Hello.\nThis is what I tried. But I didn't get a good grade(~0.25). What I would have missed.\nAlso, I hope this could help you.
\n\n-- Numeric Data--
\n\n--Date Data--
\n\n--Categorical Data--
\n\n, pointing out my mistakes, thank you.
", "messages": [ { "id": "137974", "postDate": "10/06/2016 11:15:51", "content": "Hello.\nThis is what I tried. But I didn't get a good grade(~0.25). What I would have missed.\nAlso, I hope this could help you.
\n\n-- Numeric Data--
\n\n--Date Data--
\n\n--Categorical Data--
\n\n, pointing out my mistakes, thank you.
", "rawMarkdown": "Hello.\r\nThis is what I tried. But I didn't get a good grade(~0.25). What I would have missed.\r\nAlso, I hope this could help you.\r\n\r\n-- Numeric Data--\r\n\r\n - Fourier transform,\r\n - ftrl, \r\n - logistic regression, \r\n - nearest neighbor, \r\n - SVM,\r\n - count number of NA (each station), \r\n - count Negative/positive value (each station), \r\n - difference, \r\n - min/max/avg/SD (each station)\r\n - cumulative sum\r\n....\r\n\r\n\r\n\r\n--Date Data--\r\n\r\n - working hours (max_value - min_value), \r\n - standardization,\r\n - difference\r\n...\r\n\r\n\r\n--Categorical Data--\r\n\r\n - sparse matrix \r\n\r\n\r\n, pointing out my mistakes, thank you.", "votes": null }, { "id": "137988", "postDate": "10/06/2016 13:50:45", "content": "A very prominent absence from your features are error rate features. Since train and test data sets overlap, you can determine the failure rate (e.g. on a machine level) for different time intervals and use these as features.
", "rawMarkdown": "A very prominent absence from your features are error rate features. Since train and test data sets overlap, you can determine the failure rate (e.g. on a machine level) for different time intervals and use these as features.", "votes": null }, { "id": "138000", "postDate": "10/06/2016 14:32:09", "content": "I am also in similar condition. Max CV score that I am able to attain is 0.252243.
\nThe features that I tried are -
--Numeric File--
\n1. Number of Non-NAN entries in each Id.
\n2. Number of Non-NAN entries in each Id per stationwise & linewise.
\n3. Sum of Numeric features per stationwise & linewise for each Id.
--Date File--
\n\nAs was shown by beluga in his "Date Exploration" Kernel that 0.01 = 6 min. I ordered the date file in ascending order as per station S3 since it has least NaNs (This I did so that my data would be arranged in ascending time order. So that I can find any pattern, if there is any between Time & Ids. I am assuming that Bosch has anonymized the earliest date in the data as 0). Then I calculated -
\n1. Min_Time & Max_Time for each Id.
\n2. Time_Difference (Max_Time - Min_Time) for each Id.
\n3. Since 0.01= 6 min. This gives 1 hr = 0.1, 1 day = 2.39, 1 week = 16.75, 1 month = 71.7, 1 year = 872.35.
\nFrom this I created year feature, week feature, day of week feature, hour feature. Day of week & hour feature were furthur divided into Start_Day, End_Day & Start_Hour, End_Hour respectively.
Apart from above features, I am also thinking of creating some more features mentioned below -
\n\n@Abhilash Awasthi
\n\noh, I also tried it. As mentioned by @gingerman, it is very prominent absence. \nThat's because we typically begin at the wrong starting point.
\n\nI spend lot of time, and i tried various method. But i faced false starts and mistakes.
\n\nWhat’s the best way to analysis data and feature engineering?
\n\nI am desperate for any suggestion kaggler can offer.
\n\nThanks.
", "rawMarkdown": "Abhilash Awasthi \r\n\r\noh, I also tried it. As mentioned by @gingerman, it is very prominent absence. \r\nThat's because we typically begin at the wrong starting point.\r\n\r\n----------\r\n\r\nI spend lot of time, and i tried various method. But i faced false starts and mistakes.\r\n\r\nWhat’s the best way to analysis data and feature engineering?\r\n\r\nI am desperate for any suggestion kaggler can offer. \r\n\r\nThanks.", "votes": null }, { "id": "138126", "postDate": "10/07/2016 10:02:41", "content": "I cannot call myself a seasoned Kaggler but this was my first competiton that I tried to actively compete in over a longer time. This was in part because of the unavailibility of running strong scripts directly from the platform which constantly through me back and frustrated me in other competitions. Still, it has been a challenging trip - the first weeks I have made millions of silly mistakes like wrong merge keys, putting to many features into the model at the same time, etc. But my biggest mistake has initially been thinking to much about tuning xgboost, playing around with the MCC rounding rule (most of my submissions are identical data with different roundings) and simply not realizing that it is just not that simple.
