{
  "id": 359866,
  "title": "Methods to score 0.812",
  "url": "/competitions/open-problems-multimodal/discussion/359866",
  "author_name": "Joseph Zhou",
  "post_date": "2022-10-14T02:25:11.129000",
  "votes": 23,
  "comment_count": 33,
  "views": 0,
  "content": "<p>For me, I use four models:</p>\n<ol>\n<li>MLP(both for CITEseq and Multiome)  -- 0.811</li>\n<li>MLP(both for CITEseq and Multiome but different structure and parameters) -- 0.811</li>\n<li>Catboost with MultiOutputRegressor(both for CITEseq and Multiome) -- 0.809</li>\n<li>Public one <a href=\"https://www.kaggle.com/code/pourchot/multiome-with-keras-ensemble\" target=\"_blank\">Multiome with Keras &amp; ensemble</a></li>\n</ol>\n<p>Now I am wondering how to score 0.812 with a single model, by feature engeering? different models?  or biological preprocessing for data?<br>\nLooking forward to your comments.</p>",
  "messages": [
    {
      "id": 1986315,
      "postDate": "2022-10-14T02:25:11.130Z",
      "content": "<p>For me, I use four models:</p>\n<ol>\n<li>MLP(both for CITEseq and Multiome)  -- 0.811</li>\n<li>MLP(both for CITEseq and Multiome but different structure and parameters) -- 0.811</li>\n<li>Catboost with MultiOutputRegressor(both for CITEseq and Multiome) -- 0.809</li>\n<li>Public one <a href=\"https://www.kaggle.com/code/pourchot/multiome-with-keras-ensemble\" target=\"_blank\">Multiome with Keras &amp; ensemble</a></li>\n</ol>\n<p>Now I am wondering how to score 0.812 with a single model, by feature engeering? different models?  or biological preprocessing for data?<br>\nLooking forward to your comments.</p>",
      "rawMarkdown": "For me, I use four models:\n1. MLP(both for CITEseq and Multiome)  -- 0.811\n2. MLP(both for CITEseq and Multiome but different structure and parameters) -- 0.811\n3. Catboost with MultiOutputRegressor(both for CITEseq and Multiome) -- 0.809\n4. Public one [Multiome with Keras & ensemble](https://www.kaggle.com/code/pourchot/multiome-with-keras-ensemble)\n\nNow I am wondering how to score 0.812 with a single model, by feature engeering? different models?  or biological preprocessing for data?\nLooking forward to your comments.",
      "votes": 23
    },
    {
      "id": 1986980,
      "postDate": "2022-10-14T12:28:53.490Z",
      "content": "<p>I was able to achieve 0.812 with a single model with my NN model, unfortunately no luck with ensembling till now :/ </p>",
      "rawMarkdown": "I was able to achieve 0.812 with a single model with my NN model, unfortunately no luck with ensembling till now :/ ",
      "votes": 1,
      "replies": [
        {
          "id": 1986998,
          "postDate": "2022-10-14T12:37:57.877Z",
          "content": "<p>Congrets, may I ask by feature engeering or some preprocessing?</p>",
          "rawMarkdown": "Congrets, may I ask by feature engeering or some preprocessing?",
          "votes": 2
        },
        {
          "id": 1987250,
          "postDate": "2022-10-14T15:46:44.580Z",
          "content": "<p>It's heavy on preprocessing </p>",
          "rawMarkdown": "It's heavy on preprocessing ",
          "votes": 3
        },
        {
          "id": 1987888,
          "postDate": "2022-10-14T21:50:20.853Z",
          "content": "<p>Do you use any preprocessing common for single-cell data?</p>",
          "rawMarkdown": "Do you use any preprocessing common for single-cell data?"
