Has anyone had success with under- or over-sampling techniques for classification or regression? I've tryed under-sampling the zero revenue visits and applying regression, but results were not good. The same happened with a ensemble of multiple samples.
\n\nFor oversampling I used a label \"has_revenue\" as target for Smote and treated the actual numeric revenue as a standard feature. With this approach I got much better scores, but still worse than regular models. Maybe a more refined technique could work here - at least for classification.
", "messages": [ { "id": 399734, "postDate": "2018-10-06T15:51:50.590Z", "content": "Has anyone had success with under- or over-sampling techniques for classification or regression? I've tryed under-sampling the zero revenue visits and applying regression, but results were not good. The same happened with a ensemble of multiple samples.
\n\nFor oversampling I used a label \"has_revenue\" as target for Smote and treated the actual numeric revenue as a standard feature. With this approach I got much better scores, but still worse than regular models. Maybe a more refined technique could work here - at least for classification.
", "rawMarkdown": "Has anyone had success with under- or over-sampling techniques for classification or regression? I've tryed under-sampling the zero revenue visits and applying regression, but results were not good. The same happened with a ensemble of multiple samples.\n\nFor oversampling I used a label \"has_revenue\" as target for Smote and treated the actual numeric revenue as a standard feature. With this approach I got much better scores, but still worse than regular models. Maybe a more refined technique could work here - at least for classification.\n", "votes": 2 }, { "id": 399791, "postDate": "2018-10-06T19:14:33.517Z", "content": "I tried under sampling, which gave predictions with a much better distribution (visually at least) but score was terrible (around 2 or 3)
\n\nSee here: https://www.kaggle.com/xavierbourretsicotte/lgbm-error-analysis-negative-downsampling
\n\nand here: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/67883
", "rawMarkdown": "I tried under sampling, which gave predictions with a much better distribution (visually at least) but score was terrible (around 2 or 3) \n\nSee here: https://www.kaggle.com/xavierbourretsicotte/lgbm-error-analysis-negative-downsampling\n\nand here: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/67883" }, { "id": 399815, "postDate": "2018-10-06T20:51:18.473Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 399833, "postDate": "2018-10-06T22:08:52.520Z", "content": "No false negatives is a good start - perhaps try setting all values predicted as negative (zero I presume ?) to have value zero ?
\n\nAnyways from my trial and error it seems that despite everything there is some information contained in the bounces = 1 rows, because simply removing them from the training set worsens the LB as well... mysterious
", "rawMarkdown": "No false negatives is a good start - perhaps try setting all values predicted as negative (zero I presume ?) to have value zero ? \n\nAnyways from my trial and error it seems that despite everything there is some information contained in the bounces = 1 rows, because simply removing them from the training set worsens the LB as well... mysterious" }, { "id": 400350, "postDate": "2018-10-08T06:43:43.770Z", "content": "I have just started building a model and did simple linear regression . Almost all predicted values are negative when the actual transaction value is 0 . Moreover , the magnitudes are so insanely high , I will use a different model. linear regression wont be a good choice
", "rawMarkdown": "I have just started building a model and did simple linear regression . Almost all predicted values are negative when the actual transaction value is 0 . Moreover , the magnitudes are so insanely high , I will use a different model. linear regression wont be a good choice" } ] } ], "comments": [ { "id": 399791, "author_name": "Xavier Bourret Sicotte", "author_url": "", "post_date": "2018-10-06T19:14:33.517000", "content": "I tried under sampling, which gave predictions with a much better distribution (visually at least) but score was terrible (around 2 or 3)
\n\nSee here: https://www.kaggle.com/xavierbourretsicotte/lgbm-error-analysis-negative-downsampling
\n\nand here: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/67883
", "votes": 0, "replies": [] }, { "id": 399815, "author_name": "", "author_url": "", "post_date": "2018-10-06T20:51:18.473000", "content": "", "votes": 0, "replies": [ { "id": 399833, "author_name": "Xavier Bourret Sicotte", "author_url": "", "post_date": "2018-10-06T22:08:52.520000", "content": "No false negatives is a good start - perhaps try setting all values predicted as negative (zero I presume ?) to have value zero ?
\n\nAnyways from my trial and error it seems that despite everything there is some information contained in the bounces = 1 rows, because simply removing them from the training set worsens the LB as well... mysterious
", "votes": 0, "replies": [] }, { "id": 400350, "author_name": "Abhishek", "author_url": "", "post_date": "2018-10-08T06:43:43.770000", "content": "I have just started building a model and did simple linear regression . Almost all predicted values are negative when the actual transaction value is 0 . Moreover , the magnitudes are so insanely high , I will use a different model. linear regression wont be a good choice
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "399734": "Has anyone had success with under- or over-sampling techniques for classification or regression? I've tryed under-sampling the zero revenue visits and applying regression, but results were not good. The same happened with a ensemble of multiple samples.\n\nFor oversampling I used a label \"has_revenue\" as target for Smote and treated the actual numeric revenue as a standard feature. With this approach I got much better scores, but still worse than regular models. Maybe a more refined technique could work here - at least for classification.\n", "399791": "I tried under sampling, which gave predictions with a much better distribution (visually at least) but score was terrible (around 2 or 3) \n\nSee here: https://www.kaggle.com/xavierbourretsicotte/lgbm-error-analysis-negative-downsampling\n\nand here: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/67883", "399815": "" } }