i use feature : display_id, ad_id, uuid, document_id, platform, country, state, dma, \n weekday, hour, cate, entity, topics, ad_id_2, ad_document_id, campaign_id, advertiser_id, ad_click_rate.\nalso try to compose some combination. \nlogloss stop at around 0.42 .
\n\ncould anyone give some clue?
", "messages": [ { "id": "153457", "postDate": "01/01/2017 16:30:19", "content": "i use feature : display_id, ad_id, uuid, document_id, platform, country, state, dma, \n weekday, hour, cate, entity, topics, ad_id_2, ad_document_id, campaign_id, advertiser_id, ad_click_rate.\nalso try to compose some combination. \nlogloss stop at around 0.42 .
\n\ncould anyone give some clue?
", "rawMarkdown": "i use feature : display_id, ad_id, uuid, document_id, platform, country, state, dma, \r\n weekday, hour, cate, entity, topics, ad_id_2, ad_document_id, campaign_id, advertiser_id, ad_click_rate.\r\nalso try to compose some combination. \r\nlogloss stop at around 0.42 . \r\n\r\ncould anyone give some clue?", "votes": null }, { "id": "153464", "postDate": "01/01/2017 17:00:35", "content": "use single model with ftrl.
", "rawMarkdown": "use single model with ftrl.", "votes": null }, { "id": "153520", "postDate": "01/02/2017 00:59:01", "content": "I was able to get a .64 just by training on all the data using SRK's template.
", "rawMarkdown": "I was able to get a .64 just by training on all the data using SRK's template.", "votes": null }, { "id": "153586", "postDate": "01/02/2017 11:31:13", "content": "i use his script too ... i try his script , but only got 0.53 on all the data ... actually , i just use displayid, ad_id and train for 8 epoch, and get result 0.54
", "rawMarkdown": "i use his script too ... i try his script , but only got 0.53 on all the data ... actually , i just use displayid, ad_id and train for 8 epoch, and get result 0.54", "votes": null }, { "id": "153622", "postDate": "01/02/2017 14:55:18", "content": "Display id is not a particularly good feature, try something else
", "rawMarkdown": "Display id is not a particularly good feature, try something else", "votes": null }, { "id": "153656", "postDate": "01/02/2017 17:27:33", "content": "@BIGAN\nHmmmm well i'm definitely not an expert but for me, two things come to mind. Are you sure you're training over the entire dataset? passing 8 times over 80 million datapoints must take forever to run (are you using pypy? havent had a chance to look at that). If you are training on all the data, maybe you're overfitting. The default regularization parameters are about as 'off' as they can be and running 8 passes over the data will certainly reinforce any overfitting tendencies. Not sure how strong the overfitting is since the there are only 2**20 permissible weights for many more predictors. When I trained on 100k datapoints I was able to get a .55 although that was only a 2-fold CV and not an LB score so take that with a grain of salt.
\n\nHow are you performing your CV? What is the logloss stopping thing you're doing and why?
", "rawMarkdown": "BIGAN\r\nHmmmm well i'm definitely not an expert but for me, two things come to mind. Are you sure you're training over the entire dataset? passing 8 times over 80 million datapoints must take forever to run (are you using pypy? havent had a chance to look at that). If you are training on all the data, maybe you're overfitting. The default regularization parameters are about as 'off' as they can be and running 8 passes over the data will certainly reinforce any overfitting tendencies. Not sure how strong the overfitting is since the there are only 2**20 permissible weights for many more predictors. When I trained on 100k datapoints I was able to get a .55 although that was only a 2-fold CV and not an LB score so take that with a grain of salt.\r\n\r\nHow are you performing your CV? What is the logloss stopping thing you're doing and why?", "votes": null }, { "id": "153741", "postDate": "01/03/2017 05:57:09", "content": "hi, thanks @ololo . i think ad_id feature is just like ranking ad_id with their clicked rate . i got that in kernels ranking with clicked rate can get 0.63. i am not sure whether i am right.