        },
        {
          "id": 1989099,
          "postDate": "2022-10-15T18:04:30.410Z",
          "content": "<p>No just truncatedSVD </p>",
          "rawMarkdown": "No just truncatedSVD "
        },
        {
          "id": 1989517,
          "postDate": "2022-10-16T03:36:34.523Z",
          "content": "<p>Nice we should Team with u :D .. Single model is really great between 👍 Care telling your CV scheme .. </p>",
          "rawMarkdown": "Nice we should Team with u :D .. Single model is really great between 👍 Care telling your CV scheme .. "
        },
        {
          "id": 1989589,
          "postDate": "2022-10-16T04:45:56.353Z",
          "content": "<p>Just for clarification, by single model i meant I have one model for multiome and one model for citeseq </p>",
          "rawMarkdown": "Just for clarification, by single model i meant I have one model for multiome and one model for citeseq ",
          "votes": 2
        },
        {
          "id": 1991101,
          "postDate": "2022-10-17T01:32:32.400Z",
          "content": "<p>Thats great .. single model with Multinome I was able to achieve that with a relatively simple MLP .. cite still an ensemble .</p>",
          "rawMarkdown": "Thats great .. single model with Multinome I was able to achieve that with a relatively simple MLP .. cite still an ensemble .",
          "votes": 1
        }
      ]
    },
    {
      "id": 1986975,
      "postDate": "2022-10-14T12:26:19.783Z",
      "content": "<p><a href=\"https://www.kaggle.com/takanashihumbert\" target=\"_blank\">@takanashihumbert</a> , thanks for sharing. Are you using preprocessing for any of those? Thanks</p>",
      "rawMarkdown": "@takanashihumbert , thanks for sharing. Are you using preprocessing for any of those? Thanks",
      "votes": 1,
      "replies": [
        {
          "id": 1986995,
          "postDate": "2022-10-14T12:36:06.077Z",
          "content": "<p>No special preprocessing, just TruncatedSVD.</p>",
          "rawMarkdown": "No special preprocessing, just TruncatedSVD.",
          "votes": 3
        }
      ]
    },
    {
      "id": 2028529,
      "postDate": "2022-11-14T03:11:52.443Z",
      "content": "<p>Thanks so much for sharing, very helpful!</p>",
      "rawMarkdown": "Thanks so much for sharing, very helpful!"
    },
    {
      "id": 2028502,
      "postDate": "2022-11-14T02:45:49.360Z",
      "content": "<p>Wow…it's really helpful</p>",
      "rawMarkdown": "Wow…it's really helpful"
    },
    {
      "id": 2007927,
      "postDate": "2022-10-28T15:22:03.187Z",
      "content": "<p>hi, im just rejoining this competition in these last few days. May I ask which MLP did you use? Is it from the scikit-learn package?</p>",
      "rawMarkdown": "hi, im just rejoining this competition in these last few days. May I ask which MLP did you use? Is it from the scikit-learn package?"
    },
    {
      "id": 2003046,
      "postDate": "2022-10-25T09:01:31.133Z",
      "content": "<p>Thanks for sharing your plan.I want to ask that your CatBoost model score is less than MLP,so are you still use CatBoost model in your plan?</p>",
      "rawMarkdown": "Thanks for sharing your plan.I want to ask that your CatBoost model score is less than MLP,so are you still use CatBoost model in your plan?",
      "replies": [
        {
          "id": 2005897,
          "postDate": "2022-10-27T10:06:29.817Z",
          "content": "<p>CatBoost one I used for ensemble which scored 0.813 has a weight only 0.1</p>",
          "rawMarkdown": "CatBoost one I used for ensemble which scored 0.813 has a weight only 0.1",
          "votes": 2
        },
        {
          "id": 2007891,
          "postDate": "2022-10-28T15:07:27.087Z",
          "content": "<p>How did you choose that weight if I may ask?. By trial/error checking the Public LB?</p>",
          "rawMarkdown": "How did you choose that weight if I may ask?. By trial/error checking the Public LB?"
