As we are on the end of the competition , would you mind sharing single best features used in the model till now?\nFor my case -\nSingle LGB - 98.09 with only Count & Next Click features.\nSingle NN - 97.94 with only Count & Next Click features.\ntried with Var , Skew , Mean etc.
\n\nNext Click is real hero here. Any such features found by anyone.
", "messages": [ { "id": 323986, "postDate": "2018-05-06T20:57:00.343Z", "content": "I suggest you wait for only one day. We will tell you the importance of features after the competition ends. As Fengari said, it is not fair for all the competitors if some people discuss strong features here at this moment.
", "rawMarkdown": "I suggest you wait for only one day. We will tell you the importance of features after the competition ends. As Fengari said, it is not fair for all the competitors if some people discuss strong features here at this moment.", "votes": 6, "replies": [ { "id": 323990, "postDate": "2018-05-06T21:35:30.780Z", "rawMarkdown": "", "votes": -12, "isDeleted": true }, { "id": 324098, "postDate": "2018-05-07T06:13:15.587Z", "content": "@Dirk, do you have any evidence of this?
", "rawMarkdown": "@Dirk, do you have any evidence of this? " } ] }, { "id": 323903, "postDate": "2018-05-06T16:10:18.270Z", "content": "It's not fair for the top competitors to share their golden features, anyway just two days, I am also very interesting and what to learn from this guys.
", "rawMarkdown": "It's not fair for the top competitors to share their golden features, anyway just two days, I am also very interesting and what to learn from this guys.", "votes": 4, "replies": [ { "id": 323905, "postDate": "2018-05-06T16:13:21.323Z", "content": "O yah , I agree that we are still 2 days to go. Not worth sharing at this point.
", "rawMarkdown": "O yah , I agree that we are still 2 days to go. Not worth sharing at this point." }, { "id": 323992, "postDate": "2018-05-06T21:40:11.273Z", "rawMarkdown": "", "votes": -7, "isDeleted": true }, { "id": 324017, "postDate": "2018-05-06T23:47:15.147Z", "content": "@midi303, IMHO your accusation is not fair - it is labelling people cheating without ANY evidence. If you have evidence, please put it forward.
\n\nFor what it is worth, here is a competition where the winner blew everyone away with one submission:\nhttps://www.kaggle.com/c/expedia-hotel-recommendations/leaderboard
", "rawMarkdown": "@midi303, IMHO your accusation is not fair - it is labelling people cheating without ANY evidence. If you have evidence, please put it forward.\n\nFor what it is worth, here is a competition where the winner blew everyone away with one submission:\nhttps://www.kaggle.com/c/expedia-hotel-recommendations/leaderboard", "votes": 3 } ] }, { "id": 324070, "postDate": "2018-05-07T04:49:06.727Z", "content": "I feel hard to define a good 'single' feature... I created some feature that shows great gain in lightgbm (rank top 1, without target encoding) but did not improved my LB score at all.
", "rawMarkdown": "I feel hard to define a good 'single' feature... I created some feature that shows great gain in lightgbm (rank top 1, without target encoding) but did not improved my LB score at all.", "votes": 2 }, { "id": 323902, "postDate": "2018-05-06T16:05:59.870Z", "content": "As we are on the end of the competition , would you mind sharing single best features used in the model till now?\nFor my case -\nSingle LGB - 98.09 with only Count & Next Click features.\nSingle NN - 97.94 with only Count & Next Click features.\ntried with Var , Skew , Mean etc.
\n\nNext Click is real hero here. Any such features found by anyone.
", "rawMarkdown": "As we are on the end of the competition , would you mind sharing single best features used in the model till now?\nFor my case -\nSingle LGB - 98.09 with only Count & Next Click features.\nSingle NN - 97.94 with only Count & Next Click features.\ntried with Var , Skew , Mean etc.\n\nNext Click is real hero here. Any such features found by anyone." }, { "id": 323988, "postDate": "2018-05-06T21:10:10.123Z", "content": "You might look into the output (finalfeatimportance.csv) and code of https://www.kaggle.com/alesgb/genetic-target-encoding-counts-and-lag-search-etc/ to find a great variety of automatically found features of different sorts (lags of variables by groups, target encoded counts, means, sd by groups, counts by groups in general and some others) and their ranks, that you might like to use during the last trainings. If you have the capacity to run it on the whole dataset for a while, I guess you might indeed find many interesting patterns ;)
", "rawMarkdown": "You might look into the output (finalfeatimportance.csv) and code of https://www.kaggle.com/alesgb/genetic-target-encoding-counts-and-lag-search-etc/ to find a great variety of automatically found features of different sorts (lags of variables by groups, target encoded counts, means, sd by groups, counts by groups in general and some others) and their ranks, that you might like to use during the last trainings. If you have the capacity to run it on the whole dataset for a while, I guess you might indeed find many interesting patterns ;)" }, { "id": 323958, "postDate": "2018-05-06T19:38:06.800Z", "content": "Beside the ones you mentioned: Unique Counts.
