{
  "id": 56209,
  "title": "Is it all about text features and model tuning?",
  "url": "/competitions/avito-demand-prediction/discussion/56209",
  "author_name": "",
  "post_date": "2018-05-07T19:58:18.559047800Z",
  "votes": 5,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I tried several combinations of text features, grouped mean and variance features, and VGG image features, yet so far the group and image features do not help with the public LB score at all. Model tuning, on the other hand, increases the score slightly. </p>\n\n<p>For text features, dimensionality reduction severely reduce the score on public LB and I have to keep all of them (around 20,000 so far). Is it because there are too many text features that the other features simply would not matter? Has anyone got better results by adding image features to a model with text features? It would be great if we can hear some good news on this!</p>",
  "messages": [
    {
      "id": "324571",
      "postDate": "05/07/2018 19:58:18",
      "content": "<p>I tried several combinations of text features, grouped mean and variance features, and VGG image features, yet so far the group and image features do not help with the public LB score at all. Model tuning, on the other hand, increases the score slightly. </p>\n\n<p>For text features, dimensionality reduction severely reduce the score on public LB and I have to keep all of them (around 20,000 so far). Is it because there are too many text features that the other features simply would not matter? Has anyone got better results by adding image features to a model with text features? It would be great if we can hear some good news on this!</p>",
      "rawMarkdown": "I tried several combinations of text features, grouped mean and variance features, and VGG image features, yet so far the group and image features do not help with the public LB score at all. Model tuning, on the other hand, increases the score slightly. \n\nFor text features, dimensionality reduction severely reduce the score on public LB and I have to keep all of them (around 20,000 so far). Is it because there are too many text features that the other features simply would not matter? Has anyone got better results by adding image features to a model with text features? It would be great if we can hear some good news on this!",
      "votes": null
    },
    {
      "id": "324763",
      "postDate": "05/07/2018 23:34:15",
      "content": "<p>I have the same concern as you. The text features improved my model rapidly. Adding image features seems does not help.\nI am wondering if text feature is all we need</p>",
      "rawMarkdown": "I have the same concern as you. The text features improved my model rapidly. Adding image features seems does not help.\nI am wondering if text feature is all we need",
      "votes": null
    },
    {
      "id": "325215",
      "postDate": "05/08/2018 08:25:57",
      "content": "<p>What do you mean by \"image features\"? I doubt \"raw\" features from pretrained VGG can help here.</p>",
      "rawMarkdown": "What do you mean by \"image features\"? I doubt \"raw\" features from pretrained VGG can help here.",
      "votes": null
    },
    {
      "id": "325627",
      "postDate": "05/08/2018 16:38:04",
      "content": "<p>I did not manage to increase my score with pretrained image classification models yet. However, adding image resolution to my LightGBM model slightly increased my local CV score.</p>\n\n<p>Image saturation, contrast etc. are on my list of things to try. Intuitively, these should definitely help. </p>",
      "rawMarkdown": "I did not manage to increase my score with pretrained image classification models yet. However, adding image resolution to my LightGBM model slightly increased my local CV score.\n\nImage saturation, contrast etc. are on my list of things to try. Intuitively, these should definitely help.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 324763,
      "author_name": "ngxbac",
      "author_url": "",
      "post_date": "05/07/2018 23:34:15",
      "content": "<p>I have the same concern as you. The text features improved my model rapidly. Adding image features seems does not help.\nI am wondering if text feature is all we need</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 325215,
      "author_name": "ddanevskyi",
      "author_url": "",
      "post_date": "05/08/2018 08:25:57",
      "content": "<p>What do you mean by \"image features\"? I doubt \"raw\" features from pretrained VGG can help here.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 325627,
      "author_name": "bminixhofer",
      "author_url": "",
      "post_date": "05/08/2018 16:38:04",
      "content": "<p>I did not manage to increase my score with pretrained image classification models yet. However, adding image resolution to my LightGBM model slightly increased my local CV score.</p>\n\n<p>Image saturation, contrast etc. are on my list of things to try. Intuitively, these should definitely help. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "324571": "I tried several combinations of text features, grouped mean and variance features, and VGG image features, yet so far the group and image features do not help with the public LB score at all. Model tuning, on the other hand, increases the score slightly. \n\nFor text features, dimensionality reduction severely reduce the score on public LB and I have to keep all of them (around 20,000 so far). Is it because there are too many text features that the other features simply would not matter? Has anyone got better results by adding image features to a model with text features? It would be great if we can hear some good news on this!",
    "324763": "I have the same concern as you. The text features improved my model rapidly. Adding image features seems does not help.\nI am wondering if text feature is all we need",
    "325215": "What do you mean by \"image features\"? I doubt \"raw\" features from pretrained VGG can help here.",
    "325627": "I did not manage to increase my score with pretrained image classification models yet. However, adding image resolution to my LightGBM model slightly increased my local CV score.\n\nImage saturation, contrast etc. are on my list of things to try. Intuitively, these should definitely help."
  },
  "source": "meta"
}