        },
        {
          "id": 2015770,
          "postDate": "2022-11-03T14:34:33.107Z",
          "content": "<p>Example:</p>\n<pre><code># output No.1(mlp): \nLB: 0.813, weight: 0.4\n# output No.2(mlp): \nLB: 0.812, weight: 0.25\n# output No.3(mlp): \nLB: 0.812, weight: 0.25\n# output No.4(catboost): \nLB: 0.809, weight: 0.1\n</code></pre>\n<p>just choose by feeling😄</p>",
          "rawMarkdown": "Example:\n```\n# output No.1(mlp): \nLB: 0.813, weight: 0.4\n# output No.2(mlp): \nLB: 0.812, weight: 0.25\n# output No.3(mlp): \nLB: 0.812, weight: 0.25\n# output No.4(catboost): \nLB: 0.809, weight: 0.1\n\n```\njust choose by feeling😄",
          "votes": 1
        }
      ]
    },
    {
      "id": 1989541,
      "postDate": "2022-10-16T04:01:01.043Z",
      "content": "<p>osm work dear <a href=\"https://www.kaggle.com/stajdi\" target=\"_blank\">@stajdi</a> </p>",
      "rawMarkdown": "osm work dear @stajdi "
    },
    {
      "id": 1986647,
      "postDate": "2022-10-14T07:47:43.147Z",
      "content": "<p>you forget trial and error, probably is a big factor too.. just luck/overfitting</p>",
      "rawMarkdown": "you forget trial and error, probably is a big factor too.. just luck/overfitting"
    },
    {
      "id": 1986376,
      "postDate": "2022-10-14T04:30:55.380Z",
      "content": "<p>I found Catboost very slow, I couldn't get it to run on GPU with MultiOutput. Do you have the same issue ?</p>",
      "rawMarkdown": "I found Catboost very slow, I couldn't get it to run on GPU with MultiOutput. Do you have the same issue ?",
      "replies": [
        {
          "id": 1986387,
          "postDate": "2022-10-14T04:46:11.893Z",
          "content": "<p>Yes, MultiOutput can't run on GPU, so I use:</p>\n<pre><code>parameters = {\n    'learning_rate': 0.06, \n    'depth': 8, \n    'l2_leaf_reg': 2, \n    'loss_function': 'RMSE', \n    'task_type': 'GPU', \n    'iterations': 800,\n    'od_type': 'Iter', \n    'boosting_type': 'Plain', \n    'bootstrap_type': 'Bayesian',\n    'bagging_temperature': 0.3,\n    'allow_const_label': True, \n    'random_state': 42,\n    'verbose': 0\n}\n\nmodel = MultiOutputRegressor(CatBoostRegressor(**parameters))\n</code></pre>\n<p>And this result of score is better than MultiOutput on LB.</p>",
          "rawMarkdown": "Yes, MultiOutput can't run on GPU, so I use:\n```\nparameters = {\n    'learning_rate': 0.06, \n    'depth': 8, \n    'l2_leaf_reg': 2, \n    'loss_function': 'RMSE', \n    'task_type': 'GPU', \n    'iterations': 800,\n    'od_type': 'Iter', \n    'boosting_type': 'Plain', \n    'bootstrap_type': 'Bayesian',\n    'bagging_temperature': 0.3,\n    'allow_const_label': True, \n    'random_state': 42,\n    'verbose': 0\n}\n\nmodel = MultiOutputRegressor(CatBoostRegressor(**parameters))\n```\nAnd this result of score is better than MultiOutput on LB.",
          "votes": 7
        },
        {
          "id": 1986391,
          "postDate": "2022-10-14T04:55:39.500Z",
          "content": "<p>Thanks for sharing, I was using <br>\n                      'loss_function': 'MultiRMSE',<br>\n                      'eval_metric': 'MultiRMSE',<br>\n                      'task_type': 'CPU',</p>\n<p>And it takes 10 hours to run. </p>",
          "rawMarkdown": "Thanks for sharing, I was using \n                      'loss_function': 'MultiRMSE',\n                      'eval_metric': 'MultiRMSE',\n                      'task_type': 'CPU',\n\nAnd it takes 10 hours to run. ",
          "votes": 1
        },
        {
          "id": 2015721,
          "postDate": "2022-11-03T13:52:20.643Z",
          "content": "<p><a href=\"https://www.kaggle.com/takanashihumbert\" target=\"_blank\">@takanashihumbert</a> with MultiOutputRegressor I still can't run on the GPU</p>",
          "rawMarkdown": "@takanashihumbert with MultiOutputRegressor I still can't run on the GPU"
        },
        {
          "id": 2015756,
          "postDate": "2022-11-03T14:22:56.277Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 2015759,
          "postDate": "2022-11-03T14:23:55.377Z",
          "content": "<p><a href=\"https://www.kaggle.com/tonymarkchris\" target=\"_blank\">@tonymarkchris</a> what's wrong?</p>",
          "rawMarkdown": "@tonymarkchris what's wrong?"