", "rawMarkdown": "Beside the ones you mentioned: Unique Counts.", "replies": [ { "id": 323959, "postDate": "2018-05-06T19:40:40.827Z", "content": "Unique Count - did it make significant improvement? Not for me at least.
", "rawMarkdown": "Unique Count - did it make significant improvement? Not for me at least." }, { "id": 324033, "postDate": "2018-05-07T01:53:38.970Z", "content": "It is in my top 10.
", "rawMarkdown": "It is in my top 10." } ] }, { "id": 323932, "postDate": "2018-05-06T17:53:22.290Z", "content": "just wondering how much improvement did you get from hyper-parameters tuning
", "rawMarkdown": "just wondering how much improvement did you get from hyper-parameters tuning", "replies": [ { "id": 323939, "postDate": "2018-05-06T18:27:58.730Z", "content": "We got about 0.0004 improvement by parameter tuning compared to our baseline solution (i.e. a solution with basic hyperparameters). But that heavily depends on your baseline solution ofcourse, and the learning rate you chose.
", "rawMarkdown": "We got about 0.0004 improvement by parameter tuning compared to our baseline solution (i.e. a solution with basic hyperparameters). But that heavily depends on your baseline solution ofcourse, and the learning rate you chose." }, { "id": 323948, "postDate": "2018-05-06T19:16:20.220Z", "content": "Reducing learning rate is helping scoring a little bit but making train-process dam slow.
", "rawMarkdown": "Reducing learning rate is helping scoring a little bit but making train-process dam slow." }, { "id": 323954, "postDate": "2018-05-06T19:33:48.297Z", "content": "0.0004 is quite a big edge Thanks for the comment
", "rawMarkdown": "0.0004 is quite a big edge Thanks for the comment" } ] }, { "id": 323977, "postDate": "2018-05-06T20:23:31.047Z", "rawMarkdown": "", "votes": -6, "isDeleted": true } ], "comments": [ { "id": 323986, "author_name": "Feiyang Pan", "author_url": "", "post_date": "2018-05-06T20:57:00.343000", "content": "I suggest you wait for only one day. We will tell you the importance of features after the competition ends. As Fengari said, it is not fair for all the competitors if some people discuss strong features here at this moment.
", "votes": 6, "replies": [ { "id": 323990, "author_name": "", "author_url": "", "post_date": "2018-05-06T21:35:30.780000", "content": "", "votes": -12, "replies": [] }, { "id": 324098, "author_name": "CPMP", "author_url": "", "post_date": "2018-05-07T06:13:15.587000", "content": "@Dirk, do you have any evidence of this?
", "votes": 0, "replies": [] } ] }, { "id": 323903, "author_name": "Fengari", "author_url": "", "post_date": "2018-05-06T16:10:18.270000", "content": "It's not fair for the top competitors to share their golden features, anyway just two days, I am also very interesting and what to learn from this guys.
", "votes": 4, "replies": [ { "id": 323905, "author_name": "SubikashPal", "author_url": "", "post_date": "2018-05-06T16:13:21.323000", "content": "O yah , I agree that we are still 2 days to go. Not worth sharing at this point.
", "votes": 0, "replies": [] }, { "id": 323992, "author_name": "", "author_url": "", "post_date": "2018-05-06T21:40:11.273000", "content": "", "votes": -7, "replies": [] }, { "id": 324017, "author_name": "Yifan Xie", "author_url": "", "post_date": "2018-05-06T23:47:15.147000", "content": "@midi303, IMHO your accusation is not fair - it is labelling people cheating without ANY evidence. If you have evidence, please put it forward.
