{
  "id": 63550,
  "title": "Possible improvements to attract kagglers",
  "url": "/competitions/flavours-of-physics-kernels-only/discussion/63550",
  "author_name": "",
  "post_date": "2018-08-18T04:19:47.633004700Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I'm very new to data science, but I have experience in high energy physics.  After watching the introduction video, I was entirely puzzled.  The best folks on here seem to be good at data science but aren't physicists so I can imagine that they're even more intimidated, but would probably love to help if the goals were presented more clearly.  The data science can be separated from the physics.  (I know the physics and I still couldn't make much sense of the competition)</p>\n\n<p>To be more specific, the standard model poster isn't informative to anyone, at least one of the plots in the video lacked axes, and connections to Noether's theorem are philosophically satisfying but probably off putting to good data scientists.  I think the hosts could attract more folks by making the goals more clear to a data scientist.  </p>",
  "messages": [
    {
      "id": "372012",
      "postDate": "08/18/2018 04:19:47",
      "content": "<p>I'm very new to data science, but I have experience in high energy physics.  After watching the introduction video, I was entirely puzzled.  The best folks on here seem to be good at data science but aren't physicists so I can imagine that they're even more intimidated, but would probably love to help if the goals were presented more clearly.  The data science can be separated from the physics.  (I know the physics and I still couldn't make much sense of the competition)</p>\n\n<p>To be more specific, the standard model poster isn't informative to anyone, at least one of the plots in the video lacked axes, and connections to Noether's theorem are philosophically satisfying but probably off putting to good data scientists.  I think the hosts could attract more folks by making the goals more clear to a data scientist.  </p>",
      "rawMarkdown": "I'm very new to data science, but I have experience in high energy physics.  After watching the introduction video, I was entirely puzzled.  The best folks on here seem to be good at data science but aren't physicists so I can imagine that they're even more intimidated, but would probably love to help if the goals were presented more clearly.  The data science can be separated from the physics.  (I know the physics and I still couldn't make much sense of the competition)\n\nTo be more specific, the standard model poster isn't informative to anyone, at least one of the plots in the video lacked axes, and connections to Noether's theorem are philosophically satisfying but probably off putting to good data scientists.  I think the hosts could attract more folks by making the goals more clear to a data scientist.",
      "votes": null
    },
    {
      "id": "372598",
      "postDate": "08/19/2018 20:45:23",
      "content": "<p>The physics here isn't really critical. From a machine learning standpoint it seems a standard classification situation. Binary classification, 1 or 0, signal or not. The extra constraints are a little different. The more successful approaches I've looked at seem to have ensembles that take the best accuracy models, which appear to rely on mass estimates, and ensemble those with less accurate but less correlated models that don't rely so much on mass. For whatever reason that seems to be a bad thing. <br>\nThat involves some domain knowledge but if that is passed on it is still a question of the right features handled in the usual data science ways. Not a lot of heavy physics necessarily required. Although, participating in the competition you might hope to learn a little in the area. <br>\nAlso I think this is the second go round for this competition. So a lot of this was already covered in a more standard competition about 3 years back. <a href=\"https://www.kaggle.com/c/flavours-of-physics\">Flavours of Physics: Finding τ → μμμ</a> Some participants posted on the physics involved. Because of that maybe there hasn't been so much posted for this version of the competition. <br>\nWhat I've gathered, although I am not expert in either data science or physics.</p>",
      "rawMarkdown": "The physics here isn't really critical. From a machine learning standpoint it seems a standard classification situation. Binary classification, 1 or 0, signal or not. The extra constraints are a little different. The more successful approaches I've looked at seem to have ensembles that take the best accuracy models, which appear to rely on mass estimates, and ensemble those with less accurate but less correlated models that don't rely so much on mass. For whatever reason that seems to be a bad thing.  \nThat involves some domain knowledge but if that is passed on it is still a question of the right features handled in the usual data science ways. Not a lot of heavy physics necessarily required. Although, participating in the competition you might hope to learn a little in the area.  \nAlso I think this is the second go round for this competition. So a lot of this was already covered in a more standard competition about 3 years back. [Flavours of Physics: Finding τ → μμμ](https://www.kaggle.com/c/flavours-of-physics) Some participants posted on the physics involved. Because of that maybe there hasn't been so much posted for this version of the competition.  \nWhat I've gathered, although I am not expert in either data science or physics.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 372598,
      "author_name": "mik3hall",
      "author_url": "",
      "post_date": "08/19/2018 20:45:23",
      "content": "<p>The physics here isn't really critical. From a machine learning standpoint it seems a standard classification situation. Binary classification, 1 or 0, signal or not. The extra constraints are a little different. The more successful approaches I've looked at seem to have ensembles that take the best accuracy models, which appear to rely on mass estimates, and ensemble those with less accurate but less correlated models that don't rely so much on mass. For whatever reason that seems to be a bad thing. <br>\nThat involves some domain knowledge but if that is passed on it is still a question of the right features handled in the usual data science ways. Not a lot of heavy physics necessarily required. Although, participating in the competition you might hope to learn a little in the area. <br>\nAlso I think this is the second go round for this competition. So a lot of this was already covered in a more standard competition about 3 years back. <a href=\"https://www.kaggle.com/c/flavours-of-physics\">Flavours of Physics: Finding τ → μμμ</a> Some participants posted on the physics involved. Because of that maybe there hasn't been so much posted for this version of the competition. <br>\nWhat I've gathered, although I am not expert in either data science or physics.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "372012": "I'm very new to data science, but I have experience in high energy physics.  After watching the introduction video, I was entirely puzzled.  The best folks on here seem to be good at data science but aren't physicists so I can imagine that they're even more intimidated, but would probably love to help if the goals were presented more clearly.  The data science can be separated from the physics.  (I know the physics and I still couldn't make much sense of the competition)\n\nTo be more specific, the standard model poster isn't informative to anyone, at least one of the plots in the video lacked axes, and connections to Noether's theorem are philosophically satisfying but probably off putting to good data scientists.  I think the hosts could attract more folks by making the goals more clear to a data scientist.",
    "372598": "The physics here isn't really critical. From a machine learning standpoint it seems a standard classification situation. Binary classification, 1 or 0, signal or not. The extra constraints are a little different. The more successful approaches I've looked at seem to have ensembles that take the best accuracy models, which appear to rely on mass estimates, and ensemble those with less accurate but less correlated models that don't rely so much on mass. For whatever reason that seems to be a bad thing.  \nThat involves some domain knowledge but if that is passed on it is still a question of the right features handled in the usual data science ways. Not a lot of heavy physics necessarily required. Although, participating in the competition you might hope to learn a little in the area.  \nAlso I think this is the second go round for this competition. So a lot of this was already covered in a more standard competition about 3 years back. [Flavours of Physics: Finding τ → μμμ](https://www.kaggle.com/c/flavours-of-physics) Some participants posted on the physics involved. Because of that maybe there hasn't been so much posted for this version of the competition.  \nWhat I've gathered, although I am not expert in either data science or physics."
  },
  "source": "meta"
}