![\"dd\"](\"http://imglf5.nosdn.127.net/img/OEUzbU5jenB4YlRsQWNNcnk3Um5DVnFkc0Y0eEQyMFQzRXRlRXRBWW5NakMyU28yQnc4MS9nPT0.jpeg?imageView&thumbnail=2086y226&type=jpg&quality=96&stripmeta=0&type=jpg\")
I use two models to solve this problem
\nLightGBM, it's based on Ceshine lee's script, but I add some other new features.
\na) zero crossing rate (https://en.wikipedia.org/wiki/Zero-crossing_rate), I create this feature for different windows(1, 3, 7, 14, 30, 60) and levels(item level, store*family level)
\nbelow is my implementation
def corss_zero(row):
\n return (((row[:-1] * row[1:]) == 0) * ~(row[:-1] == row[1:])).sum()
b) zero-sales counts in different windows
\n: == 0) \n : == 0) \n : == 0) \nb) lastest promo day in last 30 days
\n fo_in_row(row): \n = \n i in row: \n += \n i == : \n j \n - \n\"lastest_promo_30\": get_timespan(promo_2017, t2017, 30, 30).apply(lambda row: fo_in_row(row[::-1]), axis=1).values.ravel()
\nsmall tricks.
\nI separate Perishable and Non-Perishable variable to train different models. and this give me a little improvement.
I used the same framework recommended by @Ceshine lee, but I paid careful attention to first 5 days and last 11 days, I only add variables that both have improvement to first 5 days, last 11 days and Public Board, this give me a rather steady improvement on both public and private.
", "comments": [ { "id": 269179, "author_name": "CPMP", "author_url": "", "post_date": "2018-01-16T11:54:34.843000", "content": "Nice!
", "votes": 1, "replies": [ { "id": 269213, "author_name": "🐳鲲(China)", "author_url": "", "post_date": "2018-01-16T13:31:32.033000", "content": "thanks @CPMP, I was trying to use one year data before the same as your script in WTF but get bad result - -, but your script gives me some ideas! thanks for sharing!!
", "votes": 0, "replies": [] }, { "id": 269215, "author_name": "CPMP", "author_url": "", "post_date": "2018-01-16T13:33:20.903000", "content": "Yes, training one previous year data as I did in WTF does not work here. I lost a week trying to make it work :(
", "votes": 1, "replies": [] }, { "id": 269227, "author_name": "🐳鲲(China)", "author_url": "", "post_date": "2018-01-16T13:51:59.423000", "content": "I lost one day, lol~ I was trying to make my feedforward NN performance as good as yours, but failed, only 0.521 in private, but I should do better in the next time :）
", "votes": 1, "replies": [] }, { "id": 269230, "author_name": "CPMP", "author_url": "", "post_date": "2018-01-16T13:56:10.347000", "content": "Let's team next time!
", "votes": 1, "replies": [] }, { "id": 269237, "author_name": "🐳鲲(China)", "author_url": "", "post_date": "2018-01-16T14:09:12.040000", "content": "My door is always opening! thanks for your invitation, sir🙈!
", "votes": 1, "replies": [] } ] }, { "id": 269011, "author_name": "Gerryl", "author_url": "", "post_date": "2018-01-16T03:51:34.013000", "content": "Thanks for sharing!
", "votes": 1, "replies": [ { "id": 269071, "author_name": "🐳鲲(China)", "author_url": "", "post_date": "2018-01-16T06:15:29.120000", "content": "🙂
", "votes": 0, "replies": [] } ] }, { "id": 275225, "author_name": "shark", "author_url": "", "post_date": "2018-01-28T14:19:01.523000", "content": "hello 鲲(China).How does zero crossing rate features perform?Does them rank top 10% in feature importance?
", "votes": 0, "replies": [ { "id": 275369, "author_name": "🐳鲲(China)", "author_url": "", "post_date": "2018-01-29T00:41:57.033000", "content": "Sure, but the function that i written is not optimized. ɿ😏ɾ
", "votes": 0, "replies": [] }, { "id": 277054, "author_name": "shark", "author_url": "", "post_date": "2018-02-02T03:45:31.303000", "content": "I'm wonder how do you optimized this function.It seems amazing.
", "votes": 0, "replies": [] }, { "id": 277091, "author_name": "🐳鲲(China)", "author_url": "", "post_date": "2018-02-02T05:26:55.843000", "content": "I didn't optimized it - -!
", "votes": 0, "replies": [] } ] }, { "id": 269626, "author_name": "feng", "author_url": "", "post_date": "2018-01-17T03:18:56.293000", "content": "Thanks for sharing!\nThere some question for me, could you share some insights?