        },
        {
          "id": 2015763,
          "postDate": "2022-11-03T14:27:26.277Z",
          "content": "<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F7330788%2F917bbb03b94cc25abbe7a80193f1dbfc%2FUntitled%204.png?generation=1667485642644462&amp;alt=media\" alt=\"\"></p>",
          "rawMarkdown": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F7330788%2F917bbb03b94cc25abbe7a80193f1dbfc%2FUntitled%204.png?generation=1667485642644462&alt=media)"
        },
        {
          "id": 2015785,
          "postDate": "2022-11-03T14:44:01.263Z",
          "content": "<p>The param 'loss_function' is 'RMSE' not 'MultiRMSE' right?</p>",
          "rawMarkdown": "The param 'loss_function' is 'RMSE' not 'MultiRMSE' right?"
        },
        {
          "id": 2015817,
          "postDate": "2022-11-03T15:03:45.123Z",
          "content": "<p>thanks for your help, my params <strong>'loss_function'</strong> is indeed MultiRMSE</p>",
          "rawMarkdown": "thanks for your help, my params **'loss_function'** is indeed MultiRMSE"
        },
        {
          "id": 2015943,
          "postDate": "2022-11-03T16:32:58.010Z",
          "content": "<p>Haha without GPU takes 1 day for me 1 fold :D </p>",
          "rawMarkdown": "Haha without GPU takes 1 day for me 1 fold :D "
        },
        {
          "id": 2016461,
          "postDate": "2022-11-04T03:12:47.520Z",
          "content": "<p>May I ask why you guys are using catboost among GBDT?</p>",
          "rawMarkdown": "May I ask why you guys are using catboost among GBDT?"
        }
      ]
    },
    {
      "id": 2029793,
      "postDate": "2022-11-14T23:33:33.463Z",
      "content": "<p>thanks you for share Its help me a lot.</p>",
      "rawMarkdown": "thanks you for share Its help me a lot."
    },
    {
      "id": 2028898,
      "postDate": "2022-11-14T10:00:52.373Z",
      "content": "<p>This is GREAT!!! Thank you!!</p>",
      "rawMarkdown": "This is GREAT!!! Thank you!!"
    },
    {
      "id": 2028645,
      "postDate": "2022-11-14T06:07:58.397Z",
      "content": "<p>It's very helpful! Thank you! :)</p>",
      "rawMarkdown": "It's very helpful! Thank you! :)"
    }
  ],
  "comments": [
    {
      "id": 1986980,
      "author_name": "tarick.morty",
      "author_url": "",
      "post_date": "2022-10-14T12:28:53.490000",
      "content": "<p>I was able to achieve 0.812 with a single model with my NN model, unfortunately no luck with ensembling till now :/ </p>",
      "votes": 1,
      "replies": [
        {
          "id": 1986998,
          "author_name": "Joseph Zhou",
          "author_url": "",
          "post_date": "2022-10-14T12:37:57.877000",
          "content": "<p>Congrets, may I ask by feature engeering or some preprocessing?</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1987250,
          "author_name": "tarick.morty",
          "author_url": "",
          "post_date": "2022-10-14T15:46:44.580000",
          "content": "<p>It's heavy on preprocessing </p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 1987888,
          "author_name": "Hiram Coria 🧬",
          "author_url": "",
          "post_date": "2022-10-14T21:50:20.853000",
          "content": "<p>Do you use any preprocessing common for single-cell data?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1989099,
          "author_name": "tarick.morty",
          "author_url": "",
          "post_date": "2022-10-15T18:04:30.410000",
          "content": "<p>No just truncatedSVD </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1989517,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2022-10-16T03:36:34.523000",
          "content": "<p>Nice we should Team with u :D .. Single model is really great between 👍 Care telling your CV scheme .. </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1989589,
          "author_name": "tarick.morty",
          "author_url": "",
          "post_date": "2022-10-16T04:45:56.353000",
          "content": "<p>Just for clarification, by single model i meant I have one model for multiome and one model for citeseq </p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1991101,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2022-10-17T01:32:32.400000",
          "content": "<p>Thats great .. single model with Multinome I was able to achieve that with a relatively simple MLP .. cite still an ensemble .</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1986975,
      "author_name": "stajdi",
      "author_url": "",
      "post_date": "2022-10-14T12:26:19.783000",