\n\nTaking a step back, thinking a little I started combining the data sets especially taking into account temporal structure and the manufacturing perspective of paths through the factory. It was this thinking plus the upcoming discussion on the magical feature which made me look harder and being more successful in implementing structurally better solutions than incremental, minimal improvements due to fine-tuning of the ML-algorithm. Essentially, the latter I take it can be done in the last week, together with ensembling, stacking etc. Before you even start thinking about it one should construct simple models with strong meta-features. And to this end, this competition offers a host of opportunities. What I found especially appealing was the fact that for once one could easily understand the underlying reality despite the anonymization of the data. See for example my attempt at plotting the shopfloor: https://www.kaggle.com/gingerman/bosch-production-line-performance/shopfloor-visualization
", "rawMarkdown": "I cannot call myself a seasoned Kaggler but this was my first competiton that I tried to actively compete in over a longer time. This was in part because of the unavailibility of running strong scripts directly from the platform which constantly through me back and frustrated me in other competitions. Still, it has been a challenging trip - the first weeks I have made millions of silly mistakes like wrong merge keys, putting to many features into the model at the same time, etc. But my biggest mistake has initially been thinking to much about tuning xgboost, playing around with the MCC rounding rule (most of my submissions are identical data with different roundings) and simply not realizing that it is just not that simple.\r\n\r\nTaking a step back, thinking a little I started combining the data sets especially taking into account temporal structure and the manufacturing perspective of paths through the factory. It was this thinking plus the upcoming discussion on the magical feature which made me look harder and being more successful in implementing structurally better solutions than incremental, minimal improvements due to fine-tuning of the ML-algorithm. Essentially, the latter I take it can be done in the last week, together with ensembling, stacking etc. Before you even start thinking about it one should construct simple models with strong meta-features. And to this end, this competition offers a host of opportunities. What I found especially appealing was the fact that for once one could easily understand the underlying reality despite the anonymization of the data. See for example my attempt at plotting the shopfloor: https://www.kaggle.com/gingerman/bosch-production-line-performance/shopfloor-visualization", "votes": null }, { "id": "138197", "postDate": "10/07/2016 16:08:09", "content": "@gingerman. Thank you very much.
\n\nSomethimes, Many people lose time tuning ML-algorithm, like me.\nHowever, 'Looking Data' is always the most important thing.\nGet back to basics and see if there is any way I can improve my prediction.
", "rawMarkdown": "gingerman. Thank you very much. \r\n\r\nSomethimes, Many people lose time tuning ML-algorithm, like me.\r\nHowever, 'Looking Data' is always the most important thing.\r\nGet back to basics and see if there is any way I can improve my prediction.", "votes": null }, { "id": "138594", "postDate": "10/09/2016 17:26:28", "content": "Noob question. How do you use fourier transform and turn it into a feature? I thought about this for a long time but i cant seem to make sense of it
", "rawMarkdown": "Noob question. How do you use fourier transform and turn it into a feature? I thought about this for a long time but i cant seem to make sense of it", "votes": null }, { "id": "480131", "postDate": "02/27/2019 20:15:20", "content": "Hey did you find the answer? If, yes please share it.. It will be helpful for me
", "rawMarkdown": "Hey did you find the answer? If, yes please share it.. It will be helpful for me", "votes": null } ], "comments": [ { "id": 137988, "author_name": "gingerman", "author_url": "", "post_date": "10/06/2016 13:50:45", "content": "A very prominent absence from your features are error rate features. Since train and test data sets overlap, you can determine the failure rate (e.g. on a machine level) for different time intervals and use these as features.
", "votes": null, "replies": [] }, { "id": 138000, "author_name": "abhilashawasthi", "author_url": "", "post_date": "10/06/2016 14:32:09", "content": "I am also in similar condition. Max CV score that I am able to attain is 0.252243.
\nThe features that I tried are -
--Numeric File--
\n1. Number of Non-NAN entries in each Id.
