{ "id": 54586, "title": "Tweak to libffm to get AUC scores", "url": "/competitions/talkingdata-adtracking-fraud-detection/discussion/54586", "author_name": "", "post_date": "2018-04-15T13:32:26.400787700Z", "votes": 12, "comment_count": 9, "views": 0, "content": "

If you like libffm but hate the fact that it only displays log_loss scores and would like to have an idea of what the AUC score would be - I have tweaked the main file to use FastAUC

", "messages": [ { "id": "314390", "postDate": "04/15/2018 13:32:26", "content": "

If you like libffm but hate the fact that it only displays log_loss scores and would like to have an idea of what the AUC score would be - I have tweaked the main file to use FastAUC

", "rawMarkdown": "If you like libffm but hate the fact that it only displays log_loss scores and would like to have an idea of what the AUC score would be - I have tweaked the main file to use FastAUC", "votes": null }, { "id": "314626", "postDate": "04/16/2018 00:50:34", "content": "

There is also the xlearn library.

", "rawMarkdown": "There is also the [xlearn][1] library.\n\n\n [1]: http://xlearn-doc.readthedocs.io/en/latest/index.html", "votes": null }, { "id": "314716", "postDate": "04/16/2018 06:49:15", "content": "
\n

I have tweaked the main file to use FastAUC

\n
\n\n

This works for me only on regular-size datasets, but it throws a segmentation error on this dataset as it blows through memory. The only way I could solve this was by explicitly declaring the size of AUCtuff and working on a computer with 64Gb+ memory.

\n\n

vector<pair<float,int>> AUCtuff(150000000);

\n\n

It may be because I am working on a full dataset, so hopefully this can help others who are in the same boat.

\n\n

PS If you are wondering why 150M instead of 190M: I am doing 5-fold validation, so the largest training dataset is < 150M rows.

", "rawMarkdown": "> I have tweaked the main file to use FastAUC\n\nThis works for me only on regular-size datasets, but it throws a segmentation error on this dataset as it blows through memory. The only way I could solve this was by explicitly declaring the size of AUCtuff and working on a computer with 64Gb+ memory.\n\n`vector", "votes": null }, { "id": "314718", "postDate": "04/16/2018 06:56:17", "content": "

A way of reducing the memory might be to tweak the tweak by using

\n\n
vector<pair<float,unsigned char>>\n
", "rawMarkdown": "A way of reducing the memory might be to tweak the tweak by using \n\n vector", "votes": null }, { "id": "316834", "postDate": "04/20/2018 03:35:19", "content": "

Hi, Scirpus would you might share ffm.h with us? \nI'm not so familiar with c++

", "rawMarkdown": "Hi, Scirpus would you might share ffm.h with us? \nI'm not so familiar with c++", "votes": null }, { "id": "316945", "postDate": "04/20/2018 08:14:01", "content": "

Here for all code

", "rawMarkdown": "[Here for all code][1]\n\n\n [1]: https://github.com/guestwalk/libffm", "votes": null }, { "id": "319570", "postDate": "04/26/2018 10:27:56", "content": "

Thanks Scirpus for the tweak. Where did you declare variable ffm_problem? Did you modify ffm.h as well?

", "rawMarkdown": "Thanks Scirpus for the tweak. Where did you declare variable ffm_problem? Did you modify ffm.h as well?", "votes": null }, { "id": "319611", "postDate": "04/26/2018 12:42:04", "content": "

I only modified ffm.cpp and added the fastauc file - not sure what ffm_problem is just look at the file and it should be obvious

", "rawMarkdown": "I only modified ffm.cpp and added the fastauc file - not sure what ffm_problem is just look at the file and it should be obvious", "votes": null }, { "id": "319634", "postDate": "04/26/2018 13:45:35", "content": "

It was not that obvious but I think I found it. Thanks!