", "rawMarkdown": "hi, thanks @ololo . i think ad_id feature is just like ranking ad_id with their clicked rate . i got that in kernels ranking with clicked rate can get 0.63. i am not sure whether i am right.", "votes": null }, { "id": "153743", "postDate": "01/03/2017 06:05:56", "content": "hi , @Vape Naysh. \n1. i use pypy and training on over 80 million datapoionts. actually, only use display_id,ad_id just need to maintain w with length of about 2**20 . so it didn't take a long time.\n2. i just use holdout = 100 for cv. not accurate. maybe overfitting give a bad result , i am trying something to avoid this.\n3. i do not know \"What is the logloss stopping thing you're doing and why\" mean.
\n\nthanks for your sugessions!
", "rawMarkdown": "hi , @Vape Naysh. \r\n1. i use pypy and training on over 80 million datapoionts. actually, only use display_id,ad_id just need to maintain w with length of about 2**20 . so it didn't take a long time.\r\n2. i just use holdout = 100 for cv. not accurate. maybe overfitting give a bad result , i am trying something to avoid this.\r\n3. i do not know \"What is the logloss stopping thing you're doing and why\" mean.\r\n\r\nthanks for your sugessions!", "votes": null }, { "id": "153744", "postDate": "01/03/2017 06:06:01", "content": "hi , @Vape Naysh. \n1. i use pypy and training on over 80 million datapoionts. actually, only use display_id,ad_id just need to maintain w with length of about 2**20 . so it didn't take a long time.\n2. i just use holdout = 100 for cv. not accurate. maybe overfitting give a bad result , i am trying something to avoid this.\n3. i do not know \"What is the logloss stopping thing you're doing and why\" mean.
\n\nthanks for your sugessions!
", "rawMarkdown": "hi , @Vape Naysh. \r\n1. i use pypy and training on over 80 million datapoionts. actually, only use display_id,ad_id just need to maintain w with length of about 2**20 . so it didn't take a long time.\r\n2. i just use holdout = 100 for cv. not accurate. maybe overfitting give a bad result , i am trying something to avoid this.\r\n3. i do not know \"What is the logloss stopping thing you're doing and why\" mean.\r\n\r\nthanks for your sugessions!", "votes": null } ], "comments": [ { "id": 153464, "author_name": "", "author_url": "", "post_date": "01/01/2017 17:00:35", "content": "use single model with ftrl.
", "votes": null, "replies": [] }, { "id": 153520, "author_name": "vapenaysh", "author_url": "", "post_date": "01/02/2017 00:59:01", "content": "I was able to get a .64 just by training on all the data using SRK's template.
", "votes": null, "replies": [] }, { "id": 153586, "author_name": "", "author_url": "", "post_date": "01/02/2017 11:31:13", "content": "i use his script too ... i try his script , but only got 0.53 on all the data ... actually , i just use displayid, ad_id and train for 8 epoch, and get result 0.54
", "votes": null, "replies": [] }, { "id": 153622, "author_name": "agrigorev", "author_url": "", "post_date": "01/02/2017 14:55:18", "content": "Display id is not a particularly good feature, try something else
", "votes": null, "replies": [] }, { "id": 153656, "author_name": "vapenaysh", "author_url": "", "post_date": "01/02/2017 17:27:33", "content": "@BIGAN\nHmmmm well i'm definitely not an expert but for me, two things come to mind. Are you sure you're training over the entire dataset? passing 8 times over 80 million datapoints must take forever to run (are you using pypy? havent had a chance to look at that). If you are training on all the data, maybe you're overfitting. The default regularization parameters are about as 'off' as they can be and running 8 passes over the data will certainly reinforce any overfitting tendencies. Not sure how strong the overfitting is since the there are only 2**20 permissible weights for many more predictors. When I trained on 100k datapoints I was able to get a .55 although that was only a 2-fold CV and not an LB score so take that with a grain of salt.
\n\nHow are you performing your CV? What is the logloss stopping thing you're doing and why?
