Maybe it is the time to predict whether there is big shake-up between public LB and the private LB?
", "messages": [ { "id": "262143", "postDate": "12/25/2017 14:21:11", "content": "Maybe it is the time to predict whether there is big shake-up between public LB and the private LB?
", "rawMarkdown": "Maybe it is the time to predict whether there is big shake-up between public LB and the private LB?", "votes": null }, { "id": "262171", "postDate": "12/25/2017 16:44:29", "content": "Definitely, there will be a big shake-up.
", "rawMarkdown": "Definitely, there will be a big shake-up.", "votes": null }, { "id": "262242", "postDate": "12/25/2017 23:56:59", "content": "I would guess the shake-up will be pretty big. The private data are further ahead in time, thus harder to predict, so methods that work for the public data may not work for the private data. Also, the private data include a holiday.
", "rawMarkdown": "I would guess the shake-up will be pretty big. The private data are further ahead in time, thus harder to predict, so methods that work for the public data may not work for the private data. Also, the private data include a holiday.", "votes": null }, { "id": "262293", "postDate": "12/26/2017 04:38:53", "content": "I think there will be a big shake down at least for me. The performance for the first five days is better than the rest days.
", "rawMarkdown": "I think there will be a big shake down at least for me. The performance for the first five days is better than the rest days.", "votes": null }, { "id": "262421", "postDate": "12/26/2017 14:44:48", "content": "the data is so big, so i think it will be a vary small shake-up......
", "rawMarkdown": "the data is so big, so i think it will be a vary small shake-up......", "votes": null }, { "id": "262448", "postDate": "12/26/2017 17:04:05", "content": "Big data, but there are biases in the public/private split (see my earlier comment).
", "rawMarkdown": "Big data, but there are biases in the public/private split (see my earlier comment).", "votes": null }, { "id": "262520", "postDate": "12/27/2017 00:10:27", "content": "May be you are right , but I don't figure out why do you think the private data are further ahead in time. consider there are some difference betweent train and test, it was hard to say harder or easier.
", "rawMarkdown": "May be you are right , but I don't figure out why do you think the private data are further ahead in time. consider there are some difference betweent train and test, it was hard to say harder or easier.", "votes": null }, { "id": "262691", "postDate": "12/27/2017 16:06:23", "content": "Surely a overfitting game.
", "rawMarkdown": "Surely a overfitting game.", "votes": null }, { "id": "262696", "postDate": "12/27/2017 16:34:15", "content": "I tried to overfit leaderboard but failed，I think the shake up depends on the different method the competitors use.
", "rawMarkdown": "I tried to overfit leaderboard but failed，I think the shake up depends on the different method the competitors use.", "votes": null }, { "id": "263120", "postDate": "12/29/2017 00:50:40", "content": "Maybe the scores will be reduced, and the rank will not change a lot.
", "rawMarkdown": "Maybe the scores will be reduced, and the rank will not change a lot.", "votes": null }, { "id": "263128", "postDate": "12/29/2017 02:17:12", "content": "I've tried various gradient boosting methods and some RNNs and it looks like validation is consistent across the board. Unlike wikipedia, we don't have to take into account long-term dependency.
", "rawMarkdown": "I've tried various gradient boosting methods and some RNNs and it looks like validation is consistent across the board. Unlike wikipedia, we don't have to take into account long-term dependency.", "votes": null }, { "id": "263143", "postDate": "12/29/2017 03:27:30", "content": "这个分会变，排行榜变动不会很大的，以前玩过携程赛，是一样的感觉
", "rawMarkdown": "这个分会变，排行榜变动不会很大的，以前玩过携程赛，是一样的感觉", "votes": null }, { "id": "263273", "postDate": "12/29/2017 15:34:26", "content": "蠻認同的！慣例，公開成績是前五天的，剩餘的應該是六到十五天！
", "rawMarkdown": "蠻認同的！慣例，公開成績是前五天的，剩餘的應該是六到十五天！", "votes": null }, { "id": "263469", "postDate": "12/30/2017 11:23:49", "content": "要是这样，还真说不准，谢谢你的回复
", "rawMarkdown": "要是这样，还真说不准，谢谢你的回复", "votes": null }, { "id": "263568", "postDate": "12/30/2017 20:03:59", "content": "basically we should focus on predicting from day 6 to 16, 21 to 31 August ... correct? I am starting to make this challenge and this is a key point ...
