{ "id": 199626, "title": "Cross-validation, just for hyperparameter optimisation ? Or for training too ?", "url": "/competitions/riiid-test-answer-prediction/discussion/199626", "author_name": "", "post_date": "2020-11-26T14:11:08.512219400Z", "votes": 3, "comment_count": 4, "views": 0, "content": "

Hi !
\nI am a beginner in data science.
\nI thank cross validation was used with grid search or bayesan optimisation etc to optimize the hyperparameters.
\nBut I am not sureโ€ฆ
\nWhen we have found our optimal model, have we to use cross validation to train our model ?
\nThank in advance for your answers !
\nHave a nice day, and good luck for the competition !

", "messages": [ { "id": "1092066", "postDate": "11/26/2020 14:11:08", "content": "

Hi !
\nI am a beginner in data science.
\nI thank cross validation was used with grid search or bayesan optimisation etc to optimize the hyperparameters.
\nBut I am not sureโ€ฆ
\nWhen we have found our optimal model, have we to use cross validation to train our model ?
\nThank in advance for your answers !
\nHave a nice day, and good luck for the competition !

", "rawMarkdown": "Hi ! \nI am a beginner in data science.\nI thank cross validation was used with grid search or bayesan optimisation etc to optimize the hyperparameters.\nBut I am not sure...\nWhen we have found our optimal model, have we to use cross validation to train our model ?\nThank in advance for your answers !\nHave a nice day, and good luck for the competition !", "votes": null }, { "id": "1092077", "postDate": "11/26/2020 14:22:29", "content": "

Hi @colomba,

\n

I would say that the main point of cross-validation is to reduce overfitting. So everything you do for the model development process (model comparison, features selection, feature engineering, hyperparameters tuning, etc.) will demand cross-validation.
\nWhen you have your final model and want to train and validate or deploy it, CV is not necessary anymore since you found already the parameters that best suits your model and you just wanna have a final one.

", "rawMarkdown": "Hi @colomba, \n\nI would say that the main point of cross-validation is to reduce overfitting. So everything you do for the model development process (model comparison, features selection, feature engineering, hyperparameters tuning, etc.) will demand cross-validation.\nWhen you have your final model and want to train and validate or deploy it, CV is not necessary anymore since you found already the parameters that best suits your model and you just wanna have a final one.", "votes": null }, { "id": "1092093", "postDate": "11/26/2020 14:34:10", "content": "

@pedrocouto39 , thank you for your answer !!

", "rawMarkdown": "pedrocouto39 , thank you for your answer !!", "votes": null }, { "id": "1092508", "postDate": "11/26/2020 22:54:51", "content": "

Also a downside would be having multiple models which make everything more complex (space, etc.)

", "rawMarkdown": "Also a downside would be having multiple models which make everything more complex (space, etc.)", "votes": null }, { "id": "1093397", "postDate": "11/27/2020 17:22:26", "content": "

Totally agree, @aliabdin1. This is where it gets a bit painful to run models ๐Ÿ˜‚

", "rawMarkdown": "Totally agree, @aliabdin1. This is where it gets a bit painful to run models ๐Ÿ˜‚", "votes": null } ], "comments": [ { "id": 1092077, "author_name": "pedrocouto39", "author_url": "", "post_date": "11/26/2020 14:22:29", "content": "

Hi @colomba,

\n

I would say that the main point of cross-validation is to reduce overfitting. So everything you do for the model development process (model comparison, features selection, feature engineering, hyperparameters tuning, etc.) will demand cross-validation.
\nWhen you have your final model and want to train and validate or deploy it, CV is not necessary anymore since you found already the parameters that best suits your model and you just wanna have a final one.

", "votes": null, "replies": [ { "id": 1092508, "author_name": "aliabdin1", "author_url": "", "post_date": "11/26/2020 22:54:51", "content": "

Also a downside would be having multiple models which make everything more complex (space, etc.)

", "votes": null, "replies": [] }, { "id": 1093397, "author_name": "pedrocouto39", "author_url": "", "post_date": "11/27/2020 17:22:26", "content": "

Totally agree, @aliabdin1. This is where it gets a bit painful to run models ๐Ÿ˜‚

", "votes": null, "replies": [] } ] }, { "id": 1092093, "author_name": "colomba", "author_url": "", "post_date": "11/26/2020 14:34:10", "content": "

@pedrocouto39 , thank you for your answer !!

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1092066": "Hi ! \nI am a beginner in data science.\nI thank cross validation was used with grid search or bayesan optimisation etc to optimize the hyperparameters.\nBut I am not sure...\nWhen we have found our optimal model, have we to use cross validation to train our model ?\nThank in advance for your answers !\nHave a nice day, and good luck for the competition !", "1092077": "Hi @colomba, \n\nI would say that the main point of cross-validation is to reduce overfitting. So everything you do for the model development process (model comparison, features selection, feature engineering, hyperparameters tuning, etc.) will demand cross-validation.\nWhen you have your final model and want to train and validate or deploy it, CV is not necessary anymore since you found already the parameters that best suits your model and you just wanna have a final one.", "1092093": "pedrocouto39 , thank you for your answer !!", "1092508": "Also a downside would be having multiple models which make everything more complex (space, etc.)", "1093397": "Totally agree, @aliabdin1. This is where it gets a bit painful to run models ๐Ÿ˜‚" }, "source": "meta" }