\n\nIf there is improvement to first 5 days, last 11 days, then the improvement to the full 16 days? or why you didn't use the full 16 days to calculate the score?
\n\nI use the same validation framework, but I only use the full 16 days to evaluate, but the improve of local validation score result in the decrease of public board score.
", "votes": 0, "replies": [ { "id": 271040, "author_name": "🐳鲲(China)", "author_url": "", "post_date": "2018-01-19T15:06:46.570000", "content": "Improvement to the full 16 days is not equal to improvement to both first 5 and last 11 days, the leadboard is cut by time. Sometimes, improvement to first 5 day, but decrease to last 11 days. The private board is about the last 11 days, so we should choose features that do both good to first 5 and last 11.
", "votes": 3, "replies": [] } ] } ], "comment_count": 14, "messages": [ { "id": 268981, "postDate": "2018-01-16T02:44:34.283Z", "content": "
I use two models to solve this problem
\nLightGBM, it's based on Ceshine lee's script, but I add some other new features.
\na) zero crossing rate (https://en.wikipedia.org/wiki/Zero-crossing_rate), I create this feature for different windows(1, 3, 7, 14, 30, 60) and levels(item level, store*family level)
\nbelow is my implementation
def corss_zero(row):
\n return (((row[:-1] * row[1:]) == 0) * ~(row[:-1] == row[1:])).sum()
b) zero-sales counts in different windows
\n: == 0) \n : == 0) \n : == 0) \nb) lastest promo day in last 30 days
\n fo_in_row(row): \n = \n i in row: \n += \n i == : \n j \n - \n\"lastest_promo_30\": get_timespan(promo_2017, t2017, 30, 30).apply(lambda row: fo_in_row(row[::-1]), axis=1).values.ravel()
\nsmall tricks.
\nI separate Perishable and Non-Perishable variable to train different models. and this give me a little improvement.
I used the same framework recommended by @Ceshine lee, but I paid careful attention to first 5 days and last 11 days, I only add variables that both have improvement to first 5 days, last 11 days and Public Board, this give me a rather steady improvement on both public and private.
", "rawMarkdown": "![dd][1]\n\n# Models\nI use two models to solve this problem\n\n1. LightGBM, it's based on Ceshine lee's script, but I add some other new features.\na) zero crossing rate (https://en.wikipedia.org/wiki/Zero-crossing_rate), I create this feature for different windows(1, 3, 7, 14, 30, 60) and levels(item level, store*family level)\nbelow is my implementation \n\n def corss_zero(row):\n return (((row[:-1] * row[1:]) == 0) * ~(row[:-1] == row[1:])).sum()\n\nb) zero-sales counts in different windows \n\n \"zeros_28_2017\": (get_timespan(df_2017, t2017, 28, 28) == 0).sum(axis=1).values \n \"zeros_56_2017\": (get_timespan(df_2017, t2017, 56, 56) == 0).sum(axis=1).values \n \"zeros_84_2017\": (get_timespan(df_2017, t2017, 84, 84) == 0).sum(axis=1).values \n\nb) lastest promo day in last 30 days \n\n def fo_in_row(row): \n j = 0 \n for i in row: \n j += 1 \n if i == 1: \n return j \n return -1 \n\"lastest_promo_30\": get_timespan(promo_2017, t2017, 30, 30).apply(lambda row: fo_in_row(row[::-1]), axis=1).values.ravel() \n2. NN\nI used the simplest feed-forward neural network.\nsame features as LightGBM, for pre-processing, I used ”RankGauss“ recommended by @Michael Jahrer(https://www.kaggle.com/c/porto-seguro-safe-driver-prediction/discussion/44629), for training I used BatchNorman, big dropout and nothing else.\n\nsmall tricks.\nI separate Perishable and Non-Perishable variable to train different models. and this give me a little improvement.\n\n# Validation Framework\nI used the same framework recommended by @Ceshine lee, but I paid careful attention to first 5 days and last 11 days, I only add variables that both have improvement to first 5 days, last 11 days and Public Board, this give me a rather steady improvement on both public and private. \n\n\n [1]: http://imglf5.nosdn.127.net/img/OEUzbU5jenB4YlRsQWNNcnk3Um5DVnFkc0Y0eEQyMFQzRXRlRXRBWW5NakMyU28yQnc4MS9nPT0.jpeg?imageView&thumbnail=2086y226&type=jpg&quality=96&stripmeta=0&type=jpg", "votes": 18 }, { "id": 269179, "postDate": "2018-01-16T11:54:34.843Z", "content": "Nice!
