what is the best way to have an auc value more than the input submission files?? \nI have already tried most of the ensemble processes in my files, though I am not getting a satisfying result.
", "messages": [ { "id": "957725", "postDate": "08/04/2020 14:11:19", "content": "what is the best way to have an auc value more than the input submission files?? \nI have already tried most of the ensemble processes in my files, though I am not getting a satisfying result.
", "rawMarkdown": "what is the best way to have an auc value more than the input submission files?? \nI have already tried most of the ensemble processes in my files, though I am not getting a satisfying result.", "votes": null }, { "id": "959407", "postDate": "08/05/2020 15:00:57", "content": "Power average is the best I can think of, It would be better if all your submission files are from similar architecture with different seeds or different number of folds for better results.
", "rawMarkdown": "Power average is the best I can think of, It would be better if all your submission files are from similar architecture with different seeds or different number of folds for better results.", "votes": null }, { "id": "959475", "postDate": "08/05/2020 16:00:12", "content": "You can check out this topic for discussion on that matter.
\n\nIn case you are not satisfied with the result, the reason might be not how you ensemble predictions but what are the predictions that you combine. Remember that there are two qualities necessary for a good ensemble: strength (how well your individual models perform) and diversity (how different are the models). The performance boost is to be expected only if both of these criteria are fulfilled.
", "rawMarkdown": "You can check out [this topic](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/165653) for discussion on that matter.\n\nIn case you are not satisfied with the result, the reason might be not **how** you ensemble predictions but **what** are the predictions that you combine. Remember that there are two qualities necessary for a good ensemble: strength (how well your individual models perform) and diversity (how different are the models). The performance boost is to be expected only if both of these criteria are fulfilled.", "votes": null }, { "id": "959556", "postDate": "08/05/2020 17:31:19", "content": "The way to ensemble is important I agree. But your choice of models is more important. \nYou need to take different models with good scores.\nTo see the correlation/difference, you can put the output targets in a DataFrame and use df.corr().
", "rawMarkdown": "The way to ensemble is important I agree. But your choice of models is more important. \nYou need to take different models with good scores.\nTo see the correlation/difference, you can put the output targets in a DataFrame and use df.corr().", "votes": null }, { "id": "959952", "postDate": "08/06/2020 03:56:40", "content": "Thanks @kozodoi I will concentrate on these things during ensembling. Truly helpful suggestion for me
", "rawMarkdown": "Thanks @kozodoi I will concentrate on these things during ensembling. Truly helpful suggestion for me", "votes": null } ], "comments": [ { "id": 959407, "author_name": "bharatsahu", "author_url": "", "post_date": "08/05/2020 15:00:57", "content": "Power average is the best I can think of, It would be better if all your submission files are from similar architecture with different seeds or different number of folds for better results.
", "votes": null, "replies": [] }, { "id": 959475, "author_name": "kozodoi", "author_url": "", "post_date": "08/05/2020 16:00:12", "content": "You can check out this topic for discussion on that matter.
\n\nIn case you are not satisfied with the result, the reason might be not how you ensemble predictions but what are the predictions that you combine. Remember that there are two qualities necessary for a good ensemble: strength (how well your individual models perform) and diversity (how different are the models). The performance boost is to be expected only if both of these criteria are fulfilled.
", "votes": null, "replies": [] }, { "id": 959556, "author_name": "vicioussong", "author_url": "", "post_date": "08/05/2020 17:31:19", "content": "The way to ensemble is important I agree. But your choice of models is more important. \nYou need to take different models with good scores.\nTo see the correlation/difference, you can put the output targets in a DataFrame and use df.corr().
", "votes": null, "replies": [] }, { "id": 959952, "author_name": "ritachetadas", "author_url": "", "post_date": "08/06/2020 03:56:40", "content": "Thanks @kozodoi I will concentrate on these things during ensembling. Truly helpful suggestion for me
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "957725": "what is the best way to have an auc value more than the input submission files?? \nI have already tried most of the ensemble processes in my files, though I am not getting a satisfying result.", "959407": "Power average is the best I can think of, It would be better if all your submission files are from similar architecture with different seeds or different number of folds for better results.", "959475": "You can check out [this topic](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/165653) for discussion on that matter.\n\nIn case you are not satisfied with the result, the reason might be not **how** you ensemble predictions but **what** are the predictions that you combine. Remember that there are two qualities necessary for a good ensemble: strength (how well your individual models perform) and diversity (how different are the models). The performance boost is to be expected only if both of these criteria are fulfilled.", "959556": "The way to ensemble is important I agree. But your choice of models is more important. \nYou need to take different models with good scores.\nTo see the correlation/difference, you can put the output targets in a DataFrame and use df.corr().", "959952": "Thanks @kozodoi I will concentrate on these things during ensembling. Truly helpful suggestion for me" }, "source": "meta" }