\n\nFor what it is worth, here is a competition where the winner blew everyone away with one submission:\nhttps://www.kaggle.com/c/expedia-hotel-recommendations/leaderboard
", "votes": 3, "replies": [] } ] }, { "id": 324070, "author_name": "Cheng", "author_url": "", "post_date": "2018-05-07T04:49:06.727000", "content": "I feel hard to define a good 'single' feature... I created some feature that shows great gain in lightgbm (rank top 1, without target encoding) but did not improved my LB score at all.
", "votes": 2, "replies": [] }, { "id": 323988, "author_name": "Aleś", "author_url": "", "post_date": "2018-05-06T21:10:10.123000", "content": "You might look into the output (finalfeatimportance.csv) and code of https://www.kaggle.com/alesgb/genetic-target-encoding-counts-and-lag-search-etc/ to find a great variety of automatically found features of different sorts (lags of variables by groups, target encoded counts, means, sd by groups, counts by groups in general and some others) and their ranks, that you might like to use during the last trainings. If you have the capacity to run it on the whole dataset for a while, I guess you might indeed find many interesting patterns ;)
", "votes": 0, "replies": [] }, { "id": 323958, "author_name": "Rohit Mehra", "author_url": "", "post_date": "2018-05-06T19:38:06.800000", "content": "Beside the ones you mentioned: Unique Counts.
", "votes": 0, "replies": [ { "id": 323959, "author_name": "SubikashPal", "author_url": "", "post_date": "2018-05-06T19:40:40.827000", "content": "Unique Count - did it make significant improvement? Not for me at least.
", "votes": 0, "replies": [] }, { "id": 324033, "author_name": "Rohit Mehra", "author_url": "", "post_date": "2018-05-07T01:53:38.970000", "content": "It is in my top 10.
", "votes": 0, "replies": [] } ] }, { "id": 323932, "author_name": "zcistkidd", "author_url": "", "post_date": "2018-05-06T17:53:22.290000", "content": "just wondering how much improvement did you get from hyper-parameters tuning
", "votes": 0, "replies": [ { "id": 323939, "author_name": "Kevin", "author_url": "", "post_date": "2018-05-06T18:27:58.730000", "content": "We got about 0.0004 improvement by parameter tuning compared to our baseline solution (i.e. a solution with basic hyperparameters). But that heavily depends on your baseline solution ofcourse, and the learning rate you chose.
", "votes": 0, "replies": [] }, { "id": 323948, "author_name": "SubikashPal", "author_url": "", "post_date": "2018-05-06T19:16:20.220000", "content": "Reducing learning rate is helping scoring a little bit but making train-process dam slow.
", "votes": 0, "replies": [] }, { "id": 323954, "author_name": "zcistkidd", "author_url": "", "post_date": "2018-05-06T19:33:48.297000", "content": "0.0004 is quite a big edge Thanks for the comment
", "votes": 0, "replies": [] } ] }, { "id": 323977, "author_name": "", "author_url": "", "post_date": "2018-05-06T20:23:31.047000", "content": "", "votes": -6, "replies": [] } ], "raw_markdown_by_id": { "323986": "I suggest you wait for only one day. We will tell you the importance of features after the competition ends. As Fengari said, it is not fair for all the competitors if some people discuss strong features here at this moment.", "323903": "It's not fair for the top competitors to share their golden features, anyway just two days, I am also very interesting and what to learn from this guys.", "324070": "I feel hard to define a good 'single' feature... I created some feature that shows great gain in lightgbm (rank top 1, without target encoding) but did not improved my LB score at all.", "323902": "As we are on the end of the competition , would you mind sharing single best features used in the model till now?\nFor my case -\nSingle LGB - 98.09 with only Count & Next Click features.\nSingle NN - 97.94 with only Count & Next Click features.\ntried with Var , Skew , Mean etc.\n\nNext Click is real hero here. Any such features found by anyone.", "323988": "You might look into the output (finalfeatimportance.csv) and code of https://www.kaggle.com/alesgb/genetic-target-encoding-counts-and-lag-search-etc/ to find a great variety of automatically found features of different sorts (lags of variables by groups, target encoded counts, means, sd by groups, counts by groups in general and some others) and their ranks, that you might like to use during the last trainings. If you have the capacity to run it on the whole dataset for a while, I guess you might indeed find many interesting patterns ;)", "323958": "Beside the ones you mentioned: Unique Counts.", "323932": "just wondering how much improvement did you get from hyper-parameters tuning", "323977": "" } }