Hi all,
\nI am wondering what is the model performance without the ensembling in this competition. Maybe we could share this information in this place. I do not use the ensemble methods right now, so my best single model score on pb is 0.811.
", "messages": [ { "id": 1948430, "postDate": "2022-09-21T04:24:08.293Z", "content": "Hi all,
\nI am wondering what is the model performance without the ensembling in this competition. Maybe we could share this information in this place. I do not use the ensemble methods right now, so my best single model score on pb is 0.811.
", "rawMarkdown": "Hi all, \n\nI am wondering what is the model performance without the ensembling in this competition. Maybe we could share this information in this place. I do not use the ensemble methods right now, so my best single model score on pb is 0.811. ", "votes": 12 }, { "id": 2029012, "postDate": "2022-11-14T11:28:14.967Z", "content": "Thank you very much, it's helpful for me
", "rawMarkdown": "Thank you very much, it's helpful for me", "votes": -1 }, { "id": 1966316, "postDate": "2022-10-01T21:05:45.853Z", "content": "For those of you who mentioned MLP, are you using a reduced feature set as input? Such as truncated SVD? Or doing feature selection before MLP?
\nFor me, my best single model score is .802, which is combined CITE and Multiome, one model for each.
", "rawMarkdown": "For those of you who mentioned MLP, are you using a reduced feature set as input? Such as truncated SVD? Or doing feature selection before MLP? \n\nFor me, my best single model score is .802, which is combined CITE and Multiome, one model for each." }, { "id": 1948704, "postDate": "2022-09-21T08:29:48.103Z", "content": "Hi would you mind sharing your score on valid? citeseq and multi. There is a huge gap between my score on CV and PB. I'm looking for the answer
\nCV !~ PB : find a correct CV ? or ignore PB ?
Sure, my MULT past is 0.6677, and my CITE part is 0.8931. Both of them are based on group k-folds (3 donors).
", "rawMarkdown": "Sure, my MULT past is 0.6677, and my CITE part is 0.8931. Both of them are based on group k-folds (3 donors). ", "votes": 1 }, { "id": 1951827, "postDate": "2022-09-23T09:13:21.390Z", "content": "Thank you for sharing
", "rawMarkdown": "Thank you for sharing" }, { "id": 1959997, "postDate": "2022-09-28T12:36:32.167Z", "content": "all are mlp for these two part of data？Can I ask how many features you used？I add more features to train ,I tried 256 and 512 ，but the result seems almost the same，not much improvement ，tree models (like catboost) are also like that. I don’t know whether to focus on features or maybe I should keep turning the NN model and paras.By the way,the tree model does not seem to perform as well as the network model in my training results.😊
", "rawMarkdown": "all are mlp for these two part of data？Can I ask how many features you used？I add more features to train ,I tried 256 and 512 ，but the result seems almost the same，not much improvement ，tree models (like catboost) are also like that. I don’t know whether to focus on features or maybe I should keep turning the NN model and paras.By the way,the tree model does not seem to perform as well as the network model in my training results.😊", "votes": 1 }, { "id": 1961353, "postDate": "2022-09-29T06:04:14.217Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 1961400, "postDate": "2022-09-29T06:35:48.050Z", "content": "more features will be better in my Catboost case,but not much improvement,just a little, but in my NN model,128 performs better than 256 and 512.It seems no feature engineering to do with the gene data. May be we should focus on different models and try different paras.
", "rawMarkdown": "more features will be better in my Catboost case,but not much improvement,just a little, but in my NN model,128 performs better than 256 and 512.It seems no feature engineering to do with the gene data. May be we should focus on different models and try different paras.", "votes": 1 }, { "id": 1962885, "postDate": "2022-09-30T03:27:41.963Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 1962906, "postDate": "2022-09-30T04:00:56.937Z", "content": "Awesome!!! May I ask the 121 features are the important features(features name are relevant to targets names) in many kaggle Master mentioned？my cv score for Cite is just 0.89219 and 656 features(512+144)，maybe I should readjust my model structure.
