{ "id": 22663, "title": "Facebook Results Dataset", "url": "/competitions/facebook-v-predicting-check-ins/discussion/22663", "author_name": "", "post_date": "2016-08-03T08:36:54.133Z", "votes": 6, "comment_count": 5, "views": 1239, "content": "

I've just released a dataset containing the results of this Facebook competition. It includes:

\n\n\n\n

A couple examples of what you can do with this dataset:

\n\n

Ensemble the top team's final submissions

\n\n

\"ensemble\"

\n\n

Calculate the agreement between top submissions

\n\n

\"agreement\"

", "messages": [ { "id": "129992", "postDate": "08/03/2016 08:36:54", "content": "

I've just released a dataset containing the results of this Facebook competition. It includes:

\n\n\n\n

A couple examples of what you can do with this dataset:

\n\n

Ensemble the top team's final submissions

\n\n

\"ensemble\"

\n\n

Calculate the agreement between top submissions

\n\n

\"agreement\"

", "rawMarkdown": "I've just released [a dataset containing the results of this Facebook competition](https://www.kaggle.com/facebook/facebook-v-results). It includes:\r\n\r\n - the solution file\r\n - the final submission for each of the top 20 teams\r\n - final submissions for a small number of teams farther down the leaderboard\r\n - public/private scores for all submissions made during the competition\r\n\r\nA couple examples of what you can do with this dataset:\r\n\r\n[Ensemble the top team's final submissions](https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/ensembling-the-top-teams)\r\n\r\n[![ensemble](https://www.kaggle.io/svf/320079/dc302b86d09f5654a8ef5cf8d2157b1a/ensemble.png)](https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/ensembling-the-top-teams)\r\n\r\n[Calculate the agreement between top submissions](https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/top-submission-agreement)\r\n\r\n[![agreement](https://www.kaggle.io/svf/320080/4c9b6e6f4c979d0000d2fbe1440ecc81/ranked_agreement.png)](https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/top-submission-agreement)", "votes": null }, { "id": "130172", "postDate": "08/04/2016 14:09:35", "content": "

Hi Ben,

\n\n

What tools do you recommend for extracting the original csv's from the hd5 file? I tried using hdfview . I found something strange.

\n\n

The Rank1 values did not match the values that Tom made publicly available. When I submitted my own conversion from Rank1, the results were low (around 0.23061 on public LB). I scored my converted csv with my export from Solution.txt and got the score I expected.

\n\n

I'm new to hd5 so perhaps, I am doing something obviously wrong in using hdfview to export a text file.

\n\n

Thanks,

\n\n

-Larry

\n\n
\n\n

Edit 2: Ben, I got the same result using h5py similar to the code you provided to do the ensemble.

\n\n

Here is the code that I used. When I submitted the resultant csv, I again got the same lower than expected LB score.

\n\n
import h5py\nimport numpy as np\nimport pandas as pd\n\n\nwith h5py.File("top_submissions.h5","r") as f:\n    submission = f['Rank1'][:]\n    df_submission = pd.DataFrame(submission)\n    df_submission.columns = ['predict1','predict2','predict3']\n    df_submission['place_id'] = df_submission.predict1.map(str)+" "+df_submission.predict2.map(str)+" "+df_submission.predict3.map(str)\n    df_submission = df_submission[['place_id']]\n    df_submission.reset_index(drop=True,inplace=True)\n    df_submission.index.rename('row_id',inplace=True)\n    df_submission.to_csv("Rank1_test.csv")\n
", "rawMarkdown": "Hi Ben,\r\n\r\nWhat tools do you recommend for extracting the original csv's from the hd5 file? I tried using [hdfview][1] . I found something strange.\r\n\r\nThe Rank1 values did not match the values that Tom [made publicly available][2]. When I submitted my own conversion from Rank1, the results were low (around 0.23061 on public LB). I scored my converted csv with my export from Solution.txt and got the score I expected.\r\n\r\nI'm new to hd5 so perhaps, I am doing something obviously wrong in using hdfview to export a text file.\r\n\r\nThanks,\r\n\r\n-Larry\r\n\r\n------\r\n\r\nEdit 2: Ben, I got the same result using h5py similar to the code you [provided][3] to do the ensemble. \r\n\r\nHere is the code that I used. When I submitted the resultant csv, I again got the same lower than expected LB score.\r\n\r\n import h5py\r\n import numpy as np\r\n import pandas as pd\r\n \r\n \r\n with h5py.File(\"top_submissions.h5\",\"r\") as f:\r\n submission = f['Rank1'][:]\r\n df_submission = pd.DataFrame(submission)\r\n df_submission.columns = ['predict1','predict2','predict3']\r\n df_submission['place_id'] = df_submission.predict1.map(str)+\" \"+df_submission.predict2.map(str)+\" \"+df_submission.predict3.map(str)\r\n df_submission = df_submission[['place_id']]\r\n df_submission.reset_index(drop=True,inplace=True)\r\n df_submission.index.rename('row_id',inplace=True)\r\n df_submission.to_csv(\"Rank1_test.csv\")\r\n\r\n\r\n\r\n\r\n\r\n [1]: https://apps.ubuntu.com/cat/applications/precise/hdfview/\r\n [2]: https://www.kaggle.com/c/facebook-v-predicting-check-ins/forums/t/22078/solution-sharing/128239#post128239\r\n [3]: https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/ensembling-the-top-teams", "votes": null }, { "id": "130311", "postDate": "08/05/2016 14:02:11", "content": "

Thanks a lot.