      "content": "<p><a href=\"https://www.kaggle.com/takanashihumbert\" target=\"_blank\">@takanashihumbert</a> , thanks for sharing. Are you using preprocessing for any of those? Thanks</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1986995,
          "author_name": "Joseph Zhou",
          "author_url": "",
          "post_date": "2022-10-14T12:36:06.077000",
          "content": "<p>No special preprocessing, just TruncatedSVD.</p>",
          "votes": 3,
          "replies": []
        }
      ]
    },
    {
      "id": 2028529,
      "author_name": "",
      "author_url": "",
      "post_date": "2022-11-14T03:11:52.443000",
      "content": "<p>Thanks so much for sharing, very helpful!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2028502,
      "author_name": "",
      "author_url": "",
      "post_date": "2022-11-14T02:45:49.360000",
      "content": "<p>Wow…it's really helpful</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2007927,
      "author_name": "Leandro Destefani",
      "author_url": "",
      "post_date": "2022-10-28T15:22:03.187000",
      "content": "<p>hi, im just rejoining this competition in these last few days. May I ask which MLP did you use? Is it from the scikit-learn package?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2003046,
      "author_name": "ZhangYuan2000",
      "author_url": "",
      "post_date": "2022-10-25T09:01:31.133000",
      "content": "<p>Thanks for sharing your plan.I want to ask that your CatBoost model score is less than MLP,so are you still use CatBoost model in your plan?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 2005897,
          "author_name": "Joseph Zhou",
          "author_url": "",
          "post_date": "2022-10-27T10:06:29.817000",
          "content": "<p>CatBoost one I used for ensemble which scored 0.813 has a weight only 0.1</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 2007891,
          "author_name": "SergioMiguelM",
          "author_url": "",
          "post_date": "2022-10-28T15:07:27.087000",
          "content": "<p>How did you choose that weight if I may ask?. By trial/error checking the Public LB?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2015770,
          "author_name": "Joseph Zhou",
          "author_url": "",
          "post_date": "2022-11-03T14:34:33.107000",
          "content": "<p>Example:</p>\n<pre><code># output No.1(mlp): \nLB: 0.813, weight: 0.4\n# output No.2(mlp): \nLB: 0.812, weight: 0.25\n# output No.3(mlp): \nLB: 0.812, weight: 0.25\n# output No.4(catboost): \nLB: 0.809, weight: 0.1\n</code></pre>\n<p>just choose by feeling😄</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1989541,
      "author_name": "AbdulHadi12",
      "author_url": "",
      "post_date": "2022-10-16T04:01:01.043000",
      "content": "<p>osm work dear <a href=\"https://www.kaggle.com/stajdi\" target=\"_blank\">@stajdi</a> </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1986647,
      "author_name": "pinouche",
      "author_url": "",
      "post_date": "2022-10-14T07:47:43.147000",
      "content": "<p>you forget trial and error, probably is a big factor too.. just luck/overfitting</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1986376,
      "author_name": "Elias",
      "author_url": "",
      "post_date": "2022-10-14T04:30:55.380000",
      "content": "<p>I found Catboost very slow, I couldn't get it to run on GPU with MultiOutput. Do you have the same issue ?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1986387,
          "author_name": "Joseph Zhou",
          "author_url": "",
          "post_date": "2022-10-14T04:46:11.893000",
          "content": "<p>Yes, MultiOutput can't run on GPU, so I use:</p>\n<pre><code>parameters = {\n    'learning_rate': 0.06, \n    'depth': 8, \n    'l2_leaf_reg': 2, \n    'loss_function': 'RMSE', \n    'task_type': 'GPU', \n    'iterations': 800,\n    'od_type': 'Iter', \n    'boosting_type': 'Plain', \n    'bootstrap_type': 'Bayesian',\n    'bagging_temperature': 0.3,\n    'allow_const_label': True, \n    'random_state': 42,\n    'verbose': 0\n}\n\nmodel = MultiOutputRegressor(CatBoostRegressor(**parameters))\n</code></pre>\n<p>And this result of score is better than MultiOutput on LB.</p>",
          "votes": 7,
          "replies": []
        },
        {
          "id": 1986391,
          "author_name": "Elias",
          "author_url": "",
          "post_date": "2022-10-14T04:55:39.500000",
          "content": "<p>Thanks for sharing, I was using <br>\n                      'loss_function': 'MultiRMSE',<br>\n                      'eval_metric': 'MultiRMSE',<br>\n                      'task_type': 'CPU',</p>\n<p>And it takes 10 hours to run. </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 2015721,
