{
  "id": 369680,
  "title": "What is your best score of pure co-visit-mat method?",
  "url": "/competitions/otto-recommender-system/discussion/369680",
  "author_name": "",
  "post_date": "2022-12-01T01:35:38.348162300Z",
  "votes": 16,
  "comment_count": 8,
  "views": 0,
  "content": "<p>Without any other reranker models, has the co-visit-mat method reached its limitation?</p>\n<p>For me, I'm still using the old pipeline and get 0.586 score on LB. Both boosting-tree ranker and NN ranker didn't help.</p>",
  "messages": [
    {
      "id": "2050767",
      "postDate": "12/01/2022 01:35:38",
      "content": "<p>Without any other reranker models, has the co-visit-mat method reached its limitation?</p>\n<p>For me, I'm still using the old pipeline and get 0.586 score on LB. Both boosting-tree ranker and NN ranker didn't help.</p>",
      "rawMarkdown": "Without any other reranker models, has the co-visit-mat method reached its limitation?\n\nFor me, I'm still using the old pipeline and get 0.586 score on LB. Both boosting-tree ranker and NN ranker didn't help.",
      "votes": null
    },
    {
      "id": "2051163",
      "postDate": "12/01/2022 09:00:02",
      "content": "<p>That's… amazing!!!!</p>\n<p>Really neat to know that much can be squeezed out of those co-visiatation matrices. I haven't had a chance to play with them too much, so am on what Chris shared, which brings me to where I am currently on the LB (using output from co-visitation matrices with ensembling, without ensembling it is -0.001)</p>",
      "rawMarkdown": "That's... amazing!!!!\n\nReally neat to know that much can be squeezed out of those co-visiatation matrices. I haven't had a chance to play with them too much, so am on what Chris shared, which brings me to where I am currently on the LB (using output from co-visitation matrices with ensembling, without ensembling it is -0.001)",
      "votes": null
    },
    {
      "id": "2051411",
      "postDate": "12/01/2022 11:51:52",
      "content": "<p>if I may ask, how much time does the pipeline take?</p>",
      "rawMarkdown": "if I may ask, how much time does the pipeline take?",
      "votes": null
    },
    {
      "id": "2052154",
      "postDate": "12/02/2022 01:03:35",
      "content": "<p>I have a better-optimized numba pipeline similar to my public notebook, which spends 2min to generate one co-visitation matrix and ~10min in total including data IO.</p>",
      "rawMarkdown": "I have a better-optimized numba pipeline similar to my public notebook, which spends 2min to generate one co-visitation matrix and ~10min in total including data IO.",
      "votes": null
    },
    {
      "id": "2052155",
      "postDate": "12/02/2022 01:05:40",
      "content": "<p>I think other top competitors might use trained-model to rank the recall result. But I also think we have not reached the limitation of pure co-visitation matrix.</p>",
      "rawMarkdown": "I think other top competitors might use trained-model to rank the recall result. But I also think we have not reached the limitation of pure co-visitation matrix.",
      "votes": null
    },
    {
      "id": "2052216",
      "postDate": "12/02/2022 02:38:16",
      "content": "<p>that's crazy fast, your notebook inspired me to use numba for this competition, and I will stick with that</p>",
      "rawMarkdown": "that's crazy fast, your notebook inspired me to use numba for this competition, and I will stick with that",
      "votes": null
    },
    {
      "id": "2052219",
      "postDate": "12/02/2022 02:43:26",
      "content": "<p>Pure covisitation method achieved 0.586 is amazing …   I still word on my own recall + rank pipeline  😂😂</p>",
      "rawMarkdown": "Pure covisitation method achieved 0.586 is amazing ...   I still word on my own recall + rank pipeline  😂😂",
      "votes": null
    },
    {
      "id": "2052222",
      "postDate": "12/02/2022 02:48:05",
      "content": "<p>Actually, co-vistation also includes recall and rank, because the number of co-visited candidates are usually more than 20. We need design a tricky weighting rule the select the top-20, that is, so-called \"rerank\".</p>",
      "rawMarkdown": "Actually, co-vistation also includes recall and rank, because the number of co-visited candidates are usually more than 20. We need design a tricky weighting rule the select the top-20, that is, so-called \"rerank\".",
      "votes": null
    },
    {
      "id": "2052225",
      "postDate": "12/02/2022 02:54:26",
      "content": "<p>😃Glad my notebook can help you. Because I only have 24G GPU and <code>cudf</code> methods usually meet CUDA memory errors, so I keep using <code>numba</code> which can be as fast as <code>cudf</code> on my 16-core CPU machine.</p>",