", "rawMarkdown": "Awesome!!! May I ask the 121 features are the important features(features name are relevant to targets names) in many kaggle Master mentioned？my cv score for Cite is just 0.89219 and 656 features(512+144)，maybe I should readjust my model structure." }, { "id": 1963222, "postDate": "2022-09-30T06:56:39.863Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 1957455, "postDate": "2022-09-27T01:54:17.473Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 1957611, "postDate": "2022-09-27T04:49:46.280Z", "content": "mlp, based on keras
", "rawMarkdown": "mlp, based on keras", "votes": 1 }, { "id": 1959476, "postDate": "2022-09-28T06:29:00.913Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 2029012, "author_name": "", "author_url": "", "post_date": "2022-11-14T11:28:14.967000", "content": "Thank you very much, it's helpful for me
", "votes": -1, "replies": [] }, { "id": 1966316, "author_name": "KirkDCO", "author_url": "", "post_date": "2022-10-01T21:05:45.853000", "content": "For those of you who mentioned MLP, are you using a reduced feature set as input? Such as truncated SVD? Or doing feature selection before MLP?
\nFor me, my best single model score is .802, which is combined CITE and Multiome, one model for each.
", "votes": 0, "replies": [] }, { "id": 1948704, "author_name": "Chouchou", "author_url": "", "post_date": "2022-09-21T08:29:48.103000", "content": "Hi would you mind sharing your score on valid? citeseq and multi. There is a huge gap between my score on CV and PB. I'm looking for the answer
\nCV !~ PB : find a correct CV ? or ignore PB ?
Sure, my MULT past is 0.6677, and my CITE part is 0.8931. Both of them are based on group k-folds (3 donors).
", "votes": 1, "replies": [] }, { "id": 1951827, "author_name": "Chouchou", "author_url": "", "post_date": "2022-09-23T09:13:21.390000", "content": "Thank you for sharing
", "votes": 0, "replies": [] }, { "id": 1959997, "author_name": "Kunming Xie", "author_url": "", "post_date": "2022-09-28T12:36:32.167000", "content": "all are mlp for these two part of data？Can I ask how many features you used？I add more features to train ,I tried 256 and 512 ，but the result seems almost the same，not much improvement ，tree models (like catboost) are also like that. I don’t know whether to focus on features or maybe I should keep turning the NN model and paras.By the way,the tree model does not seem to perform as well as the network model in my training results.😊
", "votes": 1, "replies": [] }, { "id": 1961353, "author_name": "", "author_url": "", "post_date": "2022-09-29T06:04:14.217000", "content": "", "votes": 1, "replies": [] }, { "id": 1961400, "author_name": "Kunming Xie", "author_url": "", "post_date": "2022-09-29T06:35:48.050000", "content": "more features will be better in my Catboost case,but not much improvement,just a little, but in my NN model,128 performs better than 256 and 512.It seems no feature engineering to do with the gene data. May be we should focus on different models and try different paras.
", "votes": 1, "replies": [] }, { "id": 1962885, "author_name": "", "author_url": "", "post_date": "2022-09-30T03:27:41.963000", "content": "", "votes": 1, "replies": [] }, { "id": 1962906, "author_name": "Kunming Xie", "author_url": "", "post_date": "2022-09-30T04:00:56.937000", "content": "Awesome!!! May I ask the 121 features are the important features(features name are relevant to targets names) in many kaggle Master mentioned？my cv score for Cite is just 0.89219 and 656 features(512+144)，maybe I should readjust my model structure.
", "votes": 0, "replies": [] }, { "id": 1963222, "author_name": "", "author_url": "", "post_date": "2022-09-30T06:56:39.863000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1957455, "author_name": "", "author_url": "", "post_date": "2022-09-27T01:54:17.473000", "content": "", "votes": 0, "replies": [ { "id": 1957611, "author_name": "no-magic", "author_url": "", "post_date": "2022-09-27T04:49:46.280000", "content": "mlp, based on keras
", "votes": 1, "replies": [] }, { "id": 1959476, "author_name": "", "author_url": "", "post_date": "2022-09-28T06:29:00.913000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1948430": "Hi all, \n\nI am wondering what is the model performance without the ensembling in this competition. Maybe we could share this information in this place. I do not use the ensemble methods right now, so my best single model score on pb is 0.811. ", "2029012": "Thank you very much, it's helpful for me", "1966316": "For those of you who mentioned MLP, are you using a reduced feature set as input? Such as truncated SVD? Or doing feature selection before MLP? \n\nFor me, my best single model score is .802, which is combined CITE and Multiome, one model for each.", "1948704": "Hi would you mind sharing your score on valid? citeseq and multi. There is a huge gap between my score on CV and PB. I'm looking for the answer\n[CV !~ PB : find a correct CV ? or ignore PB ?](https://www.kaggle.com/competitions/open-problems-multimodal/discussion/353993)", "1957455": "" } }