{
  "id": 1196,
  "title": "Best CBD for a single method",
  "url": "/competitions/WhatDoYouKnow/discussion/1196",
  "author_name": "",
  "post_date": "2011-12-28T17:13:54.293Z",
  "votes": null,
  "comment_count": 9,
  "views": 3924,
  "content": "<p>I was wondering how well a single method performs on this dataset,&nbsp;</p>\r\n<p>My best single method, matrix factorization inside a logistic function.</p>\r\n<p><strong>Validation &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; Public Leaderboard</strong></p>\r\n<p>0.25211 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; 0.25320</p>",
  "messages": [
    {
      "id": "7536",
      "postDate": "12/28/2011 17:13:54",
      "content": "<p>I was wondering how well a single method performs on this dataset,&nbsp;</p>\r\n<p>My best single method, matrix factorization inside a logistic function.</p>\r\n<p><strong>Validation &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; Public Leaderboard</strong></p>\r\n<p>0.25211 &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; 0.25320</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "7538",
      "postDate": "12/28/2011 17:55:53",
      "content": "<p>I guess it depends on your definition of &quot;single method&quot;.&nbsp; Some of my stuff blurs the line a bit.</p>\r\n<p>Based on my own loose definition I'd say...</p>\r\n<table border=\"0\">\r\n<tbody>\r\n<tr>\r\n<td><strong>Validation&nbsp;&nbsp;</strong>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;</td>\r\n<td><strong>Leaderboard</strong></td>\r\n</tr>\r\n<tr>\r\n<td>0.251520</td>\r\n<td>\r\n<p>0.25275</p>\r\n</td>\r\n</tr>\r\n</tbody>\r\n</table>\r\n<p>By the way, I also tried logistic matrix factorization (by resurrecting some old Netflix Prize code I wrote).&nbsp; My validation and leaderboard scores were both ~0.0006 higher than yours.&nbsp; Hmm... Maybe a little more tweaking is in order.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "7540",
      "postDate": "12/28/2011 18:17:04",
      "content": "<p>Thanks Yetiman!&nbsp;How many latent factors are you using for your result?&nbsp;</p>\r\n<p>[edit]</p>\r\n<p>[quote=YetiMan;7538]</p>\r\n<p>I guess it depends on your definition of &quot;single method&quot;.&nbsp; Some of my stuff blurs the line a bit.</p>\r\n<p>Based on my own loose definition I'd say...</p>\r\n<table border=\"0\">\r\n<tbody>\r\n<tr>\r\n<td><strong>Validation&nbsp;&nbsp;</strong>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;</td>\r\n<td><strong>Leaderboard</strong></td>\r\n</tr>\r\n<tr>\r\n<td>0.251520</td>\r\n<td>\r\n<p>0.25275</p>\r\n</td>\r\n</tr>\r\n</tbody>\r\n</table>\r\n<p>By the way, I also tried logistic matrix factorization (by resurrecting some old Netflix Prize code I wrote).&nbsp; My validation and leaderboard scores were both ~0.0006 higher than yours.&nbsp; Hmm... Maybe a little more tweaking is in order.</p>\r\n<p>[/quote]</p>\r\n<p>&nbsp;</p>\r\n<p>&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "7541",
      "postDate": "12/28/2011 18:19:50",
      "content": "<p>I tried various numbers of factors but only saw miniscule improvement with factors&gt;80.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "7542",
      "postDate": "12/28/2011 18:22:42",
      "content": "<p>I use 3 as of now! I haven't tried numbers above 10 since I use sgd and haven't invested time in a cross-validation framework for learning rate and regularization. &nbsp;I am also using the group, track, subtrack, game_type and question type priors in the factorization.</p>\r\n<p>[quote=YetiMan;7541]</p>\r\n<p>I tried various numbers of factors but only saw miniscule improvement with factors&gt;80.</p>\r\n<p>[/quote]</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "7543",
      "postDate": "12/28/2011 18:33:26",
      "content": "<p>I also used (regularized) sgd for training, with learning/regularization rates chosen by trial and error.&nbsp; Not much difference there.&nbsp; But I'm using raw data with no priors, so I'm guessing that's the biggest reason for the score difference.</p>\r\n<p>By the way, the (only) submission I made to kaggle using this method alone used 32 factors.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "8090",
      "postDate": "01/13/2012 15:52:38",
      "content": "<p>what's sgd??</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "8091",
      "postDate": "01/13/2012 16:03:16",
      "content": "<p>SGD is &quot;stochastic gradient descent&quot;.&nbsp; There's plenty of web-accessible literature on this, particularly in a machine learning context (the backprop algorithm, is one example).&nbsp; You can get a basic description of how it works here: http://en.wikipedia.org/wiki/Stochastic_gradient_descent</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "8092",
      "postDate": "01/13/2012 16:24:48",
      "content": "<p>Thanks YetiMan~</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "349247",
      "postDate": "06/28/2018 00:43:35",
