{
  "id": 11329,
  "title": "Mini-batch or Online SGD",
  "url": "/competitions/avazu-ctr-prediction/discussion/11329",
  "author_name": "",
  "post_date": "2014-12-23T22:23:51.683Z",
  "votes": null,
  "comment_count": 4,
  "views": 2771,
  "content": "<p>Considering the size&nbsp;of the training dataset, seems like mini-batch or online/stochastic gradient descent will be good options to consider. Has anyone got good results on LB with these options?&nbsp;</p>\n<p>Also, anyone with experience using Apache Spark? Seems like this is a good problem to try out on Spark.</p>",
  "messages": [
    {
      "id": "61787",
      "postDate": "12/23/2014 22:23:51",
      "content": "<p>Considering the size&nbsp;of the training dataset, seems like mini-batch or online/stochastic gradient descent will be good options to consider. Has anyone got good results on LB with these options?&nbsp;</p>\n<p>Also, anyone with experience using Apache Spark? Seems like this is a good problem to try out on Spark.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "61825",
      "postDate": "12/24/2014 13:44:50",
      "content": "<p>I am looking into this now as well. Training either a log-reg or neural-net using mini-batch SGD.</p>\n\n<p>Some python code for batch-generation:&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62740",
      "postDate": "01/15/2015 06:52:49",
      "content": "<p>i have used spark for feature engineering.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62745",
      "postDate": "01/15/2015 10:06:49",
      "content": "<p>[quote=logloss;61787]</p>\n<p>Considering the size&nbsp;of the training dataset, seems like mini-batch or online/stochastic gradient descent will be good options to consider. Has anyone got good results on LB with these options?&nbsp;</p>\n<p>Also, anyone with experience using Apache Spark? Seems like this is a good problem to try out on Spark.</p>\n<p>[/quote]</p>\n\n<p>I have been thinking about the same thing. Instead of implementing the algorithms myself I'm looking for R packages for ini-batch or online/stochastic gradient descent. While this approach may not be the best for LB scores I think it's easier to learn in this manner. Appreciate any pointers.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62812",
      "postDate": "01/17/2015 05:13:57",
      "content": "<p>Guys who used Spark: did you deploy it as a cluster to EC2 or just ran it on the local machine? If cluster, how many workers you used?</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 61825,
      "author_name": "luckylwk",
      "author_url": "",
      "post_date": "12/24/2014 13:44:50",
      "content": "<p>I am looking into this now as well. Training either a log-reg or neural-net using mini-batch SGD.</p>\n\n<p>Some python code for batch-generation:&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62740,
      "author_name": "xuqiang1990",
      "author_url": "",
      "post_date": "01/15/2015 06:52:49",
      "content": "<p>i have used spark for feature engineering.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62745,
      "author_name": "hitesh",
      "author_url": "",
      "post_date": "01/15/2015 10:06:49",
      "content": "<p>[quote=logloss;61787]</p>\n<p>Considering the size&nbsp;of the training dataset, seems like mini-batch or online/stochastic gradient descent will be good options to consider. Has anyone got good results on LB with these options?&nbsp;</p>\n<p>Also, anyone with experience using Apache Spark? Seems like this is a good problem to try out on Spark.</p>\n<p>[/quote]</p>\n\n<p>I have been thinking about the same thing. Instead of implementing the algorithms myself I'm looking for R packages for ini-batch or online/stochastic gradient descent. While this approach may not be the best for LB scores I think it's easier to learn in this manner. Appreciate any pointers.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62812,
      "author_name": "yankov",
      "author_url": "",
      "post_date": "01/17/2015 05:13:57",
      "content": "<p>Guys who used Spark: did you deploy it as a cluster to EC2 or just ran it on the local machine? If cluster, how many workers you used?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "61787": "",
    "61825": "",
    "62740": "",
    "62745": "",
    "62812": ""
  },
  "source": "meta"
}