{
  "id": 285209,
  "title": "Our solution (PaRaNA) is now public!",
  "url": "/competitions/kddbr-2021/discussion/285209",
  "author_name": "Bruno Klaus",
  "post_date": "2021-11-03T19:00:32.122000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>You can find it here:<br>\n<a href=\"https://www.kaggle.com/isonettv/kdd-br-parana-solution\" target=\"_blank\">https://www.kaggle.com/isonettv/kdd-br-parana-solution</a></p>\n<p>Our model is pretty simple. We feed our input data to the edge convolutional operator.  After a few layers of EdgeConv, we obtain our final edge prediction using a Multilayer Perceptron on pairs of vertices. </p>\n<p>We use a fixed threshold for the final classification, and our loss function uses the same weighting scheme  as <a href=\"https://arxiv.org/abs/1906.01227https://arxiv.org/abs/1906.01227\" target=\"_blank\">https://arxiv.org/abs/1906.01227https://arxiv.org/abs/1906.01227</a> for the loss function. We didn't have too much time available. If we did, we'd probably created a methods that chooses the best threshold w.r.t. the validation set, and also test more hyperparameter combinations (convolutions, depth, hidden size and so on).</p>\n<p>The model is implemented using Pytorch Geometric, so we had to convert the existing data using this notebook:<br>\n<a href=\"https://www.kaggle.com/isonettv/converting-kddbr-2021-data-to-pytorch-files\" target=\"_blank\">https://www.kaggle.com/isonettv/converting-kddbr-2021-data-to-pytorch-files</a></p>",
  "messages": [
    {
      "id": 1569851,
      "postDate": "2021-11-03T19:00:32.123Z",
      "content": "<p>You can find it here:<br>\n<a href=\"https://www.kaggle.com/isonettv/kdd-br-parana-solution\" target=\"_blank\">https://www.kaggle.com/isonettv/kdd-br-parana-solution</a></p>\n<p>Our model is pretty simple. We feed our input data to the edge convolutional operator.  After a few layers of EdgeConv, we obtain our final edge prediction using a Multilayer Perceptron on pairs of vertices. </p>\n<p>We use a fixed threshold for the final classification, and our loss function uses the same weighting scheme  as <a href=\"https://arxiv.org/abs/1906.01227https://arxiv.org/abs/1906.01227\" target=\"_blank\">https://arxiv.org/abs/1906.01227https://arxiv.org/abs/1906.01227</a> for the loss function. We didn't have too much time available. If we did, we'd probably created a methods that chooses the best threshold w.r.t. the validation set, and also test more hyperparameter combinations (convolutions, depth, hidden size and so on).</p>\n<p>The model is implemented using Pytorch Geometric, so we had to convert the existing data using this notebook:<br>\n<a href=\"https://www.kaggle.com/isonettv/converting-kddbr-2021-data-to-pytorch-files\" target=\"_blank\">https://www.kaggle.com/isonettv/converting-kddbr-2021-data-to-pytorch-files</a></p>",
      "rawMarkdown": "You can find it here:\nhttps://www.kaggle.com/isonettv/kdd-br-parana-solution\n\nOur model is pretty simple. We feed our input data to the edge convolutional operator.  After a few layers of EdgeConv, we obtain our final edge prediction using a Multilayer Perceptron on pairs of vertices. \n\nWe use a fixed threshold for the final classification, and our loss function uses the same weighting scheme  as https://arxiv.org/abs/1906.01227https://arxiv.org/abs/1906.01227 for the loss function. We didn't have too much time available. If we did, we'd probably created a methods that chooses the best threshold w.r.t. the validation set, and also test more hyperparameter combinations (convolutions, depth, hidden size and so on).\n\nThe model is implemented using Pytorch Geometric, so we had to convert the existing data using this notebook:\nhttps://www.kaggle.com/isonettv/converting-kddbr-2021-data-to-pytorch-files",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1569851": "You can find it here:\nhttps://www.kaggle.com/isonettv/kdd-br-parana-solution\n\nOur model is pretty simple. We feed our input data to the edge convolutional operator.  After a few layers of EdgeConv, we obtain our final edge prediction using a Multilayer Perceptron on pairs of vertices. \n\nWe use a fixed threshold for the final classification, and our loss function uses the same weighting scheme  as https://arxiv.org/abs/1906.01227https://arxiv.org/abs/1906.01227 for the loss function. We didn't have too much time available. If we did, we'd probably created a methods that chooses the best threshold w.r.t. the validation set, and also test more hyperparameter combinations (convolutions, depth, hidden size and so on).\n\nThe model is implemented using Pytorch Geometric, so we had to convert the existing data using this notebook:\nhttps://www.kaggle.com/isonettv/converting-kddbr-2021-data-to-pytorch-files"
  }
}