", "rawMarkdown": "basically we should focus on predicting from day 6 to 16, 21 to 31 August ... correct? I am starting to make this challenge and this is a key point ...", "votes": null }, { "id": "263695", "postDate": "12/31/2017 12:43:13", "content": "排名不 (会)·eps 改变
\n\nFor eps == 0 it translates as : \"Rankings will change\" 排名会改变\nFor eps == 1 it translates as : \"Rankings will not change\" 排名不会改变\n(Of course, it also works in English, but its not as cool as forcing people to use Google Translate.)
", "rawMarkdown": "排名不 (会)·eps 改变\n\n For eps == 0 it translates as : \"Rankings will change\" 排名会改变\n For eps == 1 it translates as : \"Rankings will not change\" 排名不会改变\n\n(Of course, it also works in English, but its not as cool as forcing people to use Google Translate.)", "votes": null }, { "id": "263816", "postDate": "01/01/2018 05:10:32", "content": "hahaha, lol, just ignore the comments that you don't understand. I am Chinese as well btw. But given the rising status of China in the world, I think it's not harmful to learn some simple Chinese nowadays. Stay hungry.
", "rawMarkdown": "hahaha, lol, just ignore the comments that you don't understand. I am Chinese as well btw. But given the rising status of China in the world, I think it's not harmful to learn some simple Chinese nowadays. Stay hungry.", "votes": null }, { "id": "263817", "postDate": "01/01/2018 05:14:12", "content": "", "rawMarkdown": "", "votes": null }, { "id": "263832", "postDate": "01/01/2018 07:27:39", "content": "Given the rising status of China, Chinese people may think more globally and share their ideas in a language which everybody can understand, so that they can get other ideas from all around the world. There is no problem if you write the things like \"hello\" in your language, but the part about the competition must be in English I guess.
", "rawMarkdown": "**Given the rising status of China,** Chinese people may think more globally and share their ideas in a language which everybody can understand, so that they can get other ideas from all around the world. There is no problem if you write the things like \"hello\" in your language, but the part about the competition must be in English I guess.", "votes": null }, { "id": "263921", "postDate": "01/01/2018 16:34:59", "content": "", "rawMarkdown": "", "votes": null }, { "id": "264020", "postDate": "01/02/2018 02:34:15", "content": "I feel obliged to support @AhmetErdem here. Non-native speakers of English as of 2003 outnumbered native speakers by a ratio of 3 to 1. In a sense, English does not belong to US or UK anymore. Mandarin is nowhere near that status. So I think it's presumptuous to use Mandarin in international forums.
", "rawMarkdown": "I feel obliged to support @AhmetErdem here. [Non-native speakers of English as of 2003 outnumbered native speakers by a ratio of 3 to 1][1]. In a sense, English does not belong to US or UK anymore. Mandarin is nowhere near that status. So I think it's presumptuous to use Mandarin in international forums.\n [1]: https://www.wikiwand.com/en/English-speaking_world", "votes": null }, { "id": "264049", "postDate": "01/02/2018 04:41:59", "content": "我一开始做特征就在极力避免过拟合，但愿别掉太多。\nEnglish：I was trying my best to avoid over fitting, but I would not let it go too far.（from Google Translate）
", "rawMarkdown": "我一开始做特征就在极力避免过拟合，但愿别掉太多。\nEnglish：I was trying my best to avoid over fitting, but I would not let it go too far.（from Google Translate）", "votes": null }, { "id": "264052", "postDate": "01/02/2018 04:47:03", "content": "My English is poor.I try to use Geogle Translate.But it may be ambiguous
", "rawMarkdown": "My English is poor.I try to use Geogle Translate.But it may be ambiguous", "votes": null }, { "id": "264077", "postDate": "01/02/2018 07:19:56", "content": "langjiaoshou！
", "rawMarkdown": "langjiaoshou！", "votes": null }, { "id": "264189", "postDate": "01/02/2018 14:30:31", "content": "em....about English International language, this is because of some historical reasons I think， but Chinese is really getting its momentum, so I said \"it's not harmful to learn some simple Chinese...\" apologies if it sounds haughty. Anyway it's up to you whether to learn some words of a specific language, no one can \"force\" you to learn anything.
", "rawMarkdown": "em....about English International language, this is because of some historical reasons I think， but Chinese is really getting its momentum, so I said \"it's not harmful to learn some simple Chinese...\" apologies if it sounds haughty. Anyway it's up to you whether to learn some words of a specific language, no one can \"force\" you to learn anything.", "votes": null }, { "id": "264196", "postDate": "01/02/2018 14:45:30", "content": "你是哪位....
", "rawMarkdown": "你是哪位....", "votes": null }, { "id": "264405", "postDate": "01/03/2018 01:16:21", "content": "With all due respect, you are somewhat forcing people to learn Mandarin to avoid missing out crucial information. This is a global community, but yet it feels like a tribal warfare now. Historical reason can not justify using a language only members from your tribe can understand. Unilaterally claiming the messages are irrelevant is not good enough. Why should people trust you? I can read Mandarin. I know they are at least slightly relevant.
