{
  "id": 14697,
  "title": "Currently last with a log-loss of ~13",
  "url": "/competitions/avito-context-ad-clicks/discussion/14697",
  "author_name": "Jose Jimenez",
  "post_date": "2015-06-12T11:45:23.323000",
  "votes": 1,
  "comment_count": 0,
  "views": 606,
  "content": "<p>Yup. Tonight I somehow managed to score last on the LB. Here's what I did (so you avoid it as well):</p>\n\n<p>1. Get several joins on the tables of the db. I got about 50k samples, 25k of them with IsClick = 0 and the remaining with IsClick = 1. (I thought getting enough info of both strata was a good idea).</p>\n<p>2. Fit a GBM to the data. About 5000 trees and interaction.level 4 gave me a mmce of about 0.26.</p>\n<p>3. Predict *classes* for the test data.</p>\n<p>4. Profit.</p>\n<p>Yup. Log-loss is veeeeeeery cruel. So yeah, next time I'll be predicting probabilities with a muuuuch simpler model.</p>\n\n<p>Good luck guys!</p>",
  "messages": [
    {
      "id": 81680,
      "postDate": "2015-06-12T11:45:23.323Z",
      "content": "<p>Yup. Tonight I somehow managed to score last on the LB. Here's what I did (so you avoid it as well):</p>\n\n<p>1. Get several joins on the tables of the db. I got about 50k samples, 25k of them with IsClick = 0 and the remaining with IsClick = 1. (I thought getting enough info of both strata was a good idea).</p>\n<p>2. Fit a GBM to the data. About 5000 trees and interaction.level 4 gave me a mmce of about 0.26.</p>\n<p>3. Predict *classes* for the test data.</p>\n<p>4. Profit.</p>\n<p>Yup. Log-loss is veeeeeeery cruel. So yeah, next time I'll be predicting probabilities with a muuuuch simpler model.</p>\n\n<p>Good luck guys!</p>",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "81680": ""
  }
}