Hey there guys,
\n\nare you having success with stacking? We didn't put much effort in it so far but from preliminary trials it seems like it doesn't improve over simple blending of our best submissions.
\n\nIs someone else experiencing the same? If not, what are you guys using? Linear or nonlinear models for the stack? A mix of them? Thanks
\n\nCheers.
", "messages": [ { "id": "348115", "postDate": "06/26/2018 06:03:44", "content": "Hey there guys,
\n\nare you having success with stacking? We didn't put much effort in it so far but from preliminary trials it seems like it doesn't improve over simple blending of our best submissions.
\n\nIs someone else experiencing the same? If not, what are you guys using? Linear or nonlinear models for the stack? A mix of them? Thanks
\n\nCheers.
", "rawMarkdown": "Hey there guys,\n\nare you having success with stacking? We didn't put much effort in it so far but from preliminary trials it seems like it doesn't improve over simple blending of our best submissions. \n\nIs someone else experiencing the same? If not, what are you guys using? Linear or nonlinear models for the stack? A mix of them? Thanks\n\nCheers.", "votes": null }, { "id": "348119", "postDate": "06/26/2018 06:14:16", "content": "Stacking has been working very well for us. It's key to our solution.
", "rawMarkdown": "Stacking has been working very well for us. It's key to our solution.", "votes": null }, { "id": "348120", "postDate": "06/26/2018 06:15:53", "content": "Thanks Peter. Very interesting insight! May I ask you whether you guys are stacking only your model predictions or adding some extra-features as well?
", "rawMarkdown": "Thanks Peter. Very interesting insight! May I ask you whether you guys are stacking only your model predictions or adding some extra-features as well?", "votes": null }, { "id": "348148", "postDate": "06/26/2018 07:05:58", "content": "I like to stack with other features in addition to the model predictions, but this doesn't always work best. You have to be careful and try both ways.
", "rawMarkdown": "I like to stack with other features in addition to the model predictions, but this doesn't always work best. You have to be careful and try both ways.", "votes": null }, { "id": "348162", "postDate": "06/26/2018 07:25:22", "content": "i tried stack with other features but they have little contribution compared with model predictions... i wonder when you say 'be careful' you mean carefully select other features as input to stack?
", "rawMarkdown": "i tried stack with other features but they have little contribution compared with model predictions... i wonder when you say 'be careful' you mean carefully select other features as input to stack?", "votes": null }, { "id": "348168", "postDate": "06/26/2018 07:36:35", "content": "I mean that I found some features to be helpful, but other features to merely overfit. So I had to select features into the model one-by-one.
", "rawMarkdown": "I mean that I found some features to be helpful, but other features to merely overfit. So I had to select features into the model one-by-one.", "votes": null }, { "id": "348179", "postDate": "06/26/2018 07:57:30", "content": "Thanks that’s very useful to me.
", "rawMarkdown": "Thanks that’s very useful to me.", "votes": null }, { "id": "348220", "postDate": "06/26/2018 09:54:39", "content": "Peter you used both linear and non-linear stackers?
", "rawMarkdown": "Peter you used both linear and non-linear stackers?", "votes": null }, { "id": "348223", "postDate": "06/26/2018 10:06:22", "content": "have you checked the correlation between your models?
", "rawMarkdown": "have you checked the correlation between your models?", "votes": null }, { "id": "348224", "postDate": "06/26/2018 10:13:54", "content": "Yes, we did. Now are working to diversify our submissions including other models. Did stacking work for you Izmaylov? Did you use tree based models for that including other features? Thanks
", "rawMarkdown": "Yes, we did. Now are working to diversify our submissions including other models. Did stacking work for you Izmaylov? Did you use tree based models for that including other features? Thanks", "votes": null }, { "id": "348235", "postDate": "06/26/2018 10:46:17", "content": "Yes, stacking works great, beforehand, check the correlation. Try different structures of stacking, several linear and non-linear layers.
", "rawMarkdown": "Yes, stacking works great, beforehand, check the correlation. Try different structures of stacking, several linear and non-linear layers.", "votes": null }, { "id": "348259", "postDate": "06/26/2018 12:01:32", "content": "Thank for your sharing. How to check the correlation? any refer links?
", "rawMarkdown": "Thank for your sharing. How to check the correlation? any refer links?", "votes": null }, { "id": "348261", "postDate": "06/26/2018 12:08:14", "content": "If your models' predictions are concatenated in a pandas dataframe (one column per model) you can do it easily with df.corr(), where df is the dataframe
", "rawMarkdown": "If your models' predictions are concatenated in a pandas dataframe (one column per model) you can do it easily with df.corr(), where df is the dataframe", "votes": null }, { "id": "348446", "postDate": "06/26/2018 18:46:33", "content": "It works well in this competition.