", "rawMarkdown": "It was not that obvious but I think I found it. Thanks!", "votes": null }, { "id": "319691", "postDate": "04/26/2018 15:48:04", "content": "

Cool

", "rawMarkdown": "Cool", "votes": null } ], "comments": [ { "id": 314626, "author_name": "sijunhe9248", "author_url": "", "post_date": "04/16/2018 00:50:34", "content": "

There is also the xlearn library.

", "votes": null, "replies": [] }, { "id": 314716, "author_name": "tilii7", "author_url": "", "post_date": "04/16/2018 06:49:15", "content": "
\n

I have tweaked the main file to use FastAUC

\n
\n\n

This works for me only on regular-size datasets, but it throws a segmentation error on this dataset as it blows through memory. The only way I could solve this was by explicitly declaring the size of AUCtuff and working on a computer with 64Gb+ memory.

\n\n

vector<pair<float,int>> AUCtuff(150000000);

\n\n

It may be because I am working on a full dataset, so hopefully this can help others who are in the same boat.

\n\n

PS If you are wondering why 150M instead of 190M: I am doing 5-fold validation, so the largest training dataset is < 150M rows.

", "votes": null, "replies": [ { "id": 314718, "author_name": "scirpus", "author_url": "", "post_date": "04/16/2018 06:56:17", "content": "

A way of reducing the memory might be to tweak the tweak by using

\n\n
vector<pair<float,unsigned char>>\n
", "votes": null, "replies": [] } ] }, { "id": 316834, "author_name": "lianglu", "author_url": "", "post_date": "04/20/2018 03:35:19", "content": "

Hi, Scirpus would you might share ffm.h with us? \nI'm not so familiar with c++

", "votes": null, "replies": [] }, { "id": 316945, "author_name": "scirpus", "author_url": "", "post_date": "04/20/2018 08:14:01", "content": "

Here for all code

", "votes": null, "replies": [] }, { "id": 319570, "author_name": "phansoks", "author_url": "", "post_date": "04/26/2018 10:27:56", "content": "

Thanks Scirpus for the tweak. Where did you declare variable ffm_problem? Did you modify ffm.h as well?

", "votes": null, "replies": [ { "id": 319611, "author_name": "scirpus", "author_url": "", "post_date": "04/26/2018 12:42:04", "content": "

I only modified ffm.cpp and added the fastauc file - not sure what ffm_problem is just look at the file and it should be obvious

", "votes": null, "replies": [] }, { "id": 319634, "author_name": "phansoks", "author_url": "", "post_date": "04/26/2018 13:45:35", "content": "

It was not that obvious but I think I found it. Thanks!

", "votes": null, "replies": [] }, { "id": 319691, "author_name": "scirpus", "author_url": "", "post_date": "04/26/2018 15:48:04", "content": "

Cool

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "314390": "If you like libffm but hate the fact that it only displays log_loss scores and would like to have an idea of what the AUC score would be - I have tweaked the main file to use FastAUC", "314626": "There is also the [xlearn][1] library.\n\n\n [1]: http://xlearn-doc.readthedocs.io/en/latest/index.html", "314716": "> I have tweaked the main file to use FastAUC\n\nThis works for me only on regular-size datasets, but it throws a segmentation error on this dataset as it blows through memory. The only way I could solve this was by explicitly declaring the size of AUCtuff and working on a computer with 64Gb+ memory.\n\n`vector", "314718": "A way of reducing the memory might be to tweak the tweak by using \n\n vector", "316834": "Hi, Scirpus would you might share ffm.h with us? \nI'm not so familiar with c++", "316945": "[Here for all code][1]\n\n\n [1]: https://github.com/guestwalk/libffm", "319570": "Thanks Scirpus for the tweak. Where did you declare variable ffm_problem? Did you modify ffm.h as well?", "319611": "I only modified ffm.cpp and added the fastauc file - not sure what ffm_problem is just look at the file and it should be obvious", "319634": "It was not that obvious but I think I found it. Thanks!", "319691": "Cool" }, "source": "meta" }