      "rawMarkdown": "😃Glad my notebook can help you. Because I only have 24G GPU and `cudf` methods usually meet CUDA memory errors, so I keep using `numba` which can be as fast as `cudf` on my 16-core CPU machine.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2051163,
      "author_name": "radek1",
      "author_url": "",
      "post_date": "12/01/2022 09:00:02",
      "content": "<p>That's… amazing!!!!</p>\n<p>Really neat to know that much can be squeezed out of those co-visiatation matrices. I haven't had a chance to play with them too much, so am on what Chris shared, which brings me to where I am currently on the LB (using output from co-visitation matrices with ensembling, without ensembling it is -0.001)</p>",
      "votes": null,
      "replies": [
        {
          "id": 2052155,
          "author_name": "carnozhao",
          "author_url": "",
          "post_date": "12/02/2022 01:05:40",
          "content": "<p>I think other top competitors might use trained-model to rank the recall result. But I also think we have not reached the limitation of pure co-visitation matrix.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2051411,
      "author_name": "nikhilmishradev",
      "author_url": "",
      "post_date": "12/01/2022 11:51:52",
      "content": "<p>if I may ask, how much time does the pipeline take?</p>",
      "votes": null,
      "replies": [
        {
          "id": 2052154,
          "author_name": "carnozhao",
          "author_url": "",
          "post_date": "12/02/2022 01:03:35",
          "content": "<p>I have a better-optimized numba pipeline similar to my public notebook, which spends 2min to generate one co-visitation matrix and ~10min in total including data IO.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 2052216,
          "author_name": "nikhilmishradev",
          "author_url": "",
          "post_date": "12/02/2022 02:38:16",
          "content": "<p>that's crazy fast, your notebook inspired me to use numba for this competition, and I will stick with that</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 2052225,
          "author_name": "carnozhao",
          "author_url": "",
          "post_date": "12/02/2022 02:54:26",
          "content": "<p>😃Glad my notebook can help you. Because I only have 24G GPU and <code>cudf</code> methods usually meet CUDA memory errors, so I keep using <code>numba</code> which can be as fast as <code>cudf</code> on my 16-core CPU machine.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2052219,
      "author_name": "evilpsycho42",
      "author_url": "",
      "post_date": "12/02/2022 02:43:26",
      "content": "<p>Pure covisitation method achieved 0.586 is amazing …   I still word on my own recall + rank pipeline  😂😂</p>",
      "votes": null,
      "replies": [
        {
          "id": 2052222,
          "author_name": "carnozhao",
          "author_url": "",
          "post_date": "12/02/2022 02:48:05",
          "content": "<p>Actually, co-vistation also includes recall and rank, because the number of co-visited candidates are usually more than 20. We need design a tricky weighting rule the select the top-20, that is, so-called \"rerank\".</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2050767": "Without any other reranker models, has the co-visit-mat method reached its limitation?\n\nFor me, I'm still using the old pipeline and get 0.586 score on LB. Both boosting-tree ranker and NN ranker didn't help.",
    "2051163": "That's... amazing!!!!\n\nReally neat to know that much can be squeezed out of those co-visiatation matrices. I haven't had a chance to play with them too much, so am on what Chris shared, which brings me to where I am currently on the LB (using output from co-visitation matrices with ensembling, without ensembling it is -0.001)",
    "2051411": "if I may ask, how much time does the pipeline take?",
    "2052154": "I have a better-optimized numba pipeline similar to my public notebook, which spends 2min to generate one co-visitation matrix and ~10min in total including data IO.",
    "2052155": "I think other top competitors might use trained-model to rank the recall result. But I also think we have not reached the limitation of pure co-visitation matrix.",
    "2052216": "that's crazy fast, your notebook inspired me to use numba for this competition, and I will stick with that",
    "2052219": "Pure covisitation method achieved 0.586 is amazing ...   I still word on my own recall + rank pipeline  😂😂",
    "2052222": "Actually, co-vistation also includes recall and rank, because the number of co-visited candidates are usually more than 20. We need design a tricky weighting rule the select the top-20, that is, so-called \"rerank\".",
    "2052225": "😃Glad my notebook can help you. Because I only have 24G GPU and `cudf` methods usually meet CUDA memory errors, so I keep using `numba` which can be as fast as `cudf` on my 16-core CPU machine."
  },
  "source": "meta"
}