", "rawMarkdown": "It works well in this competition.", "votes": null }, { "id": "348483", "postDate": "06/26/2018 19:45:24", "content": "@den3b: We'll do a full write-up and full code share after the competition is over and you can see.
", "rawMarkdown": "den3b: We'll do a full write-up and full code share after the competition is over and you can see.", "votes": null }, { "id": "348599", "postDate": "06/27/2018 01:51:09", "content": "It's the key for our model in the last week.
", "rawMarkdown": "It's the key for our model in the last week.", "votes": null }, { "id": "348646", "postDate": "06/27/2018 03:46:52", "content": "df.corr(), got it, thanks
", "rawMarkdown": "df.corr(), got it, thanks", "votes": null }, { "id": "348659", "postDate": "06/27/2018 04:29:44", "content": "Thanks :) Yeah, it started working for us as well.
", "rawMarkdown": "Thanks :) Yeah, it started working for us as well.", "votes": null }, { "id": "348660", "postDate": "06/27/2018 04:30:34", "content": "Yeah Peter thanks. It started working for us as well, I should have dedicated more time to it. But it's all good, glad to see we can still improve on the leaderboard.
", "rawMarkdown": "Yeah Peter thanks. It started working for us as well, I should have dedicated more time to it. But it's all good, glad to see we can still improve on the leaderboard.", "votes": null }, { "id": "348664", "postDate": "06/27/2018 05:01:52", "content": "Glad to hear it! There's still some time left and either way it will still make you stronger for the next competition.
", "rawMarkdown": "Glad to hear it! There's still some time left and either way it will still make you stronger for the next competition.", "votes": null }, { "id": "348735", "postDate": "06/27/2018 07:29:06", "content": "Are you guys using different algorithms for stacking? How do the models differentiate?
", "rawMarkdown": "Are you guys using different algorithms for stacking? How do the models differentiate?", "votes": null }, { "id": "348917", "postDate": "06/27/2018 14:15:32", "content": "Hey Kazanova.
\nI consider you to be one of the best stackers in the world,
\nso I really hope you can elaborate some of your stacking techniques after this competition...
stacknet is very cool tool for me at real work, also for this competition.
", "rawMarkdown": "stacknet is very cool tool for me at real work, also for this competition.", "votes": null } ], "comments": [ { "id": 348119, "author_name": "peterhurford", "author_url": "", "post_date": "06/26/2018 06:14:16", "content": "Stacking has been working very well for us. It's key to our solution.
", "votes": null, "replies": [ { "id": 348120, "author_name": "den3b81", "author_url": "", "post_date": "06/26/2018 06:15:53", "content": "Thanks Peter. Very interesting insight! May I ask you whether you guys are stacking only your model predictions or adding some extra-features as well?
", "votes": null, "replies": [] }, { "id": 348148, "author_name": "peterhurford", "author_url": "", "post_date": "06/26/2018 07:05:58", "content": "I like to stack with other features in addition to the model predictions, but this doesn't always work best. You have to be careful and try both ways.
", "votes": null, "replies": [] }, { "id": 348162, "author_name": "yimacs", "author_url": "", "post_date": "06/26/2018 07:25:22", "content": "i tried stack with other features but they have little contribution compared with model predictions... i wonder when you say 'be careful' you mean carefully select other features as input to stack?
", "votes": null, "replies": [] }, { "id": 348168, "author_name": "peterhurford", "author_url": "", "post_date": "06/26/2018 07:36:35", "content": "I mean that I found some features to be helpful, but other features to merely overfit. So I had to select features into the model one-by-one.
", "votes": null, "replies": [] }, { "id": 348179, "author_name": "yimacs", "author_url": "", "post_date": "06/26/2018 07:57:30", "content": "Thanks that’s very useful to me.
", "votes": null, "replies": [] }, { "id": 348220, "author_name": "den3b81", "author_url": "", "post_date": "06/26/2018 09:54:39", "content": "Peter you used both linear and non-linear stackers?
", "votes": null, "replies": [] }, { "id": 348483, "author_name": "peterhurford", "author_url": "", "post_date": "06/26/2018 19:45:24", "content": "@den3b: We'll do a full write-up and full code share after the competition is over and you can see.
", "votes": null, "replies": [] }, { "id": 348660, "author_name": "den3b81", "author_url": "", "post_date": "06/27/2018 04:30:34", "content": "Yeah Peter thanks. It started working for us as well, I should have dedicated more time to it. But it's all good, glad to see we can still improve on the leaderboard.
", "votes": null, "replies": [] }, { "id": 348664, "author_name": "peterhurford", "author_url": "", "post_date": "06/27/2018 05:01:52", "content": "Glad to hear it! There's still some time left and either way it will still make you stronger for the next competition.
