{
  "id": 248247,
  "title": "Which strategy was better for this competition?",
  "url": "/competitions/coleridgeinitiative-show-us-the-data/discussion/248247",
  "author_name": "Yust",
  "post_date": "2021-06-23T00:18:17.795000",
  "votes": 0,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I think there were two major directions this competition could have gone.<br>\n1) String matching<br>\n2) Prediction by machine learning</p>\n<p>I thought that (1) had the possibility of overfitting to PublicLB, so I chose (1) + (2) and (2) only as my final submit.</p>\n<p>In the end, was matching (1) effective in this competition?<br>\nAlso, what was the main key to a good machine learning model in (2)? (Amount of training data?)</p>",
  "messages": [
    {
      "id": 1362259,
      "postDate": "2021-06-23T10:14:19.497Z",
      "content": "<p>Its not the string matching that is giving high scores but  simple intelligent rules which reduces the sample space  plus NER.</p>",
      "rawMarkdown": "Its not the string matching that is giving high scores but  simple intelligent rules which reduces the sample space  plus NER."
    },
    {
      "id": 1361603,
      "postDate": "2021-06-23T00:29:08.747Z",
      "content": "<p>Scores for me:</p>\n<ol>\n<li>With only string matching - 0.022</li>\n<li>Spacy custom NER - 0.17</li>\n<li>Ensemble of two custom NER models (both individually scoring around 0.17) : 0.305 (unfortunately I didn't choose this for final submission thinking that it was missing some outputs for sample submission)</li>\n</ol>\n<p>I think key was to predict less false positives as they were heavily penalized in the metric, and doing false negatives was fine as the factor for it was low. 😅</p>",
      "rawMarkdown": "Scores for me:\n1. With only string matching - 0.022\n2. Spacy custom NER - 0.17\n3. Ensemble of two custom NER models (both individually scoring around 0.17) : 0.305 (unfortunately I didn't choose this for final submission thinking that it was missing some outputs for sample submission)\n\nI think key was to predict less false positives as they were heavily penalized in the metric, and doing false negatives was fine as the factor for it was low. 😅",
      "replies": [
        {
          "id": 1361879,
          "postDate": "2021-06-23T05:29:42.817Z",
          "content": "<p>Just curious, what is 1.7 if f1 lies in [0, 1]? </p>",
          "rawMarkdown": "Just curious, what is 1.7 if f1 lies in [0, 1]? ",
          "votes": 1
        },
        {
          "id": 1361975,
          "postDate": "2021-06-23T06:32:57.807Z",
          "content": "<p><a href=\"https://www.kaggle.com/lightnessofbein\" target=\"_blank\">@lightnessofbein</a> my bad, thanks for pointing out. It was 0.17. Edited 🙂</p>",
          "rawMarkdown": "@lightnessofbein my bad, thanks for pointing out. It was 0.17. Edited 🙂",
          "votes": 1
        }
      ]
    },
    {
      "id": 1361599,
      "postDate": "2021-06-23T00:18:17.797Z",
      "content": "<p>I think there were two major directions this competition could have gone.<br>\n1) String matching<br>\n2) Prediction by machine learning</p>\n<p>I thought that (1) had the possibility of overfitting to PublicLB, so I chose (1) + (2) and (2) only as my final submit.</p>\n<p>In the end, was matching (1) effective in this competition?<br>\nAlso, what was the main key to a good machine learning model in (2)? (Amount of training data?)</p>",
      "rawMarkdown": "I think there were two major directions this competition could have gone.\n1) String matching\n2) Prediction by machine learning\n\nI thought that (1) had the possibility of overfitting to PublicLB, so I chose (1) + (2) and (2) only as my final submit.\n\nIn the end, was matching (1) effective in this competition?\nAlso, what was the main key to a good machine learning model in (2)? (Amount of training data?)"
    },
    {
      "id": 1361601,
      "postDate": "2021-06-23T00:27:00.630Z",
      "rawMarkdown": "",
      "votes": -3,
      "isDeleted": true,
      "replies": [
        {
          "id": 1361725,
          "postDate": "2021-06-23T03:22:24.317Z",
          "content": "<p>How do you know that?</p>",
          "rawMarkdown": "How do you know that?"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1362259,
      "author_name": "Equilan",
      "author_url": "",
      "post_date": "2021-06-23T10:14:19.497000",
      "content": "<p>Its not the string matching that is giving high scores but  simple intelligent rules which reduces the sample space  plus NER.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1361603,
      "author_name": "Shivam Miglani",
      "author_url": "",
      "post_date": "2021-06-23T00:29:08.747000",
      "content": "<p>Scores for me:</p>\n<ol>\n<li>With only string matching - 0.022</li>\n<li>Spacy custom NER - 0.17</li>\n<li>Ensemble of two custom NER models (both individually scoring around 0.17) : 0.305 (unfortunately I didn't choose this for final submission thinking that it was missing some outputs for sample submission)</li>\n</ol>\n<p>I think key was to predict less false positives as they were heavily penalized in the metric, and doing false negatives was fine as the factor for it was low. 😅</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1361879,
          "author_name": "Shrek_fan",
          "author_url": "",
          "post_date": "2021-06-23T05:29:42.817000",
          "content": "<p>Just curious, what is 1.7 if f1 lies in [0, 1]? </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1361975,
          "author_name": "Shivam Miglani",
          "author_url": "",
          "post_date": "2021-06-23T06:32:57.807000",
          "content": "<p><a href=\"https://www.kaggle.com/lightnessofbein\" target=\"_blank\">@lightnessofbein</a> my bad, thanks for pointing out. It was 0.17. Edited 🙂</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1361601,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-06-23T00:27:00.630000",
      "content": "",
      "votes": -3,
      "replies": [
        {
          "id": 1361725,
          "author_name": "Pablo Carbó",
          "author_url": "",
          "post_date": "2021-06-23T03:22:24.317000",
          "content": "<p>How do you know that?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1362259": "Its not the string matching that is giving high scores but  simple intelligent rules which reduces the sample space  plus NER.",
    "1361603": "Scores for me:\n1. With only string matching - 0.022\n2. Spacy custom NER - 0.17\n3. Ensemble of two custom NER models (both individually scoring around 0.17) : 0.305 (unfortunately I didn't choose this for final submission thinking that it was missing some outputs for sample submission)\n\nI think key was to predict less false positives as they were heavily penalized in the metric, and doing false negatives was fine as the factor for it was low. 😅",
    "1361599": "I think there were two major directions this competition could have gone.\n1) String matching\n2) Prediction by machine learning\n\nI thought that (1) had the possibility of overfitting to PublicLB, so I chose (1) + (2) and (2) only as my final submit.\n\nIn the end, was matching (1) effective in this competition?\nAlso, what was the main key to a good machine learning model in (2)? (Amount of training data?)",
    "1361601": ""
  }
}