\n\n

I will spend time to study.

", "rawMarkdown": "Thanks a lot.\r\n\r\nI will spend time to study.", "votes": null }, { "id": "130362", "postDate": "08/06/2016 06:49:02", "content": "

I think there is some problem in solution file as well. Shouldn't it give 1.00 on lb? When I do this:

\n\n
with h5py.File("top_submissions.h5","r") as f:\n    s= f["Solution"][:]\n    df = pd.DataFrame(s)\n    df.columns = ["row_id","place_id","priv/publ"]\n    df[["row_id","place_id"]].to_csv("sol.csv",index=False)\n
\n\n

and then submit sol.csv, I get only 0.36982. Am I doing something wrong?

", "rawMarkdown": "I think there is some problem in solution file as well. Shouldn't it give 1.00 on lb? When I do this:\r\n\r\n with h5py.File(\"top_submissions.h5\",\"r\") as f:\r\n s= f[\"Solution\"][:]\r\n df = pd.DataFrame(s)\r\n df.columns = [\"row_id\",\"place_id\",\"priv/publ\"]\r\n df[[\"row_id\",\"place_id\"]].to_csv(\"sol.csv\",index=False)\r\n\r\nand then submit sol.csv, I get only 0.36982. Am I doing something wrong?", "votes": null }, { "id": "130775", "postDate": "08/10/2016 08:44:10", "content": "

Really cool post-competition analysis and data set.

\n\n

I suggest notifying all competition participants over email, as otherwise +95% of them will probably never spot this.

\n\n

I guess some people might find the release of submission metadata slightly violating their privacy, as it enables competitors to data mine their strategies for parameter tuning and ensembling, as well as the timing for entering each stage.

", "rawMarkdown": "Really cool post-competition analysis and data set. \r\n\r\nI suggest notifying all competition participants over email, as otherwise +95% of them will probably never spot this. \r\n\r\nI guess some people might find the release of submission metadata slightly violating their privacy, as it enables competitors to data mine their strategies for parameter tuning and ensembling, as well as the timing for entering each stage.", "votes": null }, { "id": "135596", "postDate": "09/15/2016 06:01:49", "content": "

There are so many people commenting here. Anyone please tell me how to get the proper solution dataset as the method I mentioned earlier has something wrong.

", "rawMarkdown": "There are so many people commenting here. Anyone please tell me how to get the proper solution dataset as the method I mentioned earlier has something wrong.", "votes": null } ], "comments": [ { "id": 130172, "author_name": "larryfreeman", "author_url": "", "post_date": "08/04/2016 14:09:35", "content": "

Hi Ben,

\n\n

What tools do you recommend for extracting the original csv's from the hd5 file? I tried using hdfview . I found something strange.

\n\n

The Rank1 values did not match the values that Tom made publicly available. When I submitted my own conversion from Rank1, the results were low (around 0.23061 on public LB). I scored my converted csv with my export from Solution.txt and got the score I expected.

\n\n

I'm new to hd5 so perhaps, I am doing something obviously wrong in using hdfview to export a text file.

\n\n

Thanks,

\n\n

-Larry

\n\n
\n\n

Edit 2: Ben, I got the same result using h5py similar to the code you provided to do the ensemble.

\n\n

Here is the code that I used. When I submitted the resultant csv, I again got the same lower than expected LB score.

\n\n
import h5py\nimport numpy as np\nimport pandas as pd\n\n\nwith h5py.File("top_submissions.h5","r") as f:\n    submission = f['Rank1'][:]\n    df_submission = pd.DataFrame(submission)\n    df_submission.columns = ['predict1','predict2','predict3']\n    df_submission['place_id'] = df_submission.predict1.map(str)+" "+df_submission.predict2.map(str)+" "+df_submission.predict3.map(str)\n    df_submission = df_submission[['place_id']]\n    df_submission.reset_index(drop=True,inplace=True)\n    df_submission.index.rename('row_id',inplace=True)\n    df_submission.to_csv("Rank1_test.csv")\n
", "votes": null, "replies": [] }, { "id": 130311, "author_name": "wowfattie", "author_url": "", "post_date": "08/05/2016 14:02:11", "content": "

Thanks a lot.

\n\n

I will spend time to study.

", "votes": null, "replies": [] }, { "id": 130362, "author_name": "chopra", "author_url": "", "post_date": "08/06/2016 06:49:02", "content": "

I think there is some problem in solution file as well. Shouldn't it give 1.00 on lb? When I do this:

\n\n
with h5py.File("top_submissions.h5","r") as f:\n    s= f["Solution"][:]\n    df = pd.DataFrame(s)\n    df.columns = ["row_id","place_id","priv/publ"]\n    df[["row_id","place_id"]].to_csv("sol.csv",index=False)\n
\n\n

and then submit sol.csv, I get only 0.36982. Am I doing something wrong?