", "rawMarkdown": "With all due respect, you are somewhat forcing people to learn Mandarin to avoid missing out crucial information. This is a global community, but yet it feels like a tribal warfare now. Historical reason can not justify using a language only members from your tribe can understand. Unilaterally claiming the messages are irrelevant is not good enough. Why should people trust you? I can read Mandarin. I know they are at least slightly relevant.", "votes": null }, { "id": "264418", "postDate": "01/03/2018 02:16:49", "content": "Dear ALL, I would suggest to finish the discussion as the original goal is to discuss the technical staff. I believe everyone should follow the role as Kaggle is an international platform, and please don't be the trouble maker.
", "rawMarkdown": "Dear ALL, I would suggest to finish the discussion as the original goal is to discuss the technical staff. I believe everyone should follow the role as Kaggle is an international platform, and please don't be the trouble maker.", "votes": null }, { "id": "264466", "postDate": "01/03/2018 04:55:52", "content": "Take it easy, man. Not trying to convincing anyone, it’s tired to convincing people on the internet. I was simply express my opinion. No one(I suppose) read every single comment in the discussion page in every competition he take part in, yet lots of kagglers perform well in the these competitions. So I said ignore the comment if you think it’s tiresome to do some clicks on google translate. On the other hand, If you google it, it’s good, because every time you google something you don’t know you get some knowledge, that’s how people learn things.
", "rawMarkdown": "Take it easy, man. Not trying to convincing anyone, it’s tired to convincing people on the internet. I was simply express my opinion. No one(I suppose) read every single comment in the discussion page in every competition he take part in, yet lots of kagglers perform well in the these competitions. So I said ignore the comment if you think it’s tiresome to do some clicks on google translate. On the other hand, If you google it, it’s good, because every time you google something you don’t know you get some knowledge, that’s how people learn things.", "votes": null }, { "id": "264502", "postDate": "01/03/2018 08:15:18", "content": "I saw your comment in somebody's CSDN blog yesterday, so I click into your blog. And I am surprised to find that Your Kaggle Username is as same as your blog Username. I just feel your experience is very inspirational, and I am a student learning machine learning myself for 7 months.
", "rawMarkdown": "I saw your comment in somebody's CSDN blog yesterday, so I click into your blog. And I am surprised to find that Your Kaggle Username is as same as your blog Username. I just feel your experience is very inspirational, and I am a student learning machine learning myself for 7 months.", "votes": null }, { "id": "264503", "postDate": "01/03/2018 08:16:41", "content": "why are you so diao !
", "rawMarkdown": "why are you so diao !", "votes": null }, { "id": "264554", "postDate": "01/03/2018 11:25:19", "content": "Common growth
", "rawMarkdown": "Common growth", "votes": null }, { "id": "264850", "postDate": "01/04/2018 02:17:37", "content": "Typical characteristic of time series forecasting: the further ahead of time, the harder it is to follow the pattern. Maybe we should analysis the sales data on August, 2016 to check the uncertain shake-up. Since August is an off-season of sales, I guess the shake-up of the whole private data is not too big. But for a specific day, it's hard to say.
", "rawMarkdown": "Typical characteristic of time series forecasting: the further ahead of time, the harder it is to follow the pattern. Maybe we should analysis the sales data on August, 2016 to check the uncertain shake-up. Since August is an off-season of sales, I guess the shake-up of the whole private data is not too big. But for a specific day, it's hard to say.", "votes": null }, { "id": "267531", "postDate": "01/11/2018 16:32:31", "content": "I think there are going to be some very large moves downwards for some people on the private board. Because of the missing promotion information, it is very easy to overfit promotion information (ie. if you fill with zeros, promo 0 -> promo 1 always means something was sold in the train set, which is not necessarily true in the test set EXCEPT on the very first day). It's also easiest to overfit the start of the time period (ie. the public board) with these types of features, which means you have to be extremely careful with methods like gradient boosting. Even cross-validation is tricky, because if you have any leaks of these types of features then cross-validation will not be accurate or consistent. It's also hard to know how to handle this (see @Feng's thread on missing info).
\n\nUnlike @Steven, I have not seen consistent validation between methods, or even within methods when I use any facet of the promotion information. With RNN I can validate from .56-.59 (not using the 40000 added item/store combos) depending on how I handle promotion information while achieving identical .511 public LB scores....
