{
  "id": 313749,
  "title": "Single Model Traditional Machine Learning Approaches",
  "url": "/competitions/h-and-m-personalized-fashion-recommendations/discussion/313749",
  "author_name": "",
  "post_date": "2022-03-18T16:29:04.827507700Z",
  "votes": 48,
  "comment_count": 2,
  "views": 0,
  "content": "<p>As far as I see, most of the top solutions are either rule based or some modelling on pre-selected candidate recommendations. I have worked hard to get my pure PyTorch model to the same level but it has only 0.0232 validation score and 0.0181 LB. I believe the main challenge is the high dimensionality and the sparsity caused by that. I share my neural network solution here: <a href=\"https://www.kaggle.com/aerdem4/h-m-pure-pytorch-baseline\" target=\"_blank\">https://www.kaggle.com/aerdem4/h-m-pure-pytorch-baseline</a></p>\n<p>I am looking forward to see if other Kagglers can come up with more tricks to improve this solution and how far a single model approach can go. I hope this notebook can help as a baseline. And I am happy to discuss your modelling ideas with other kind of models.</p>",
  "messages": [
    {
      "id": "1728180",
      "postDate": "03/18/2022 16:29:04",
      "content": "<p>As far as I see, most of the top solutions are either rule based or some modelling on pre-selected candidate recommendations. I have worked hard to get my pure PyTorch model to the same level but it has only 0.0232 validation score and 0.0181 LB. I believe the main challenge is the high dimensionality and the sparsity caused by that. I share my neural network solution here: <a href=\"https://www.kaggle.com/aerdem4/h-m-pure-pytorch-baseline\" target=\"_blank\">https://www.kaggle.com/aerdem4/h-m-pure-pytorch-baseline</a></p>\n<p>I am looking forward to see if other Kagglers can come up with more tricks to improve this solution and how far a single model approach can go. I hope this notebook can help as a baseline. And I am happy to discuss your modelling ideas with other kind of models.</p>",
      "rawMarkdown": "As far as I see, most of the top solutions are either rule based or some modelling on pre-selected candidate recommendations. I have worked hard to get my pure PyTorch model to the same level but it has only 0.0232 validation score and 0.0181 LB. I believe the main challenge is the high dimensionality and the sparsity caused by that. I share my neural network solution here: https://www.kaggle.com/aerdem4/h-m-pure-pytorch-baseline\n\nI am looking forward to see if other Kagglers can come up with more tricks to improve this solution and how far a single model approach can go. I hope this notebook can help as a baseline. And I am happy to discuss your modelling ideas with other kind of models.",
      "votes": null
    },
    {
      "id": "1729608",
      "postDate": "03/20/2022 09:23:55",
      "content": "<p>Nice work! <br>\nI try to modify the model but LB is only 0.0198.😂</p>",
      "rawMarkdown": "Nice work! \nI try to modify the model but LB is only 0.0198.😂",
      "votes": null
    },
    {
      "id": "1764094",
      "postDate": "04/22/2022 06:35:56",
      "content": "<p>You can also initialize your article embeddings with such unsupervised approaches instead of training them from random initialization: <a href=\"https://www.kaggle.com/code/aerdem4/h-m-rapids-article2vec\" target=\"_blank\">https://www.kaggle.com/code/aerdem4/h-m-rapids-article2vec</a></p>",
      "rawMarkdown": "You can also initialize your article embeddings with such unsupervised approaches instead of training them from random initialization: https://www.kaggle.com/code/aerdem4/h-m-rapids-article2vec",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1729608,
      "author_name": "yangranran",
      "author_url": "",
      "post_date": "03/20/2022 09:23:55",
      "content": "<p>Nice work! <br>\nI try to modify the model but LB is only 0.0198.😂</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1764094,
      "author_name": "aerdem4",
      "author_url": "",
      "post_date": "04/22/2022 06:35:56",
      "content": "<p>You can also initialize your article embeddings with such unsupervised approaches instead of training them from random initialization: <a href=\"https://www.kaggle.com/code/aerdem4/h-m-rapids-article2vec\" target=\"_blank\">https://www.kaggle.com/code/aerdem4/h-m-rapids-article2vec</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1728180": "As far as I see, most of the top solutions are either rule based or some modelling on pre-selected candidate recommendations. I have worked hard to get my pure PyTorch model to the same level but it has only 0.0232 validation score and 0.0181 LB. I believe the main challenge is the high dimensionality and the sparsity caused by that. I share my neural network solution here: https://www.kaggle.com/aerdem4/h-m-pure-pytorch-baseline\n\nI am looking forward to see if other Kagglers can come up with more tricks to improve this solution and how far a single model approach can go. I hope this notebook can help as a baseline. And I am happy to discuss your modelling ideas with other kind of models.",
    "1729608": "Nice work! \nI try to modify the model but LB is only 0.0198.😂",
    "1764094": "You can also initialize your article embeddings with such unsupervised approaches instead of training them from random initialization: https://www.kaggle.com/code/aerdem4/h-m-rapids-article2vec"
  },
  "source": "meta"
}