{
  "id": 25676,
  "title": "Vowpal wabbit vs FFM",
  "url": "/competitions/outbrain-click-prediction/discussion/25676",
  "author_name": "SachinTyagi",
  "post_date": "2016-11-22T06:46:19.113000",
  "votes": 3,
  "comment_count": 0,
  "views": 531,
  "content": "<p>Hi, </p>\n\n<p>I was going through the documentation of vowpal wabbit and from it it looked like that with certain options one can get it to mimic the behavior of an FFM. </p>\n\n<p>For instance, if we use matrix factorization with \"--rank\" option and then specify the quadratic interaction among all namespaces using -q :: then it appears to me that the underlying learning should not be too different from what LibFFM is doing. </p>\n\n<p>However, I want to get an opinion of other Kagglers here whether this line of thinking is correct. In what respects is FFM different from VW used with above options? (Is it that VW learns only single vector for each namespace (like FM) regardless of number of interactions with other features? Or something else?)</p>\n\n<p>Thanks in advance.</p>",
  "messages": [
    {
      "id": 145994,
      "postDate": "2016-11-22T06:46:19.113Z",
      "content": "<p>Hi, </p>\n\n<p>I was going through the documentation of vowpal wabbit and from it it looked like that with certain options one can get it to mimic the behavior of an FFM. </p>\n\n<p>For instance, if we use matrix factorization with \"--rank\" option and then specify the quadratic interaction among all namespaces using -q :: then it appears to me that the underlying learning should not be too different from what LibFFM is doing. </p>\n\n<p>However, I want to get an opinion of other Kagglers here whether this line of thinking is correct. In what respects is FFM different from VW used with above options? (Is it that VW learns only single vector for each namespace (like FM) regardless of number of interactions with other features? Or something else?)</p>\n\n<p>Thanks in advance.</p>",
      "rawMarkdown": "Hi, \r\n\r\nI was going through the documentation of vowpal wabbit and from it it looked like that with certain options one can get it to mimic the behavior of an FFM. \r\n\r\nFor instance, if we use matrix factorization with \"--rank\" option and then specify the quadratic interaction among all namespaces using -q :: then it appears to me that the underlying learning should not be too different from what LibFFM is doing. \r\n\r\nHowever, I want to get an opinion of other Kagglers here whether this line of thinking is correct. In what respects is FFM different from VW used with above options? (Is it that VW learns only single vector for each namespace (like FM) regardless of number of interactions with other features? Or something else?)\r\n\r\nThanks in advance.",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "145994": "Hi, \r\n\r\nI was going through the documentation of vowpal wabbit and from it it looked like that with certain options one can get it to mimic the behavior of an FFM. \r\n\r\nFor instance, if we use matrix factorization with \"--rank\" option and then specify the quadratic interaction among all namespaces using -q :: then it appears to me that the underlying learning should not be too different from what LibFFM is doing. \r\n\r\nHowever, I want to get an opinion of other Kagglers here whether this line of thinking is correct. In what respects is FFM different from VW used with above options? (Is it that VW learns only single vector for each namespace (like FM) regardless of number of interactions with other features? Or something else?)\r\n\r\nThanks in advance."
  }
}