", "rawMarkdown": "I think there are going to be some very large moves downwards for some people on the private board. Because of the missing promotion information, it is very easy to overfit promotion information (ie. if you fill with zeros, promo 0 -> promo 1 always means something was sold in the train set, which is not necessarily true in the test set EXCEPT on the very first day). It's also easiest to overfit the start of the time period (ie. the public board) with these types of features, which means you have to be extremely careful with methods like gradient boosting. Even cross-validation is tricky, because if you have any leaks of these types of features then cross-validation will not be accurate or consistent. It's also hard to know how to handle this (see @Feng's thread on missing info).\n\nUnlike @Steven, I have not seen consistent validation between methods, or even within methods when I use any facet of the promotion information. With RNN I can validate from .56-.59 (not using the 40000 added item/store combos) depending on how I handle promotion information while achieving identical .511 public LB scores....", "votes": null }, { "id": "267602", "postDate": "01/11/2018 21:00:53", "content": "That is a pretty old post form my first couple models. From my LB position I'm definitely stuck at the .511 barrier with everyone but the top 5-10 teams here. ;)
", "rawMarkdown": "That is a pretty old post form my first couple models. From my LB position I'm definitely stuck at the .511 barrier with everyone but the top 5-10 teams here. ;)", "votes": null }, { "id": "267641", "postDate": "01/11/2018 23:39:49", "content": "Exactly. I got stuck on this problem for a long time and I gave up...
", "rawMarkdown": "Exactly. I got stuck on this problem for a long time and I gave up...", "votes": null }, { "id": "267959", "postDate": "01/12/2018 21:55:05", "content": "Would you like to give a hint to all of us stuck in this insane trap?
", "rawMarkdown": "Would you like to give a hint to all of us stuck in this insane trap?", "votes": null }, { "id": "268134", "postDate": "01/13/2018 13:56:10", "content": "Don't worry about the PLB too much-- @Andy Harless is right
", "rawMarkdown": "Don't worry about the PLB too much-- @Andy Harless is right", "votes": null }, { "id": "268276", "postDate": "01/14/2018 01:42:11", "content": "@piupiu, wish I had thought of your avatar first : (
", "rawMarkdown": "piupiu, wish I had thought of your avatar first : (", "votes": null }, { "id": "268280", "postDate": "01/14/2018 01:52:29", "content": "@jmbull, This is the common goal of most kaggler，wish you finish it soon : )
", "rawMarkdown": "jmbull, This is the common goal of most kaggler，wish you finish it soon : )", "votes": null }, { "id": "268283", "postDate": "01/14/2018 02:03:59", "content": "My wish is that of caterpillars. Your teaching is that with patience I may soar above the earth on my own power.
", "rawMarkdown": "My wish is that of caterpillars. Your teaching is that with patience I may soar above the earth on my own power.", "votes": null }, { "id": "268472", "postDate": "01/14/2018 15:14:52", "content": "I totally agree with the overfitting problem. I have tried to do some regularization with the drop out layer on my NN models but it works to a certain extend only.
", "rawMarkdown": "I totally agree with the overfitting problem. I have tried to do some regularization with the drop out layer on my NN models but it works to a certain extend only.", "votes": null }, { "id": "268473", "postDate": "01/14/2018 15:19:32", "content": "@Andy I have been thinking of this too. But I feel that if we tackle this forecasting problem as a regression angle which is like Ceshine Lee's approach, the error might not increase(as much) as the date goes further as compared to traditional ARIMA model which relies on the predicted previous days.
", "rawMarkdown": "Andy I have been thinking of this too. But I feel that if we tackle this forecasting problem as a regression angle which is like Ceshine Lee's approach, the error might not increase(as much) as the date goes further as compared to traditional ARIMA model which relies on the predicted previous days.", "votes": null } ], "comments": [ { "id": 262171, "author_name": "cczaixian", "author_url": "", "post_date": "12/25/2017 16:44:29", "content": "Definitely, there will be a big shake-up.
", "votes": null, "replies": [] }, { "id": 262242, "author_name": "aharless", "author_url": "", "post_date": "12/25/2017 23:56:59", "content": "I would guess the shake-up will be pretty big. The private data are further ahead in time, thus harder to predict, so methods that work for the public data may not work for the private data. Also, the private data include a holiday.
", "votes": null, "replies": [ { "id": 263568, "author_name": "pietromarinelli", "author_url": "", "post_date": "12/30/2017 20:03:59", "content": "basically we should focus on predicting from day 6 to 16, 21 to 31 August ... correct? I am starting to make this challenge and this is a key point ...