      "content": "<p>Triton-SD, Sorry for ask here.\n¿Are you from Cali? </p>",
      "rawMarkdown": "Triton-SD, Sorry for ask here.\n¿Are you from Cali?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 7538,
      "author_name": "yetiman",
      "author_url": "",
      "post_date": "12/28/2011 17:55:53",
      "content": "<p>I guess it depends on your definition of &quot;single method&quot;.&nbsp; Some of my stuff blurs the line a bit.</p>\r\n<p>Based on my own loose definition I'd say...</p>\r\n<table border=\"0\">\r\n<tbody>\r\n<tr>\r\n<td><strong>Validation&nbsp;&nbsp;</strong>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;</td>\r\n<td><strong>Leaderboard</strong></td>\r\n</tr>\r\n<tr>\r\n<td>0.251520</td>\r\n<td>\r\n<p>0.25275</p>\r\n</td>\r\n</tr>\r\n</tbody>\r\n</table>\r\n<p>By the way, I also tried logistic matrix factorization (by resurrecting some old Netflix Prize code I wrote).&nbsp; My validation and leaderboard scores were both ~0.0006 higher than yours.&nbsp; Hmm... Maybe a little more tweaking is in order.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 7540,
      "author_name": "tritonsd",
      "author_url": "",
      "post_date": "12/28/2011 18:17:04",
      "content": "<p>Thanks Yetiman!&nbsp;How many latent factors are you using for your result?&nbsp;</p>\r\n<p>[edit]</p>\r\n<p>[quote=YetiMan;7538]</p>\r\n<p>I guess it depends on your definition of &quot;single method&quot;.&nbsp; Some of my stuff blurs the line a bit.</p>\r\n<p>Based on my own loose definition I'd say...</p>\r\n<table border=\"0\">\r\n<tbody>\r\n<tr>\r\n<td><strong>Validation&nbsp;&nbsp;</strong>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;</td>\r\n<td><strong>Leaderboard</strong></td>\r\n</tr>\r\n<tr>\r\n<td>0.251520</td>\r\n<td>\r\n<p>0.25275</p>\r\n</td>\r\n</tr>\r\n</tbody>\r\n</table>\r\n<p>By the way, I also tried logistic matrix factorization (by resurrecting some old Netflix Prize code I wrote).&nbsp; My validation and leaderboard scores were both ~0.0006 higher than yours.&nbsp; Hmm... Maybe a little more tweaking is in order.</p>\r\n<p>[/quote]</p>\r\n<p>&nbsp;</p>\r\n<p>&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 7541,
      "author_name": "yetiman",
      "author_url": "",
      "post_date": "12/28/2011 18:19:50",
      "content": "<p>I tried various numbers of factors but only saw miniscule improvement with factors&gt;80.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 7542,
      "author_name": "tritonsd",
      "author_url": "",
      "post_date": "12/28/2011 18:22:42",
      "content": "<p>I use 3 as of now! I haven't tried numbers above 10 since I use sgd and haven't invested time in a cross-validation framework for learning rate and regularization. &nbsp;I am also using the group, track, subtrack, game_type and question type priors in the factorization.</p>\r\n<p>[quote=YetiMan;7541]</p>\r\n<p>I tried various numbers of factors but only saw miniscule improvement with factors&gt;80.</p>\r\n<p>[/quote]</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 7543,
      "author_name": "yetiman",
      "author_url": "",
      "post_date": "12/28/2011 18:33:26",
      "content": "<p>I also used (regularized) sgd for training, with learning/regularization rates chosen by trial and error.&nbsp; Not much difference there.&nbsp; But I'm using raw data with no priors, so I'm guessing that's the biggest reason for the score difference.</p>\r\n<p>By the way, the (only) submission I made to kaggle using this method alone used 32 factors.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 8090,
      "author_name": "bootstrap",
      "author_url": "",
      "post_date": "01/13/2012 15:52:38",
      "content": "<p>what's sgd??</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 8091,
      "author_name": "yetiman",
      "author_url": "",
      "post_date": "01/13/2012 16:03:16",
      "content": "<p>SGD is &quot;stochastic gradient descent&quot;.&nbsp; There's plenty of web-accessible literature on this, particularly in a machine learning context (the backprop algorithm, is one example).&nbsp; You can get a basic description of how it works here: http://en.wikipedia.org/wiki/Stochastic_gradient_descent</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 8092,
      "author_name": "bootstrap",
      "author_url": "",
      "post_date": "01/13/2012 16:24:48",
      "content": "<p>Thanks YetiMan~</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 349247,
      "author_name": "mvargas1144",
      "author_url": "",
      "post_date": "06/28/2018 00:43:35",
      "content": "<p>Triton-SD, Sorry for ask here.\n¿Are you from Cali? </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "7536": "",
    "7538": "",
    "7540": "",
    "7541": "",
    "7542": "",
    "7543": "",
    "8090": "",
    "8091": "",
    "8092": "",
    "349247": "Triton-SD, Sorry for ask here.\n¿Are you from Cali?"
  },
  "source": "meta"
}