{
  "id": 366394,
  "title": "my summary(53th)",
  "url": "/competitions/open-problems-multimodal/discussion/366394",
  "author_name": "",
  "post_date": "2022-11-16T02:26:49.860973600Z",
  "votes": 8,
  "comment_count": 2,
  "views": 0,
  "content": "<p>we tried many models and some model or ensembled model are useful,the first month,we mainly focused on tree models:<strong>lightgbm and catboost</strong>, lightgbm not support multi targets and must use the** sklearn.preprocess.multioutput**, and running very slowly, and so does catboost, and the results of tree models are basically about 0.808/0.809 , which is lower than MLP. Actually,tree models are useful and good at the tabular dataset in most case. But in this competition,it's not easy to do some feature engineering due to gene features and we don't know the what the gene features represent for. last two months,we mainly use *<em>MLP</em>* and keep on adjusting paras and structrues of net to get score 0.813,and ensembled different MLP model results with different weight to get a better score. I learned a lot from Kaggle, especially many kaggle master's ideas in Disscussions area,you can learn much new views from others.🤓 </p>",
  "messages": [
    {
      "id": "2031331",
      "postDate": "11/16/2022 02:26:49",
      "content": "<p>we tried many models and some model or ensembled model are useful,the first month,we mainly focused on tree models:<strong>lightgbm and catboost</strong>, lightgbm not support multi targets and must use the** sklearn.preprocess.multioutput**, and running very slowly, and so does catboost, and the results of tree models are basically about 0.808/0.809 , which is lower than MLP. Actually,tree models are useful and good at the tabular dataset in most case. But in this competition,it's not easy to do some feature engineering due to gene features and we don't know the what the gene features represent for. last two months,we mainly use *<em>MLP</em>* and keep on adjusting paras and structrues of net to get score 0.813,and ensembled different MLP model results with different weight to get a better score. I learned a lot from Kaggle, especially many kaggle master's ideas in Disscussions area,you can learn much new views from others.🤓 </p>",
      "rawMarkdown": "we tried many models and some model or ensembled model are useful,the first month,we mainly focused on tree models:**lightgbm and catboost**, lightgbm not support multi targets and must use the** sklearn.preprocess.multioutput**, and running very slowly, and so does catboost, and the results of tree models are basically about 0.808/0.809 , which is lower than MLP. Actually,tree models are useful and good at the tabular dataset in most case. But in this competition,it's not easy to do some feature engineering due to gene features and we don't know the what the gene features represent for. last two months,we mainly use **MLP** and keep on adjusting paras and structrues of net to get score 0.813,and ensembled different MLP model results with different weight to get a better score. I learned a lot from Kaggle, especially many kaggle master's ideas in Disscussions area,you can learn much new views from others.🤓",
      "votes": null
    },
    {
      "id": "2031338",
      "postDate": "11/16/2022 02:33:14",
      "content": "<p>catboost support multi targets,and the loss_function and eval_metric must be :MultiRMSE and it's fast than lgb in this competition</p>",
      "rawMarkdown": "catboost support multi targets,and the loss_function and eval_metric must be :MultiRMSE and it's fast than lgb in this competition",
      "votes": null
    },
    {
      "id": "2031496",
      "postDate": "11/16/2022 05:49:12",
      "content": "<p>wow, thank you for sharing, this is something new!</p>",
      "rawMarkdown": "wow, thank you for sharing, this is something new!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2031338,
      "author_name": "kunmingxie",
      "author_url": "",
      "post_date": "11/16/2022 02:33:14",
      "content": "<p>catboost support multi targets,and the loss_function and eval_metric must be :MultiRMSE and it's fast than lgb in this competition</p>",
      "votes": null,
      "replies": [
        {
          "id": 2031496,
          "author_name": "akmalmir",
          "author_url": "",
          "post_date": "11/16/2022 05:49:12",
          "content": "<p>wow, thank you for sharing, this is something new!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2031331": "we tried many models and some model or ensembled model are useful,the first month,we mainly focused on tree models:**lightgbm and catboost**, lightgbm not support multi targets and must use the** sklearn.preprocess.multioutput**, and running very slowly, and so does catboost, and the results of tree models are basically about 0.808/0.809 , which is lower than MLP. Actually,tree models are useful and good at the tabular dataset in most case. But in this competition,it's not easy to do some feature engineering due to gene features and we don't know the what the gene features represent for. last two months,we mainly use **MLP** and keep on adjusting paras and structrues of net to get score 0.813,and ensembled different MLP model results with different weight to get a better score. I learned a lot from Kaggle, especially many kaggle master's ideas in Disscussions area,you can learn much new views from others.🤓",
    "2031338": "catboost support multi targets,and the loss_function and eval_metric must be :MultiRMSE and it's fast than lgb in this competition",
    "2031496": "wow, thank you for sharing, this is something new!"
  },
  "source": "meta"
}