", "votes": null, "replies": [] }, { "id": 264850, "author_name": "yaotang", "author_url": "", "post_date": "01/04/2018 02:17:37", "content": "Typical characteristic of time series forecasting: the further ahead of time, the harder it is to follow the pattern. Maybe we should analysis the sales data on August, 2016 to check the uncertain shake-up. Since August is an off-season of sales, I guess the shake-up of the whole private data is not too big. But for a specific day, it's hard to say.
", "votes": null, "replies": [] }, { "id": 268473, "author_name": "sbongo", "author_url": "", "post_date": "01/14/2018 15:19:32", "content": "@Andy I have been thinking of this too. But I feel that if we tackle this forecasting problem as a regression angle which is like Ceshine Lee's approach, the error might not increase(as much) as the date goes further as compared to traditional ARIMA model which relies on the predicted previous days.
", "votes": null, "replies": [] } ] }, { "id": 262293, "author_name": "wuzuping", "author_url": "", "post_date": "12/26/2017 04:38:53", "content": "I think there will be a big shake down at least for me. The performance for the first five days is better than the rest days.
", "votes": null, "replies": [] }, { "id": 262421, "author_name": "pureheart", "author_url": "", "post_date": "12/26/2017 14:44:48", "content": "the data is so big, so i think it will be a vary small shake-up......
", "votes": null, "replies": [ { "id": 262448, "author_name": "aharless", "author_url": "", "post_date": "12/26/2017 17:04:05", "content": "Big data, but there are biases in the public/private split (see my earlier comment).
", "votes": null, "replies": [] }, { "id": 262520, "author_name": "pureheart", "author_url": "", "post_date": "12/27/2017 00:10:27", "content": "May be you are right , but I don't figure out why do you think the private data are further ahead in time. consider there are some difference betweent train and test, it was hard to say harder or easier.
", "votes": null, "replies": [] }, { "id": 268276, "author_name": "jmbull", "author_url": "", "post_date": "01/14/2018 01:42:11", "content": "@piupiu, wish I had thought of your avatar first : (
", "votes": null, "replies": [] }, { "id": 268280, "author_name": "pureheart", "author_url": "", "post_date": "01/14/2018 01:52:29", "content": "@jmbull, This is the common goal of most kaggler，wish you finish it soon : )
", "votes": null, "replies": [] }, { "id": 268283, "author_name": "jmbull", "author_url": "", "post_date": "01/14/2018 02:03:59", "content": "My wish is that of caterpillars. Your teaching is that with patience I may soar above the earth on my own power.
", "votes": null, "replies": [] } ] }, { "id": 262691, "author_name": "fengari", "author_url": "", "post_date": "12/27/2017 16:06:23", "content": "Surely a overfitting game.
", "votes": null, "replies": [] }, { "id": 262696, "author_name": "h4211819", "author_url": "", "post_date": "12/27/2017 16:34:15", "content": "I tried to overfit leaderboard but failed，I think the shake up depends on the different method the competitors use.
", "votes": null, "replies": [] }, { "id": 263120, "author_name": "zhoukj", "author_url": "", "post_date": "12/29/2017 00:50:40", "content": "Maybe the scores will be reduced, and the rank will not change a lot.
", "votes": null, "replies": [ { "id": 263128, "author_name": "stevenknguyen", "author_url": "", "post_date": "12/29/2017 02:17:12", "content": "I've tried various gradient boosting methods and some RNNs and it looks like validation is consistent across the board. Unlike wikipedia, we don't have to take into account long-term dependency.
", "votes": null, "replies": [] }, { "id": 263817, "author_name": "zhoukj", "author_url": "", "post_date": "01/01/2018 05:14:12", "content": "", "votes": null, "replies": [] } ] }, { "id": 263143, "author_name": "zhilangtaosha", "author_url": "", "post_date": "12/29/2017 03:27:30", "content": "这个分会变，排行榜变动不会很大的，以前玩过携程赛，是一样的感觉
", "votes": null, "replies": [ { "id": 263273, "author_name": "kueipo", "author_url": "", "post_date": "12/29/2017 15:34:26", "content": "蠻認同的！慣例，公開成績是前五天的，剩餘的應該是六到十五天！
", "votes": null, "replies": [] }, { "id": 263469, "author_name": "zhilangtaosha", "author_url": "", "post_date": "12/30/2017 11:23:49", "content": "要是这样，还真说不准，谢谢你的回复
", "votes": null, "replies": [] }, { "id": 264077, "author_name": "shangweichen", "author_url": "", "post_date": "01/02/2018 07:19:56", "content": "langjiaoshou！
", "votes": null, "replies": [] }, { "id": 264196, "author_name": "zhilangtaosha", "author_url": "", "post_date": "01/02/2018 14:45:30", "content": "你是哪位....