", "votes": null, "replies": [] }, { "id": 348735, "author_name": "cranberry93", "author_url": "", "post_date": "06/27/2018 07:29:06", "content": "Are you guys using different algorithms for stacking? How do the models differentiate?
", "votes": null, "replies": [] } ] }, { "id": 348223, "author_name": "izmaylov", "author_url": "", "post_date": "06/26/2018 10:06:22", "content": "have you checked the correlation between your models?
", "votes": null, "replies": [ { "id": 348224, "author_name": "den3b81", "author_url": "", "post_date": "06/26/2018 10:13:54", "content": "Yes, we did. Now are working to diversify our submissions including other models. Did stacking work for you Izmaylov? Did you use tree based models for that including other features? Thanks
", "votes": null, "replies": [] }, { "id": 348235, "author_name": "izmaylov", "author_url": "", "post_date": "06/26/2018 10:46:17", "content": "Yes, stacking works great, beforehand, check the correlation. Try different structures of stacking, several linear and non-linear layers.
", "votes": null, "replies": [] }, { "id": 348259, "author_name": "yyqing", "author_url": "", "post_date": "06/26/2018 12:01:32", "content": "Thank for your sharing. How to check the correlation? any refer links?
", "votes": null, "replies": [] }, { "id": 348261, "author_name": "den3b81", "author_url": "", "post_date": "06/26/2018 12:08:14", "content": "If your models' predictions are concatenated in a pandas dataframe (one column per model) you can do it easily with df.corr(), where df is the dataframe
", "votes": null, "replies": [] }, { "id": 348646, "author_name": "yyqing", "author_url": "", "post_date": "06/27/2018 03:46:52", "content": "df.corr(), got it, thanks
", "votes": null, "replies": [] } ] }, { "id": 348446, "author_name": "kazanova", "author_url": "", "post_date": "06/26/2018 18:46:33", "content": "It works well in this competition.
", "votes": null, "replies": [ { "id": 348659, "author_name": "den3b81", "author_url": "", "post_date": "06/27/2018 04:29:44", "content": "Thanks :) Yeah, it started working for us as well.
", "votes": null, "replies": [] }, { "id": 348917, "author_name": "pocketsuteado", "author_url": "", "post_date": "06/27/2018 14:15:32", "content": "Hey Kazanova.
\nI consider you to be one of the best stackers in the world,
\nso I really hope you can elaborate some of your stacking techniques after this competition...
stacknet is very cool tool for me at real work, also for this competition.
", "votes": null, "replies": [] } ] }, { "id": 348599, "author_name": "naivelamb", "author_url": "", "post_date": "06/27/2018 01:51:09", "content": "It's the key for our model in the last week.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "348115": "Hey there guys,\n\nare you having success with stacking? We didn't put much effort in it so far but from preliminary trials it seems like it doesn't improve over simple blending of our best submissions. \n\nIs someone else experiencing the same? If not, what are you guys using? Linear or nonlinear models for the stack? A mix of them? Thanks\n\nCheers.", "348119": "Stacking has been working very well for us. It's key to our solution.", "348120": "Thanks Peter. Very interesting insight! May I ask you whether you guys are stacking only your model predictions or adding some extra-features as well?", "348148": "I like to stack with other features in addition to the model predictions, but this doesn't always work best. You have to be careful and try both ways.", "348162": "i tried stack with other features but they have little contribution compared with model predictions... i wonder when you say 'be careful' you mean carefully select other features as input to stack?", "348168": "I mean that I found some features to be helpful, but other features to merely overfit. So I had to select features into the model one-by-one.", "348179": "Thanks that’s very useful to me.", "348220": "Peter you used both linear and non-linear stackers?", "348223": "have you checked the correlation between your models?", "348224": "Yes, we did. Now are working to diversify our submissions including other models. Did stacking work for you Izmaylov? Did you use tree based models for that including other features? Thanks", "348235": "Yes, stacking works great, beforehand, check the correlation. Try different structures of stacking, several linear and non-linear layers.", "348259": "Thank for your sharing. How to check the correlation? any refer links?", "348261": "If your models' predictions are concatenated in a pandas dataframe (one column per model) you can do it easily with df.corr(), where df is the dataframe", "348446": "It works well in this competition.", "348483": "den3b: We'll do a full write-up and full code share after the competition is over and you can see.", "348599": "It's the key for our model in the last week.", "348646": "df.corr(), got it, thanks", "348659": "Thanks :) Yeah, it started working for us as well.", "348660": "Yeah Peter thanks. It started working for us as well, I should have dedicated more time to it. But it's all good, glad to see we can still improve on the leaderboard.", "348664": "Glad to hear it! There's still some time left and either way it will still make you stronger for the next competition.", "348735": "Are you guys using different algorithms for stacking? How do the models differentiate?", "348917": "Hey Kazanova.