{
  "id": 21357,
  "title": "Naive Bayes ",
  "url": "/competitions/facebook-v-predicting-check-ins/discussion/21357",
  "author_name": "",
  "post_date": "2016-06-01T09:13:53.340Z",
  "votes": null,
  "comment_count": 2,
  "views": 847,
  "content": "<p>Given that there is so much data, has anyone tried to use &quot;naive bayes&quot; to predict the probablity of place-ids given x,y,time etc.?</p>\n\n<p>Ie P(place = place_x |x=x1, y=y1,t=t1...) = P(place = place_x) * P(x=x1|place=place_x) * P(y=y1 | place = place_x) ....</p>\n\n<p>I think the Naive Bayes assumption of the features being independent seems to hold looking at the attached scatter plots below. </p>\n\n<p>One is for a random subset of 100 samples and the other is for all samples with accuracy &gt; 50%. The first one shows that features are totally scattered w.r.t each other (except for accuracy, which need not be considered as a feature). The latter has too many points but clearly there is no linear relationship.</p>",
  "messages": [
    {
      "id": "122104",
      "postDate": "06/01/2016 09:13:53",
      "content": "<p>Given that there is so much data, has anyone tried to use &quot;naive bayes&quot; to predict the probablity of place-ids given x,y,time etc.?</p>\n\n<p>Ie P(place = place_x |x=x1, y=y1,t=t1...) = P(place = place_x) * P(x=x1|place=place_x) * P(y=y1 | place = place_x) ....</p>\n\n<p>I think the Naive Bayes assumption of the features being independent seems to hold looking at the attached scatter plots below. </p>\n\n<p>One is for a random subset of 100 samples and the other is for all samples with accuracy &gt; 50%. The first one shows that features are totally scattered w.r.t each other (except for accuracy, which need not be considered as a feature). The latter has too many points but clearly there is no linear relationship.</p>",
      "rawMarkdown": "Given that there is so much data, has anyone tried to use \"naive bayes\" to predict the probablity of place-ids given x,y,time etc.?\r\n\r\nIe P(place = place_x |x=x1, y=y1,t=t1...) = P(place = place_x) * P(x=x1|place=place_x) * P(y=y1 | place = place_x) ....\r\n\r\nI think the Naive Bayes assumption of the features being independent seems to hold looking at the attached scatter plots below. \r\n\r\nOne is for a random subset of 100 samples and the other is for all samples with accuracy > 50%. The first one shows that features are totally scattered w.r.t each other (except for accuracy, which need not be considered as a feature). The latter has too many points but clearly there is no linear relationship.",
      "votes": null
    },
    {
      "id": "122106",
      "postDate": "06/01/2016 09:35:24",
      "content": "<p>I think naive bayes is worth trying, although my brief attempt with it yielded worse results than the public scripts. One major problem I had was with the prediction time of sklearn's implementation, it was taking about 8 seconds do predict 1000 test observations, which would be &gt;60000 seconds for the complete test set. While that's possible, it's a bit too long for testing purposes.</p>",
      "rawMarkdown": "I think naive bayes is worth trying, although my brief attempt with it yielded worse results than the public scripts. One major problem I had was with the prediction time of sklearn's implementation, it was taking about 8 seconds do predict 1000 test observations, which would be >60000 seconds for the complete test set. While that's possible, it's a bit too long for testing purposes.",
      "votes": null
    },
    {
      "id": "122134",
      "postDate": "06/01/2016 13:47:45",
      "content": "<p>I have tried. ~0.400 on LB. Is your local testing group split by time interval? accuracy &gt; 50% is slightly &quot;too good&quot; I guess?</p>",
      "rawMarkdown": "I have tried. ~0.400 on LB. Is your local testing group split by time interval? accuracy > 50% is slightly \"too good\" I guess?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 122106,
      "author_name": "qwang88",
      "author_url": "",
      "post_date": "06/01/2016 09:35:24",
      "content": "<p>I think naive bayes is worth trying, although my brief attempt with it yielded worse results than the public scripts. One major problem I had was with the prediction time of sklearn's implementation, it was taking about 8 seconds do predict 1000 test observations, which would be &gt;60000 seconds for the complete test set. While that's possible, it's a bit too long for testing purposes.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 122134,
      "author_name": "ma350365879",
      "author_url": "",
      "post_date": "06/01/2016 13:47:45",
      "content": "<p>I have tried. ~0.400 on LB. Is your local testing group split by time interval? accuracy &gt; 50% is slightly &quot;too good&quot; I guess?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "122104": "Given that there is so much data, has anyone tried to use \"naive bayes\" to predict the probablity of place-ids given x,y,time etc.?\r\n\r\nIe P(place = place_x |x=x1, y=y1,t=t1...) = P(place = place_x) * P(x=x1|place=place_x) * P(y=y1 | place = place_x) ....\r\n\r\nI think the Naive Bayes assumption of the features being independent seems to hold looking at the attached scatter plots below. \r\n\r\nOne is for a random subset of 100 samples and the other is for all samples with accuracy > 50%. The first one shows that features are totally scattered w.r.t each other (except for accuracy, which need not be considered as a feature). The latter has too many points but clearly there is no linear relationship.",
    "122106": "I think naive bayes is worth trying, although my brief attempt with it yielded worse results than the public scripts. One major problem I had was with the prediction time of sklearn's implementation, it was taking about 8 seconds do predict 1000 test observations, which would be >60000 seconds for the complete test set. While that's possible, it's a bit too long for testing purposes.",
    "122134": "I have tried. ~0.400 on LB. Is your local testing group split by time interval? accuracy > 50% is slightly \"too good\" I guess?"
  },
  "source": "meta"
}