", "votes": null, "replies": [] }, { "id": 130775, "author_name": "frederik", "author_url": "", "post_date": "08/10/2016 08:44:10", "content": "

Really cool post-competition analysis and data set.

\n\n

I suggest notifying all competition participants over email, as otherwise +95% of them will probably never spot this.

\n\n

I guess some people might find the release of submission metadata slightly violating their privacy, as it enables competitors to data mine their strategies for parameter tuning and ensembling, as well as the timing for entering each stage.

", "votes": null, "replies": [] }, { "id": 135596, "author_name": "chopra", "author_url": "", "post_date": "09/15/2016 06:01:49", "content": "

There are so many people commenting here. Anyone please tell me how to get the proper solution dataset as the method I mentioned earlier has something wrong.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "129992": "I've just released [a dataset containing the results of this Facebook competition](https://www.kaggle.com/facebook/facebook-v-results). It includes:\r\n\r\n - the solution file\r\n - the final submission for each of the top 20 teams\r\n - final submissions for a small number of teams farther down the leaderboard\r\n - public/private scores for all submissions made during the competition\r\n\r\nA couple examples of what you can do with this dataset:\r\n\r\n[Ensemble the top team's final submissions](https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/ensembling-the-top-teams)\r\n\r\n[![ensemble](https://www.kaggle.io/svf/320079/dc302b86d09f5654a8ef5cf8d2157b1a/ensemble.png)](https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/ensembling-the-top-teams)\r\n\r\n[Calculate the agreement between top submissions](https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/top-submission-agreement)\r\n\r\n[![agreement](https://www.kaggle.io/svf/320080/4c9b6e6f4c979d0000d2fbe1440ecc81/ranked_agreement.png)](https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/top-submission-agreement)", "130172": "Hi Ben,\r\n\r\nWhat tools do you recommend for extracting the original csv's from the hd5 file? I tried using [hdfview][1] . I found something strange.\r\n\r\nThe Rank1 values did not match the values that Tom [made publicly available][2]. When I submitted my own conversion from Rank1, the results were low (around 0.23061 on public LB). I scored my converted csv with my export from Solution.txt and got the score I expected.\r\n\r\nI'm new to hd5 so perhaps, I am doing something obviously wrong in using hdfview to export a text file.\r\n\r\nThanks,\r\n\r\n-Larry\r\n\r\n------\r\n\r\nEdit 2: Ben, I got the same result using h5py similar to the code you [provided][3] to do the ensemble. \r\n\r\nHere is the code that I used. When I submitted the resultant csv, I again got the same lower than expected LB score.\r\n\r\n import h5py\r\n import numpy as np\r\n import pandas as pd\r\n \r\n \r\n with h5py.File(\"top_submissions.h5\",\"r\") as f:\r\n submission = f['Rank1'][:]\r\n df_submission = pd.DataFrame(submission)\r\n df_submission.columns = ['predict1','predict2','predict3']\r\n df_submission['place_id'] = df_submission.predict1.map(str)+\" \"+df_submission.predict2.map(str)+\" \"+df_submission.predict3.map(str)\r\n df_submission = df_submission[['place_id']]\r\n df_submission.reset_index(drop=True,inplace=True)\r\n df_submission.index.rename('row_id',inplace=True)\r\n df_submission.to_csv(\"Rank1_test.csv\")\r\n\r\n\r\n\r\n\r\n\r\n [1]: https://apps.ubuntu.com/cat/applications/precise/hdfview/\r\n [2]: https://www.kaggle.com/c/facebook-v-predicting-check-ins/forums/t/22078/solution-sharing/128239#post128239\r\n [3]: https://www.kaggle.com/benhamner/d/facebook/facebook-v-results/ensembling-the-top-teams", "130311": "Thanks a lot.\r\n\r\nI will spend time to study.", "130362": "I think there is some problem in solution file as well. Shouldn't it give 1.00 on lb? When I do this:\r\n\r\n with h5py.File(\"top_submissions.h5\",\"r\") as f:\r\n s= f[\"Solution\"][:]\r\n df = pd.DataFrame(s)\r\n df.columns = [\"row_id\",\"place_id\",\"priv/publ\"]\r\n df[[\"row_id\",\"place_id\"]].to_csv(\"sol.csv\",index=False)\r\n\r\nand then submit sol.csv, I get only 0.36982. Am I doing something wrong?", "130775": "Really cool post-competition analysis and data set. \r\n\r\nI suggest notifying all competition participants over email, as otherwise +95% of them will probably never spot this. \r\n\r\nI guess some people might find the release of submission metadata slightly violating their privacy, as it enables competitors to data mine their strategies for parameter tuning and ensembling, as well as the timing for entering each stage.", "135596": "There are so many people commenting here. Anyone please tell me how to get the proper solution dataset as the method I mentioned earlier has something wrong." }, "source": "meta" }