          "author_name": "DJ_Xia",
          "author_url": "",
          "post_date": "2022-11-03T13:52:20.643000",
          "content": "<p><a href=\"https://www.kaggle.com/takanashihumbert\" target=\"_blank\">@takanashihumbert</a> with MultiOutputRegressor I still can't run on the GPU</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2015756,
          "author_name": "",
          "author_url": "",
          "post_date": "2022-11-03T14:22:56.277000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2015759,
          "author_name": "Joseph Zhou",
          "author_url": "",
          "post_date": "2022-11-03T14:23:55.377000",
          "content": "<p><a href=\"https://www.kaggle.com/tonymarkchris\" target=\"_blank\">@tonymarkchris</a> what's wrong?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2015763,
          "author_name": "DJ_Xia",
          "author_url": "",
          "post_date": "2022-11-03T14:27:26.277000",
          "content": "<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F7330788%2F917bbb03b94cc25abbe7a80193f1dbfc%2FUntitled%204.png?generation=1667485642644462&amp;alt=media\" alt=\"\"></p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2015785,
          "author_name": "Joseph Zhou",
          "author_url": "",
          "post_date": "2022-11-03T14:44:01.263000",
          "content": "<p>The param 'loss_function' is 'RMSE' not 'MultiRMSE' right?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2015817,
          "author_name": "DJ_Xia",
          "author_url": "",
          "post_date": "2022-11-03T15:03:45.123000",
          "content": "<p>thanks for your help, my params <strong>'loss_function'</strong> is indeed MultiRMSE</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2015943,
          "author_name": "Gaurav Rawat",
          "author_url": "",
          "post_date": "2022-11-03T16:32:58.010000",
          "content": "<p>Haha without GPU takes 1 day for me 1 fold :D </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2016461,
          "author_name": "wakaka",
          "author_url": "",
          "post_date": "2022-11-04T03:12:47.520000",
          "content": "<p>May I ask why you guys are using catboost among GBDT?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2029793,
      "author_name": "",
      "author_url": "",
      "post_date": "2022-11-14T23:33:33.463000",
      "content": "<p>thanks you for share Its help me a lot.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2028898,
      "author_name": "",
      "author_url": "",
      "post_date": "2022-11-14T10:00:52.373000",
      "content": "<p>This is GREAT!!! Thank you!!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 2028645,
      "author_name": "Yanchen",
      "author_url": "",
      "post_date": "2022-11-14T06:07:58.397000",
      "content": "<p>It's very helpful! Thank you! :)</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1986315": "For me, I use four models:\n1. MLP(both for CITEseq and Multiome)  -- 0.811\n2. MLP(both for CITEseq and Multiome but different structure and parameters) -- 0.811\n3. Catboost with MultiOutputRegressor(both for CITEseq and Multiome) -- 0.809\n4. Public one [Multiome with Keras & ensemble](https://www.kaggle.com/code/pourchot/multiome-with-keras-ensemble)\n\nNow I am wondering how to score 0.812 with a single model, by feature engeering? different models?  or biological preprocessing for data?\nLooking forward to your comments.",
    "1986980": "I was able to achieve 0.812 with a single model with my NN model, unfortunately no luck with ensembling till now :/ ",
    "1986975": "@takanashihumbert , thanks for sharing. Are you using preprocessing for any of those? Thanks",
    "2028529": "Thanks so much for sharing, very helpful!",
    "2028502": "Wow…it's really helpful",
    "2007927": "hi, im just rejoining this competition in these last few days. May I ask which MLP did you use? Is it from the scikit-learn package?",
    "2003046": "Thanks for sharing your plan.I want to ask that your CatBoost model score is less than MLP,so are you still use CatBoost model in your plan?",
    "1989541": "osm work dear @stajdi ",
    "1986647": "you forget trial and error, probably is a big factor too.. just luck/overfitting",
    "1986376": "I found Catboost very slow, I couldn't get it to run on GPU with MultiOutput. Do you have the same issue ?",
    "2029793": "thanks you for share Its help me a lot.",
    "2028898": "This is GREAT!!! Thank you!!",
    "2028645": "It's very helpful! Thank you! :)"
  }
}