", "rawMarkdown": "Nice!", "votes": 1, "replies": [ { "id": 269213, "postDate": "2018-01-16T13:31:32.033Z", "content": "thanks @CPMP, I was trying to use one year data before the same as your script in WTF but get bad result - -, but your script gives me some ideas! thanks for sharing!!
", "rawMarkdown": "thanks @CPMP, I was trying to use one year data before the same as your script in WTF but get bad result - -, but your script gives me some ideas! thanks for sharing!!" }, { "id": 269215, "postDate": "2018-01-16T13:33:20.903Z", "content": "Yes, training one previous year data as I did in WTF does not work here. I lost a week trying to make it work :(
", "rawMarkdown": "Yes, training one previous year data as I did in WTF does not work here. I lost a week trying to make it work :(", "votes": 1 }, { "id": 269227, "postDate": "2018-01-16T13:51:59.423Z", "content": "I lost one day, lol~ I was trying to make my feedforward NN performance as good as yours, but failed, only 0.521 in private, but I should do better in the next time :）
", "rawMarkdown": "I lost one day, lol~ I was trying to make my feedforward NN performance as good as yours, but failed, only 0.521 in private, but I should do better in the next time :） ", "votes": 1 }, { "id": 269230, "postDate": "2018-01-16T13:56:10.347Z", "content": "Let's team next time!
", "rawMarkdown": "Let's team next time!", "votes": 1 }, { "id": 269237, "postDate": "2018-01-16T14:09:12.040Z", "content": "My door is always opening! thanks for your invitation, sir🙈!
", "rawMarkdown": "My door is always opening! thanks for your invitation, sir🙈!", "votes": 1 } ] }, { "id": 269011, "postDate": "2018-01-16T03:51:34.013Z", "content": "Thanks for sharing!
", "rawMarkdown": "Thanks for sharing!", "votes": 1, "replies": [ { "id": 269071, "postDate": "2018-01-16T06:15:29.120Z", "content": "🙂
", "rawMarkdown": "🙂" } ] }, { "id": 275225, "postDate": "2018-01-28T14:19:01.523Z", "content": "hello 鲲(China).How does zero crossing rate features perform?Does them rank top 10% in feature importance?
", "rawMarkdown": "hello 鲲(China).How does zero crossing rate features perform?Does them rank top 10% in feature importance?\n", "replies": [ { "id": 275369, "postDate": "2018-01-29T00:41:57.033Z", "content": "Sure, but the function that i written is not optimized. ɿ😏ɾ
", "rawMarkdown": "Sure, but the function that i written is not optimized. ɿ😏ɾ" }, { "id": 277054, "postDate": "2018-02-02T03:45:31.303Z", "content": "I'm wonder how do you optimized this function.It seems amazing.
", "rawMarkdown": "I'm wonder how do you optimized this function.It seems amazing." }, { "id": 277091, "postDate": "2018-02-02T05:26:55.843Z", "content": "I didn't optimized it - -!
", "rawMarkdown": "I didn't optimized it - -!" } ] }, { "id": 269626, "postDate": "2018-01-17T03:18:56.293Z", "content": "Thanks for sharing!\nThere some question for me, could you share some insights?
\n\nIf there is improvement to first 5 days, last 11 days, then the improvement to the full 16 days? or why you didn't use the full 16 days to calculate the score?
\n\nI use the same validation framework, but I only use the full 16 days to evaluate, but the improve of local validation score result in the decrease of public board score.
", "rawMarkdown": "Thanks for sharing!\nThere some question for me, could you share some insights?\n\nIf there is improvement to first 5 days, last 11 days, then the improvement to the full 16 days? or why you didn't use the full 16 days to calculate the score?\n\nI use the same validation framework, but I only use the full 16 days to evaluate, but the improve of local validation score result in the decrease of public board score. \n", "replies": [ { "id": 271040, "postDate": "2018-01-19T15:06:46.570Z", "content": "Improvement to the full 16 days is not equal to improvement to both first 5 and last 11 days, the leadboard is cut by time. Sometimes, improvement to first 5 day, but decrease to last 11 days. The private board is about the last 11 days, so we should choose features that do both good to first 5 and last 11.
", "rawMarkdown": "Improvement to the full 16 days is not equal to improvement to both first 5 and last 11 days, the leadboard is cut by time. Sometimes, improvement to first 5 day, but decrease to last 11 days. The private board is about the last 11 days, so we should choose features that do both good to first 5 and last 11.", "votes": 3 } ] } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T08:37:05Z", "competition_id": 7391, "competition_slug": "favorita-grocery-sales-forecasting", "team_id": 1014225, "solution_url": "" }