{
  "id": 57102,
  "title": "0.2276 - Keras RNN with FastText Embeddings - Starter - Much scope for Improvement",
  "url": "/competitions/avito-demand-prediction/discussion/57102",
  "author_name": "Shanth",
  "post_date": "2018-05-19T05:22:53.785000",
  "votes": 8,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Link to the Public Kernel  <a href=\"https://www.kaggle.com/shanth84/avito-fast-text-keras-model/code\"><strong><em>0.2276 AVITO RNN STARTER</em></strong></a></p>\n\n<p>Scores 0.2276 on Leaderboard.  Starter code with plenty of scope for improvement. </p>\n\n<ol>\n<li>RNN with Fast Text Embeddings </li>\n<li><p>Summary of Approach </p>\n\n<ul><li>Combined Title and Description as a single text field </li>\n<li>Label Encoding of all the Text fields </li>\n<li>Combined param1, param2 and param3 for a single categorical field \n( Param 2 and 3 have close to 50% of NaN values ) </li>\n<li>Normalized price and item_seq_number using a simple logarithmic transformation </li>\n<li>Memory optimization as far as possible </li>\n<li>Prediction in batches to ensure no memory overflows </li>\n<li>NO TUNING WAS DONE  SO FAR</li></ul></li>\n<li><p>Scope for Improvement </p>\n\n<ul><li>More feature engineering  using activation_date and image_top_1 variables </li>\n<li>Network tuning </li>\n<li>Change RNN architecture </li></ul></li>\n</ol>\n\n<p>Welcome to all kinds of input and feedback</p>\n\n<p>Regards\nShanth </p>",
  "messages": [
    {
      "id": 330556,
      "postDate": "2018-05-19T05:22:53.787Z",
      "content": "<p>Link to the Public Kernel  <a href=\"https://www.kaggle.com/shanth84/avito-fast-text-keras-model/code\"><strong><em>0.2276 AVITO RNN STARTER</em></strong></a></p>\n\n<p>Scores 0.2276 on Leaderboard.  Starter code with plenty of scope for improvement. </p>\n\n<ol>\n<li>RNN with Fast Text Embeddings </li>\n<li><p>Summary of Approach </p>\n\n<ul><li>Combined Title and Description as a single text field </li>\n<li>Label Encoding of all the Text fields </li>\n<li>Combined param1, param2 and param3 for a single categorical field \n( Param 2 and 3 have close to 50% of NaN values ) </li>\n<li>Normalized price and item_seq_number using a simple logarithmic transformation </li>\n<li>Memory optimization as far as possible </li>\n<li>Prediction in batches to ensure no memory overflows </li>\n<li>NO TUNING WAS DONE  SO FAR</li></ul></li>\n<li><p>Scope for Improvement </p>\n\n<ul><li>More feature engineering  using activation_date and image_top_1 variables </li>\n<li>Network tuning </li>\n<li>Change RNN architecture </li></ul></li>\n</ol>\n\n<p>Welcome to all kinds of input and feedback</p>\n\n<p>Regards\nShanth </p>",
      "rawMarkdown": "Link to the Public Kernel  [***0.2276 AVITO RNN STARTER***][1]\n\nScores 0.2276 on Leaderboard.  Starter code with plenty of scope for improvement. \n\n1.  RNN with Fast Text Embeddings \n2.  Summary of Approach \n \n  - Combined Title and Description as a single text field \n  - Label Encoding of all the Text fields \n  - Combined param1, param2 and param3 for a single categorical field \n     ( Param 2 and 3 have close to 50% of NaN values ) \n  - Normalized price and item_seq_number using a simple logarithmic transformation \n  -  Memory optimization as far as possible \n  -  Prediction in batches to ensure no memory overflows \n  -  NO TUNING WAS DONE  SO FAR\n\n3.  Scope for Improvement \n\n   -  More feature engineering  using activation_date and image_top_1 variables \n   -  Network tuning \n   -  Change RNN architecture \n\nWelcome to all kinds of input and feedback\n\nRegards\nShanth \n\n\n  [1]: https://www.kaggle.com/shanth84/avito-fast-text-keras-model/code",
      "votes": 8
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "330556": "Link to the Public Kernel  [***0.2276 AVITO RNN STARTER***][1]\n\nScores 0.2276 on Leaderboard.  Starter code with plenty of scope for improvement. \n\n1.  RNN with Fast Text Embeddings \n2.  Summary of Approach \n \n  - Combined Title and Description as a single text field \n  - Label Encoding of all the Text fields \n  - Combined param1, param2 and param3 for a single categorical field \n     ( Param 2 and 3 have close to 50% of NaN values ) \n  - Normalized price and item_seq_number using a simple logarithmic transformation \n  -  Memory optimization as far as possible \n  -  Prediction in batches to ensure no memory overflows \n  -  NO TUNING WAS DONE  SO FAR\n\n3.  Scope for Improvement \n\n   -  More feature engineering  using activation_date and image_top_1 variables \n   -  Network tuning \n   -  Change RNN architecture \n\nWelcome to all kinds of input and feedback\n\nRegards\nShanth \n\n\n  [1]: https://www.kaggle.com/shanth84/avito-fast-text-keras-model/code"
  }
}