Does anyone use R ？
", "messages": [ { "id": "64014", "postDate": "02/11/2015 10:45:40", "content": "Does anyone use R ？
", "rawMarkdown": "", "votes": null }, { "id": "64037", "postDate": "02/11/2015 15:23:53", "content": "I tried using this code as a last ditch attempt to get a better score. It didn't do that well. Maybe 0.4, 0.41 public. Albeit I think it could possibly break 0.4 if you let it run for a week at lower granularity. It also takes forever to run, but it runs.
\nIt's basically greedy coordinate descent for functionals.
", "rawMarkdown": "", "votes": null }, { "id": "64089", "postDate": "02/12/2015 03:43:54", "content": "[quote=Mike Kim;64037]
\nI tried using this code as a last ditch attempt to get a better score. It didn't do that well. Maybe 0.4, 0.41 public. Albeit I think it could possibly break 0.4 if you let it run for a week at lower granularity. It also takes forever to run, but it runs.
\n[/quote]
\n\nThank U Mike Kim
", "rawMarkdown": "", "votes": null }, { "id": "64159", "postDate": "02/13/2015 09:48:58", "content": "If you can excuse/tolerate the poor quality of the code, these two files illustrate an exploration of the CTR histogram approach in R + xgboost, which gave a private LB score of 0.400 before I switched to vw and tinrtgu's code. Loading the data files is a little slow. The main model trains in about 40mins.
\nThis thread explains the approach
\n\nThe code could be made far cleaner if you have better-than-laptop hardware, e.g. the distinction between 'general' and 'key' variables could be dropped and script (a) scrapped entirely.
", "rawMarkdown": "", "votes": null }, { "id": "64231", "postDate": "02/14/2015 15:44:30", "content": "I have 0.398 on private LB. I used gbm on approximately 10 random subsamples (~900000 each). Features with many levels were grouped by k-means into 70 levels. Predictions from different gbm models were averaged. See absolutely ugly code in attachment. Please excuse me in advance for code quality.
\nBefore the couple of days to the competition end I found package RSofia that implement Google algorithm sofia-ml I didn't try it but it looks like very well suited for this competition task.
", "rawMarkdown": "", "votes": null }, { "id": "64251", "postDate": "02/15/2015 00:29:24", "content": "I used R for feature engineering (thanks data.table) and XGboost to train one of my models. But due to some bug in xgboost or problem in my PC, many time that i runned xgboost to train it crached while trainning. I realized that it occurs with big datasets. I tried to train a 40M, 30M, 20M, 15M, 10M instances dataset but it all crashed before the end of trainning process. Then when I started to train using 8M instances it didn't crashed anymore. But using only 8M instances I experienced low performance... bagging 2 xgboost models I got 0.395 in public LB. if I could train using the whole dataset the performance could be better.
\nJust bagging that 0.395 gbm model with some other 0.389x FTRL models I got that 0.3867645/0.3848830 public/private LB ;-D
", "rawMarkdown": "", "votes": null }, { "id": "65107", "postDate": "02/28/2015 05:24:25", "content": "[quote=lewis ml;64159]
\nIf you can excuse/tolerate the poor quality of the code, these two files illustrate an exploration of the CTR histogram approach in R + xgboost, which gave a private LB score of 0.400 before I switched to vw and tinrtgu's code. Loading the data files is a little slow. The main model trains in about 40mins.
\nThis thread explains the approach
\n\nThe code could be made far cleaner if you have better-than-laptop hardware, e.g. the distinction between 'general' and 'key' variables could be dropped and script (a) scrapped entirely.
\n[/quote]
\nlewis ml,thank you so much!
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 64037, "author_name": "mikeskim", "author_url": "", "post_date": "02/11/2015 15:23:53", "content": "I tried using this code as a last ditch attempt to get a better score. It didn't do that well. Maybe 0.4, 0.41 public. Albeit I think it could possibly break 0.4 if you let it run for a week at lower granularity. It also takes forever to run, but it runs.
\nIt's basically greedy coordinate descent for functionals.
", "votes": null, "replies": [] }, { "id": 64089, "author_name": "rowanyuan", "author_url": "", "post_date": "02/12/2015 03:43:54", "content": "[quote=Mike Kim;64037]
\nI tried using this code as a last ditch attempt to get a better score. It didn't do that well. Maybe 0.4, 0.41 public. Albeit I think it could possibly break 0.4 if you let it run for a week at lower granularity. It also takes forever to run, but it runs.
\n[/quote]
\n\nThank U Mike Kim
", "votes": null, "replies": [] }, { "id": 64159, "author_name": "cartographic", "author_url": "", "post_date": "02/13/2015 09:48:58", "content": "If you can excuse/tolerate the poor quality of the code, these two files illustrate an exploration of the CTR histogram approach in R + xgboost, which gave a private LB score of 0.400 before I switched to vw and tinrtgu's code. Loading the data files is a little slow. The main model trains in about 40mins.
\nThis thread explains the approach
\n\nThe code could be made far cleaner if you have better-than-laptop hardware, e.g. the distinction between 'general' and 'key' variables could be dropped and script (a) scrapped entirely.
", "votes": null, "replies": [] }, { "id": 64231, "author_name": "gdemin", "author_url": "", "post_date": "02/14/2015 15:44:30", "content": "I have 0.398 on private LB. I used gbm on approximately 10 random subsamples (~900000 each). Features with many levels were grouped by k-means into 70 levels. Predictions from different gbm models were averaged. See absolutely ugly code in attachment. Please excuse me in advance for code quality.
\nBefore the couple of days to the competition end I found package RSofia that implement Google algorithm sofia-ml I didn't try it but it looks like very well suited for this competition task.
", "votes": null, "replies": [] }, { "id": 64251, "author_name": "titericz", "author_url": "", "post_date": "02/15/2015 00:29:24", "content": "I used R for feature engineering (thanks data.table) and XGboost to train one of my models. But due to some bug in xgboost or problem in my PC, many time that i runned xgboost to train it crached while trainning. I realized that it occurs with big datasets. I tried to train a 40M, 30M, 20M, 15M, 10M instances dataset but it all crashed before the end of trainning process. Then when I started to train using 8M instances it didn't crashed anymore. But using only 8M instances I experienced low performance... bagging 2 xgboost models I got 0.395 in public LB. if I could train using the whole dataset the performance could be better.
\nJust bagging that 0.395 gbm model with some other 0.389x FTRL models I got that 0.3867645/0.3848830 public/private LB ;-D
", "votes": null, "replies": [] }, { "id": 65107, "author_name": "rowanyuan", "author_url": "", "post_date": "02/28/2015 05:24:25", "content": "[quote=lewis ml;64159]
\nIf you can excuse/tolerate the poor quality of the code, these two files illustrate an exploration of the CTR histogram approach in R + xgboost, which gave a private LB score of 0.400 before I switched to vw and tinrtgu's code. Loading the data files is a little slow. The main model trains in about 40mins.
\nThis thread explains the approach
\n\nThe code could be made far cleaner if you have better-than-laptop hardware, e.g. the distinction between 'general' and 'key' variables could be dropped and script (a) scrapped entirely.
\n[/quote]
\nlewis ml,thank you so much!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "64014": "", "64037": "", "64089": "", "64159": "", "64231": "", "64251": "", "65107": "" }, "source": "meta" }