", "votes": null, "replies": [] }, { "id": 153741, "author_name": "", "author_url": "", "post_date": "01/03/2017 05:57:09", "content": "hi, thanks @ololo . i think ad_id feature is just like ranking ad_id with their clicked rate . i got that in kernels ranking with clicked rate can get 0.63. i am not sure whether i am right.
", "votes": null, "replies": [] }, { "id": 153743, "author_name": "", "author_url": "", "post_date": "01/03/2017 06:05:56", "content": "hi , @Vape Naysh. \n1. i use pypy and training on over 80 million datapoionts. actually, only use display_id,ad_id just need to maintain w with length of about 2**20 . so it didn't take a long time.\n2. i just use holdout = 100 for cv. not accurate. maybe overfitting give a bad result , i am trying something to avoid this.\n3. i do not know \"What is the logloss stopping thing you're doing and why\" mean.
\n\nthanks for your sugessions!
", "votes": null, "replies": [] }, { "id": 153744, "author_name": "", "author_url": "", "post_date": "01/03/2017 06:06:01", "content": "hi , @Vape Naysh. \n1. i use pypy and training on over 80 million datapoionts. actually, only use display_id,ad_id just need to maintain w with length of about 2**20 . so it didn't take a long time.\n2. i just use holdout = 100 for cv. not accurate. maybe overfitting give a bad result , i am trying something to avoid this.\n3. i do not know \"What is the logloss stopping thing you're doing and why\" mean.
\n\nthanks for your sugessions!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "153457": "i use feature : display_id, ad_id, uuid, document_id, platform, country, state, dma, \r\n weekday, hour, cate, entity, topics, ad_id_2, ad_document_id, campaign_id, advertiser_id, ad_click_rate.\r\nalso try to compose some combination. \r\nlogloss stop at around 0.42 . \r\n\r\ncould anyone give some clue?", "153464": "use single model with ftrl.", "153520": "I was able to get a .64 just by training on all the data using SRK's template.", "153586": "i use his script too ... i try his script , but only got 0.53 on all the data ... actually , i just use displayid, ad_id and train for 8 epoch, and get result 0.54", "153622": "Display id is not a particularly good feature, try something else", "153656": "BIGAN\r\nHmmmm well i'm definitely not an expert but for me, two things come to mind. Are you sure you're training over the entire dataset? passing 8 times over 80 million datapoints must take forever to run (are you using pypy? havent had a chance to look at that). If you are training on all the data, maybe you're overfitting. The default regularization parameters are about as 'off' as they can be and running 8 passes over the data will certainly reinforce any overfitting tendencies. Not sure how strong the overfitting is since the there are only 2**20 permissible weights for many more predictors. When I trained on 100k datapoints I was able to get a .55 although that was only a 2-fold CV and not an LB score so take that with a grain of salt.\r\n\r\nHow are you performing your CV? What is the logloss stopping thing you're doing and why?", "153741": "hi, thanks @ololo . i think ad_id feature is just like ranking ad_id with their clicked rate . i got that in kernels ranking with clicked rate can get 0.63. i am not sure whether i am right.", "153743": "hi , @Vape Naysh. \r\n1. i use pypy and training on over 80 million datapoionts. actually, only use display_id,ad_id just need to maintain w with length of about 2**20 . so it didn't take a long time.\r\n2. i just use holdout = 100 for cv. not accurate. maybe overfitting give a bad result , i am trying something to avoid this.\r\n3. i do not know \"What is the logloss stopping thing you're doing and why\" mean.\r\n\r\nthanks for your sugessions!", "153744": "hi , @Vape Naysh. \r\n1. i use pypy and training on over 80 million datapoionts. actually, only use display_id,ad_id just need to maintain w with length of about 2**20 . so it didn't take a long time.\r\n2. i just use holdout = 100 for cv. not accurate. maybe overfitting give a bad result , i am trying something to avoid this.\r\n3. i do not know \"What is the logloss stopping thing you're doing and why\" mean.\r\n\r\nthanks for your sugessions!" }, "source": "meta" }