      "id": 3049499,
                      "author_name": "eu1234",
                      "author_url": "",
                      "post_date": "11/19/2024 07:09:53",
                      "content": "<p>Try for the start to apply the existing ones on this data and you'll meet all the additional problems so you'll have the answer if it's just about the type of the model. The short answer is no, because we have to deal with a lot more and the type of the model is just one of the questions.</p>",
                      "votes": null,
                      "replies": [
                        {
                          "id": 3050184,
                          "author_name": "cgl029",
                          "author_url": "",
                          "post_date": "11/19/2024 22:44:28",
                          "content": "<p>I appreciate your comment.</p>",
                          "votes": null,
                          "replies": []
                        }
                      ]
                    }
                  ]
                }
              ]
            }
          ]
        }
      ]
    },
    {
      "id": 3049829,
      "author_name": "natanlabarrere",
      "author_url": "",
      "post_date": "11/19/2024 14:51:14",
      "content": "<p>I'm quite sure \"inventing new models\" will be very rare in competitions such as these. To answer your question, no, we're not \"supposed to\" do that, but hey, if you have a big idea, are able to implement it, and it works, by all means, congratulations. I'm far from first positions here but I've made some creative solutions I'm hoping will make it there, and I do expect some creative geniuses in the top positions (maybe not \"new model\" level but still), but I also expect most top solutions to be a very good application of known models, with tweaks.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3050185,
          "author_name": "cgl029",
          "author_url": "",
          "post_date": "11/19/2024 22:44:42",
          "content": "<p>Thank you for sharing your opinion.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3049149": "I just wanted to step back and ask a fundamental question about what we are supposed to do in a Kaggle contest.\n\nThere are a set of known models that can be applied to a contest like this one. We may try all these models and find the best one. \n\nStill, can we say that we know why this model works best? Isn't it that we just picked a model that gives the best scores?\n\nIn the past contests, I saw winners used a model that I didn't even know it existed. Then, is it that we need to rummage papers on ML or DL models that deal with similar datasets?\n\nI wish to know if it is possible to learn the patterns of the given data and invent a new model that best fits the data. Are we supposed to invent a new model or tweak existing models? Is it what all contests are about?",
    "3049166": "\"invent a new model\" -- to what extent do you consider a model is a \"new model\"?",
    "3049183": "Hard to define one, but let's say one needing a name to distinguish from other models.",
    "3049307": "For example, like when transformer was invented in the \"Attention is all you need\" paper? It would be too ambitious :)",
    "3049322": "Yes, it would be too ambitious. Almost zero possibility. So, now I am wondering, the purpose of contestees is to apply existing models? Is this contest all about?",
    "3049499": "Try for the start to apply the existing ones on this data and you'll meet all the additional problems so you'll have the answer if it's just about the type of the model. The short answer is no, because we have to deal with a lot more and the type of the model is just one of the questions.",
    "3049829": "I'm quite sure \"inventing new models\" will be very rare in competitions such as these. To answer your question, no, we're not \"supposed to\" do that, but hey, if you have a big idea, are able to implement it, and it works, by all means, congratulations. I'm far from first positions here but I've made some creative solutions I'm hoping will make it there, and I do expect some creative geniuses in the top positions (maybe not \"new model\" level but still), but I also expect most top solutions to be a very good application of known models, with tweaks.",
    "3050184": "I appreciate your comment.",
    "3050185": "Thank you for sharing your opinion."
  },
  "source": "meta"
}