", "votes": null, "replies": [] }, { "id": 264502, "author_name": "shangweichen", "author_url": "", "post_date": "01/03/2018 08:15:18", "content": "I saw your comment in somebody's CSDN blog yesterday, so I click into your blog. And I am surprised to find that Your Kaggle Username is as same as your blog Username. I just feel your experience is very inspirational, and I am a student learning machine learning myself for 7 months.
", "votes": null, "replies": [] }, { "id": 264554, "author_name": "zhilangtaosha", "author_url": "", "post_date": "01/03/2018 11:25:19", "content": "Common growth
", "votes": null, "replies": [] } ] }, { "id": 263695, "author_name": "batangas", "author_url": "", "post_date": "12/31/2017 12:43:13", "content": "排名不 (会)·eps 改变
\n\nFor eps == 0 it translates as : \"Rankings will change\" 排名会改变\nFor eps == 1 it translates as : \"Rankings will not change\" 排名不会改变\n(Of course, it also works in English, but its not as cool as forcing people to use Google Translate.)
", "votes": null, "replies": [ { "id": 263816, "author_name": "houxudan63", "author_url": "", "post_date": "01/01/2018 05:10:32", "content": "hahaha, lol, just ignore the comments that you don't understand. I am Chinese as well btw. But given the rising status of China in the world, I think it's not harmful to learn some simple Chinese nowadays. Stay hungry.
", "votes": null, "replies": [] }, { "id": 263832, "author_name": "aerdem4", "author_url": "", "post_date": "01/01/2018 07:27:39", "content": "Given the rising status of China, Chinese people may think more globally and share their ideas in a language which everybody can understand, so that they can get other ideas from all around the world. There is no problem if you write the things like \"hello\" in your language, but the part about the competition must be in English I guess.
", "votes": null, "replies": [] }, { "id": 263921, "author_name": "houxudan63", "author_url": "", "post_date": "01/01/2018 16:34:59", "content": "", "votes": null, "replies": [] }, { "id": 264020, "author_name": "ceshine", "author_url": "", "post_date": "01/02/2018 02:34:15", "content": "I feel obliged to support @AhmetErdem here. Non-native speakers of English as of 2003 outnumbered native speakers by a ratio of 3 to 1. In a sense, English does not belong to US or UK anymore. Mandarin is nowhere near that status. So I think it's presumptuous to use Mandarin in international forums.
", "votes": null, "replies": [] }, { "id": 264052, "author_name": "xuanhahaha", "author_url": "", "post_date": "01/02/2018 04:47:03", "content": "My English is poor.I try to use Geogle Translate.But it may be ambiguous
", "votes": null, "replies": [] }, { "id": 264189, "author_name": "houxudan63", "author_url": "", "post_date": "01/02/2018 14:30:31", "content": "em....about English International language, this is because of some historical reasons I think， but Chinese is really getting its momentum, so I said \"it's not harmful to learn some simple Chinese...\" apologies if it sounds haughty. Anyway it's up to you whether to learn some words of a specific language, no one can \"force\" you to learn anything.
", "votes": null, "replies": [] }, { "id": 264405, "author_name": "ceshine", "author_url": "", "post_date": "01/03/2018 01:16:21", "content": "With all due respect, you are somewhat forcing people to learn Mandarin to avoid missing out crucial information. This is a global community, but yet it feels like a tribal warfare now. Historical reason can not justify using a language only members from your tribe can understand. Unilaterally claiming the messages are irrelevant is not good enough. Why should people trust you? I can read Mandarin. I know they are at least slightly relevant.
", "votes": null, "replies": [] }, { "id": 264466, "author_name": "houxudan63", "author_url": "", "post_date": "01/03/2018 04:55:52", "content": "Take it easy, man. Not trying to convincing anyone, it’s tired to convincing people on the internet. I was simply express my opinion. No one(I suppose) read every single comment in the discussion page in every competition he take part in, yet lots of kagglers perform well in the these competitions. So I said ignore the comment if you think it’s tiresome to do some clicks on google translate. On the other hand, If you google it, it’s good, because every time you google something you don’t know you get some knowledge, that’s how people learn things.
", "votes": null, "replies": [] } ] }, { "id": 264049, "author_name": "xuanhahaha", "author_url": "", "post_date": "01/02/2018 04:41:59", "content": "我一开始做特征就在极力避免过拟合，但愿别掉太多。\nEnglish：I was trying my best to avoid over fitting, but I would not let it go too far.（from Google Translate）
", "votes": null, "replies": [ { "id": 264503, "author_name": "shangweichen", "author_url": "", "post_date": "01/03/2018 08:16:41", "content": "why are you so diao !
