{
  "id": 364022,
  "title": "Guidance on what algos to try from one of the authors of Microsoft Recommenders repo",
  "url": "/competitions/otto-recommender-system/discussion/364022",
  "author_name": "Miguel Fierro",
  "post_date": "2022-11-04T06:45:17.020000",
  "votes": 15,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi folks, I'm Miguel Fierro and I'm one of the authors of <a href=\"https://github.com/microsoft/recommenders/\" target=\"_blank\">Microsoft Recommenders repo</a>. </p>\n<p>Today I've <a href=\"https://www.linkedin.com/posts/miguelgfierro_ai-machinelearning-kaggle-activity-6994184636822130688-q1Cr?utm_source=share&amp;utm_medium=member_desktop\" target=\"_blank\">posted on LinkedIn</a> some advice on what algorithms to use and how to approach the competition. </p>\n<p>Here are some algos that you can start with:</p>\n<ul>\n<li>BPR: good baseline that uses pair-wise ranking.</li>\n<li>LightGCN: graphical nets include the interaction in the embeddings</li>\n<li>SASRec &amp; SSEPT: transformer-based recommenders.</li>\n<li>SLi-Rec: Sequential recommender based on a modified LSTM</li>\n<li>Other sequential recos like NextItNet, GRU4Rec, Caser, A2SVD, or SUM.</li>\n</ul>\n<p>Some other ideas:<br>\nYou can try to modify the loss function of each of the algos to be as similar as possible to the evaluation metric of the challenge.</p>\n<p>BONUS idea only for Kagglers: <br>\nUse LightGBM, it can solve 95% of problems in life :-) For example, you can generate embeddings with some of the algos listed above and then use LightGBM on top.</p>\n<p>Good luck!</p>",
  "messages": [
    {
      "id": 2016631,
      "postDate": "2022-11-04T06:45:17.020Z",
      "content": "<p>Hi folks, I'm Miguel Fierro and I'm one of the authors of <a href=\"https://github.com/microsoft/recommenders/\" target=\"_blank\">Microsoft Recommenders repo</a>. </p>\n<p>Today I've <a href=\"https://www.linkedin.com/posts/miguelgfierro_ai-machinelearning-kaggle-activity-6994184636822130688-q1Cr?utm_source=share&amp;utm_medium=member_desktop\" target=\"_blank\">posted on LinkedIn</a> some advice on what algorithms to use and how to approach the competition. </p>\n<p>Here are some algos that you can start with:</p>\n<ul>\n<li>BPR: good baseline that uses pair-wise ranking.</li>\n<li>LightGCN: graphical nets include the interaction in the embeddings</li>\n<li>SASRec &amp; SSEPT: transformer-based recommenders.</li>\n<li>SLi-Rec: Sequential recommender based on a modified LSTM</li>\n<li>Other sequential recos like NextItNet, GRU4Rec, Caser, A2SVD, or SUM.</li>\n</ul>\n<p>Some other ideas:<br>\nYou can try to modify the loss function of each of the algos to be as similar as possible to the evaluation metric of the challenge.</p>\n<p>BONUS idea only for Kagglers: <br>\nUse LightGBM, it can solve 95% of problems in life :-) For example, you can generate embeddings with some of the algos listed above and then use LightGBM on top.</p>\n<p>Good luck!</p>",
      "rawMarkdown": "Hi folks, I'm Miguel Fierro and I'm one of the authors of [Microsoft Recommenders repo](https://github.com/microsoft/recommenders/). \n\nToday I've [posted on LinkedIn](https://www.linkedin.com/posts/miguelgfierro_ai-machinelearning-kaggle-activity-6994184636822130688-q1Cr?utm_source=share&utm_medium=member_desktop) some advice on what algorithms to use and how to approach the competition. \n\nHere are some algos that you can start with:\n- BPR: good baseline that uses pair-wise ranking.\n- LightGCN: graphical nets include the interaction in the embeddings\n- SASRec & SSEPT: transformer-based recommenders.\n- SLi-Rec: Sequential recommender based on a modified LSTM\n- Other sequential recos like NextItNet, GRU4Rec, Caser, A2SVD, or SUM.\n\nSome other ideas:\nYou can try to modify the loss function of each of the algos to be as similar as possible to the evaluation metric of the challenge.\n\nBONUS idea only for Kagglers: \nUse LightGBM, it can solve 95% of problems in life :-) For example, you can generate embeddings with some of the algos listed above and then use LightGBM on top.\n\nGood luck!",
      "votes": 15
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2016631": "Hi folks, I'm Miguel Fierro and I'm one of the authors of [Microsoft Recommenders repo](https://github.com/microsoft/recommenders/). \n\nToday I've [posted on LinkedIn](https://www.linkedin.com/posts/miguelgfierro_ai-machinelearning-kaggle-activity-6994184636822130688-q1Cr?utm_source=share&utm_medium=member_desktop) some advice on what algorithms to use and how to approach the competition. \n\nHere are some algos that you can start with:\n- BPR: good baseline that uses pair-wise ranking.\n- LightGCN: graphical nets include the interaction in the embeddings\n- SASRec & SSEPT: transformer-based recommenders.\n- SLi-Rec: Sequential recommender based on a modified LSTM\n- Other sequential recos like NextItNet, GRU4Rec, Caser, A2SVD, or SUM.\n\nSome other ideas:\nYou can try to modify the loss function of each of the algos to be as similar as possible to the evaluation metric of the challenge.\n\nBONUS idea only for Kagglers: \nUse LightGBM, it can solve 95% of problems in life :-) For example, you can generate embeddings with some of the algos listed above and then use LightGBM on top.\n\nGood luck!"
  }
}