\n2. Number of Non-NAN entries in each Id per stationwise & linewise.
\n3. Sum of Numeric features per stationwise & linewise for each Id.
--Date File--
\n\nAs was shown by beluga in his "Date Exploration" Kernel that 0.01 = 6 min. I ordered the date file in ascending order as per station S3 since it has least NaNs (This I did so that my data would be arranged in ascending time order. So that I can find any pattern, if there is any between Time & Ids. I am assuming that Bosch has anonymized the earliest date in the data as 0). Then I calculated -
\n1. Min_Time & Max_Time for each Id.
\n2. Time_Difference (Max_Time - Min_Time) for each Id.
\n3. Since 0.01= 6 min. This gives 1 hr = 0.1, 1 day = 2.39, 1 week = 16.75, 1 month = 71.7, 1 year = 872.35.
\nFrom this I created year feature, week feature, day of week feature, hour feature. Day of week & hour feature were furthur divided into Start_Day, End_Day & Start_Hour, End_Hour respectively.
Apart from above features, I am also thinking of creating some more features mentioned below -
\n\n@Abhilash Awasthi
\n\noh, I also tried it. As mentioned by @gingerman, it is very prominent absence. \nThat's because we typically begin at the wrong starting point.
\n\nI spend lot of time, and i tried various method. But i faced false starts and mistakes.
\n\nWhat’s the best way to analysis data and feature engineering?
\n\nI am desperate for any suggestion kaggler can offer.
\n\nThanks.
", "votes": null, "replies": [] }, { "id": 138126, "author_name": "gingerman", "author_url": "", "post_date": "10/07/2016 10:02:41", "content": "I cannot call myself a seasoned Kaggler but this was my first competiton that I tried to actively compete in over a longer time. This was in part because of the unavailibility of running strong scripts directly from the platform which constantly through me back and frustrated me in other competitions. Still, it has been a challenging trip - the first weeks I have made millions of silly mistakes like wrong merge keys, putting to many features into the model at the same time, etc. But my biggest mistake has initially been thinking to much about tuning xgboost, playing around with the MCC rounding rule (most of my submissions are identical data with different roundings) and simply not realizing that it is just not that simple.
\n\nTaking a step back, thinking a little I started combining the data sets especially taking into account temporal structure and the manufacturing perspective of paths through the factory. It was this thinking plus the upcoming discussion on the magical feature which made me look harder and being more successful in implementing structurally better solutions than incremental, minimal improvements due to fine-tuning of the ML-algorithm. Essentially, the latter I take it can be done in the last week, together with ensembling, stacking etc. Before you even start thinking about it one should construct simple models with strong meta-features. And to this end, this competition offers a host of opportunities. What I found especially appealing was the fact that for once one could easily understand the underlying reality despite the anonymization of the data. See for example my attempt at plotting the shopfloor: https://www.kaggle.com/gingerman/bosch-production-line-performance/shopfloor-visualization
", "votes": null, "replies": [] }, { "id": 138197, "author_name": "niks2000", "author_url": "", "post_date": "10/07/2016 16:08:09", "content": "@gingerman. Thank you very much.
\n\nSomethimes, Many people lose time tuning ML-algorithm, like me.\nHowever, 'Looking Data' is always the most important thing.\nGet back to basics and see if there is any way I can improve my prediction.