", "votes": null, "replies": [] } ] }, { "id": 264418, "author_name": "kelexu", "author_url": "", "post_date": "01/03/2018 02:16:49", "content": "Dear ALL, I would suggest to finish the discussion as the original goal is to discuss the technical staff. I believe everyone should follow the role as Kaggle is an international platform, and please don't be the trouble maker.
", "votes": null, "replies": [] }, { "id": 267531, "author_name": "rhgrossm", "author_url": "", "post_date": "01/11/2018 16:32:31", "content": "I think there are going to be some very large moves downwards for some people on the private board. Because of the missing promotion information, it is very easy to overfit promotion information (ie. if you fill with zeros, promo 0 -> promo 1 always means something was sold in the train set, which is not necessarily true in the test set EXCEPT on the very first day). It's also easiest to overfit the start of the time period (ie. the public board) with these types of features, which means you have to be extremely careful with methods like gradient boosting. Even cross-validation is tricky, because if you have any leaks of these types of features then cross-validation will not be accurate or consistent. It's also hard to know how to handle this (see @Feng's thread on missing info).
\n\nUnlike @Steven, I have not seen consistent validation between methods, or even within methods when I use any facet of the promotion information. With RNN I can validate from .56-.59 (not using the 40000 added item/store combos) depending on how I handle promotion information while achieving identical .511 public LB scores....
", "votes": null, "replies": [ { "id": 267602, "author_name": "stevenknguyen", "author_url": "", "post_date": "01/11/2018 21:00:53", "content": "That is a pretty old post form my first couple models. From my LB position I'm definitely stuck at the .511 barrier with everyone but the top 5-10 teams here. ;)
", "votes": null, "replies": [] }, { "id": 267641, "author_name": "beedata", "author_url": "", "post_date": "01/11/2018 23:39:49", "content": "Exactly. I got stuck on this problem for a long time and I gave up...
", "votes": null, "replies": [] }, { "id": 267959, "author_name": "stevenknguyen", "author_url": "", "post_date": "01/12/2018 21:55:05", "content": "Would you like to give a hint to all of us stuck in this insane trap?
", "votes": null, "replies": [] }, { "id": 268134, "author_name": "rhgrossm", "author_url": "", "post_date": "01/13/2018 13:56:10", "content": "Don't worry about the PLB too much-- @Andy Harless is right
", "votes": null, "replies": [] }, { "id": 268472, "author_name": "sbongo", "author_url": "", "post_date": "01/14/2018 15:14:52", "content": "I totally agree with the overfitting problem. I have tried to do some regularization with the drop out layer on my NN models but it works to a certain extend only.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "262143": "Maybe it is the time to predict whether there is big shake-up between public LB and the private LB?", "262171": "Definitely, there will be a big shake-up.", "262242": "I would guess the shake-up will be pretty big. The private data are further ahead in time, thus harder to predict, so methods that work for the public data may not work for the private data. Also, the private data include a holiday.", "262293": "I think there will be a big shake down at least for me. The performance for the first five days is better than the rest days.", "262421": "the data is so big, so i think it will be a vary small shake-up......", "262448": "Big data, but there are biases in the public/private split (see my earlier comment).", "262520": "May be you are right , but I don't figure out why do you think the private data are further ahead in time. consider there are some difference betweent train and test, it was hard to say harder or easier.", "262691": "Surely a overfitting game.", "262696": "I tried to overfit leaderboard but failed，I think the shake up depends on the different method the competitors use.", "263120": "Maybe the scores will be reduced, and the rank will not change a lot.", "263128": "I've tried various gradient boosting methods and some RNNs and it looks like validation is consistent across the board. Unlike wikipedia, we don't have to take into account long-term dependency.", "263143": "这个分会变，排行榜变动不会很大的，以前玩过携程赛，是一样的感觉", "263273": "蠻認同的！慣例，公開成績是前五天的，剩餘的應該是六到十五天！", "263469": "要是这样，还真说不准，谢谢你的回复", "263568": "basically we should focus on predicting from day 6 to 16, 21 to 31 August ... correct? I am starting to make this challenge and this is a key point ...", "263695": "排名不 (会)·eps 改变\n\n For eps == 0 it translates as : \"Rankings will change\" 排名会改变\n For eps == 1 it translates as : \"Rankings will not change\" 排名不会改变\n\n(Of course, it also works in English, but its not as cool as forcing people to use Google Translate.)", "263816": "hahaha, lol, just ignore the comments that you don't understand. I am Chinese as well btw. But given the rising status of China in the world, I think it's not harmful to learn some simple Chinese nowadays. Stay hungry.", "263817": "", "263832": "**Given the rising status of China,** Chinese people may think more globally and share their ideas in a language which everybody can understand, so that they can get other ideas from all around the world. There is no problem if you write the things like \"hello\" in your language, but the part about the competition must be in English I guess.", "263921": "", "264020": "I feel obliged to support @AhmetErdem here. [Non-native speakers of English as of 2003 outnumbered native speakers by a ratio of 3 to 1][1]. In a sense, English does not belong to US or UK anymore. Mandarin is nowhere near that status. So I think it's presumptuous to use Mandarin in international forums.