", "votes": null, "replies": [] }, { "id": 138594, "author_name": "abriosi", "author_url": "", "post_date": "10/09/2016 17:26:28", "content": "Noob question. How do you use fourier transform and turn it into a feature? I thought about this for a long time but i cant seem to make sense of it
", "votes": null, "replies": [ { "id": 480131, "author_name": "karthikmgk", "author_url": "", "post_date": "02/27/2019 20:15:20", "content": "Hey did you find the answer? If, yes please share it.. It will be helpful for me
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "137974": "Hello.\r\nThis is what I tried. But I didn't get a good grade(~0.25). What I would have missed.\r\nAlso, I hope this could help you.\r\n\r\n-- Numeric Data--\r\n\r\n - Fourier transform,\r\n - ftrl, \r\n - logistic regression, \r\n - nearest neighbor, \r\n - SVM,\r\n - count number of NA (each station), \r\n - count Negative/positive value (each station), \r\n - difference, \r\n - min/max/avg/SD (each station)\r\n - cumulative sum\r\n....\r\n\r\n\r\n\r\n--Date Data--\r\n\r\n - working hours (max_value - min_value), \r\n - standardization,\r\n - difference\r\n...\r\n\r\n\r\n--Categorical Data--\r\n\r\n - sparse matrix \r\n\r\n\r\n, pointing out my mistakes, thank you.", "137988": "A very prominent absence from your features are error rate features. Since train and test data sets overlap, you can determine the failure rate (e.g. on a machine level) for different time intervals and use these as features.", "138000": "I am also in similar condition. Max CV score that I am able to attain is 0.252243. \r\nThe features that I tried are - \r\n \r\n--Numeric File-- \r\n1. Number of Non-NAN entries in each Id. \r\n2. Number of Non-NAN entries in each Id per stationwise & linewise. \r\n3. Sum of Numeric features per stationwise & linewise for each Id. \r\n\r\n--Date File-- \r\n\r\nAs was shown by beluga in his \"Date Exploration\" Kernel that 0.01 = 6 min. I ordered the date file in ascending order as per station S3 since it has least NaNs (This I did so that my data would be arranged in ascending time order. So that I can find any pattern, if there is any between Time & Ids. I am assuming that Bosch has anonymized the earliest date in the data as 0). Then I calculated - \r\n1. Min_Time & Max_Time for each Id. \r\n2. Time_Difference (Max_Time - Min_Time) for each Id. \r\n3. Since 0.01= 6 min. This gives 1 hr = 0.1, 1 day = 2.39, 1 week = 16.75, 1 month = 71.7, 1 year = 872.35. \r\nFrom this I created year feature, week feature, day of week feature, hour feature. Day of week & hour feature were furthur divided into Start_Day, End_Day & Start_Hour, End_Hour respectively. \r\n\r\nApart from above features, I am also thinking of creating some more features mentioned below - \r\n \r\n1. If product passes through S32 then 1 else 0. Similarly if it passes through S32 but not S34 then 1 else 0, and S32 but not S33 then 1 else 0. \r\n2. Grouping Ids on the basis of time. Since train & test set both shares same time period. So if there is a faulty product made at a particular time, then there is a fair chance that the products manufactured near the same timeline will be faulty.", "138103": "Abhilash Awasthi \r\n\r\noh, I also tried it. As mentioned by @gingerman, it is very prominent absence. \r\nThat's because we typically begin at the wrong starting point.\r\n\r\n----------\r\n\r\nI spend lot of time, and i tried various method. But i faced false starts and mistakes.\r\n\r\nWhat’s the best way to analysis data and feature engineering?\r\n\r\nI am desperate for any suggestion kaggler can offer. \r\n\r\nThanks.", "138126": "I cannot call myself a seasoned Kaggler but this was my first competiton that I tried to actively compete in over a longer time. This was in part because of the unavailibility of running strong scripts directly from the platform which constantly through me back and frustrated me in other competitions. Still, it has been a challenging trip - the first weeks I have made millions of silly mistakes like wrong merge keys, putting to many features into the model at the same time, etc. But my biggest mistake has initially been thinking to much about tuning xgboost, playing around with the MCC rounding rule (most of my submissions are identical data with different roundings) and simply not realizing that it is just not that simple.\r\n\r\nTaking a step back, thinking a little I started combining the data sets especially taking into account temporal structure and the manufacturing perspective of paths through the factory. It was this thinking plus the upcoming discussion on the magical feature which made me look harder and being more successful in implementing structurally better solutions than incremental, minimal improvements due to fine-tuning of the ML-algorithm. Essentially, the latter I take it can be done in the last week, together with ensembling, stacking etc. Before you even start thinking about it one should construct simple models with strong meta-features. And to this end, this competition offers a host of opportunities. What I found especially appealing was the fact that for once one could easily understand the underlying reality despite the anonymization of the data. See for example my attempt at plotting the shopfloor: https://www.kaggle.com/gingerman/bosch-production-line-performance/shopfloor-visualization", "138197": "gingerman. Thank you very much. \r\n\r\nSomethimes, Many people lose time tuning ML-algorithm, like me.\r\nHowever, 'Looking Data' is always the most important thing.\r\nGet back to basics and see if there is any way I can improve my prediction.", "138594": "Noob question. How do you use fourier transform and turn it into a feature? I thought about this for a long time but i cant seem to make sense of it", "480131": "Hey did you find the answer? If, yes please share it.. It will be helpful for me" }, "source": "meta" }