\n [1]: https://www.wikiwand.com/en/English-speaking_world", "264049": "我一开始做特征就在极力避免过拟合，但愿别掉太多。\nEnglish：I was trying my best to avoid over fitting, but I would not let it go too far.（from Google Translate）", "264052": "My English is poor.I try to use Geogle Translate.But it may be ambiguous", "264077": "langjiaoshou！", "264189": "em....about English International language, this is because of some historical reasons I think， but Chinese is really getting its momentum, so I said \"it's not harmful to learn some simple Chinese...\" apologies if it sounds haughty. Anyway it's up to you whether to learn some words of a specific language, no one can \"force\" you to learn anything.", "264196": "你是哪位....", "264405": "With all due respect, you are somewhat forcing people to learn Mandarin to avoid missing out crucial information. This is a global community, but yet it feels like a tribal warfare now. Historical reason can not justify using a language only members from your tribe can understand. Unilaterally claiming the messages are irrelevant is not good enough. Why should people trust you? I can read Mandarin. I know they are at least slightly relevant.", "264418": "Dear ALL, I would suggest to finish the discussion as the original goal is to discuss the technical staff. I believe everyone should follow the role as Kaggle is an international platform, and please don't be the trouble maker.", "264466": "Take it easy, man. Not trying to convincing anyone, it’s tired to convincing people on the internet. I was simply express my opinion. No one(I suppose) read every single comment in the discussion page in every competition he take part in, yet lots of kagglers perform well in the these competitions. So I said ignore the comment if you think it’s tiresome to do some clicks on google translate. On the other hand, If you google it, it’s good, because every time you google something you don’t know you get some knowledge, that’s how people learn things.", "264502": "I saw your comment in somebody's CSDN blog yesterday, so I click into your blog. And I am surprised to find that Your Kaggle Username is as same as your blog Username. I just feel your experience is very inspirational, and I am a student learning machine learning myself for 7 months.", "264503": "why are you so diao !", "264554": "Common growth", "264850": "Typical characteristic of time series forecasting: the further ahead of time, the harder it is to follow the pattern. Maybe we should analysis the sales data on August, 2016 to check the uncertain shake-up. Since August is an off-season of sales, I guess the shake-up of the whole private data is not too big. But for a specific day, it's hard to say.", "267531": "I think there are going to be some very large moves downwards for some people on the private board. Because of the missing promotion information, it is very easy to overfit promotion information (ie. if you fill with zeros, promo 0 -> promo 1 always means something was sold in the train set, which is not necessarily true in the test set EXCEPT on the very first day). It's also easiest to overfit the start of the time period (ie. the public board) with these types of features, which means you have to be extremely careful with methods like gradient boosting. Even cross-validation is tricky, because if you have any leaks of these types of features then cross-validation will not be accurate or consistent. It's also hard to know how to handle this (see @Feng's thread on missing info).\n\nUnlike @Steven, I have not seen consistent validation between methods, or even within methods when I use any facet of the promotion information. With RNN I can validate from .56-.59 (not using the 40000 added item/store combos) depending on how I handle promotion information while achieving identical .511 public LB scores....", "267602": "That is a pretty old post form my first couple models. From my LB position I'm definitely stuck at the .511 barrier with everyone but the top 5-10 teams here. ;)", "267641": "Exactly. I got stuck on this problem for a long time and I gave up...", "267959": "Would you like to give a hint to all of us stuck in this insane trap?", "268134": "Don't worry about the PLB too much-- @Andy Harless is right", "268276": "piupiu, wish I had thought of your avatar first : (", "268280": "jmbull, This is the common goal of most kaggler，wish you finish it soon : )", "268283": "My wish is that of caterpillars. Your teaching is that with patience I may soar above the earth on my own power.", "268472": "I totally agree with the overfitting problem. I have tried to do some regularization with the drop out layer on my NN models but it works to a certain extend only.", "268473": "Andy I have been thinking of this too. But I feel that if we tackle this forecasting problem as a regression angle which is like Ceshine Lee's approach, the error might not increase(as much) as the date goes further as compared to traditional ARIMA model which relies on the predicted previous days." }, "source": "meta" }