{ "id": 10829, "title": "50,000 clicks?!?", "url": "/competitions/avazu-ctr-prediction/discussion/10829", "author_name": "", "post_date": "2014-11-03T07:33:53.363Z", "votes": null, "comment_count": 9, "views": 2625, "content": "

I was doing a little exploratory analysis on the dataset, and noticed that a number IP addresses have >10000 clicks (one with over 50000). Is this common in this sort of data?

\n

    device_ip    sum Clicks
 1:  04245874 13369
 2:  7e4e1dd3 13361
 3:  395f7c7e 13746
 4:  82a24eb2 30382
 5:  c1b8122b 50088
 6:  2b34f027 13532
 7:  51270002 13197
 8:  13402a51 13280
 9:  ad072495 13840
10:  0a60ae81 13571
11:  c7438153 13513
12:  57a7a546 14123
13:  81a912f8 17760
14:  024d7e1e 29139
15:  e8676f9d 13486
16:  ea70f5ed 13264
17:  374dfb81 13349
18:  c4df515c 12991
19:  4fb19f5f 11280

", "messages": [ { "id": "57232", "postDate": "11/03/2014 07:33:53", "content": "

I was doing a little exploratory analysis on the dataset, and noticed that a number IP addresses have >10000 clicks (one with over 50000). Is this common in this sort of data?

\n

    device_ip    sum Clicks
 1:  04245874 13369
 2:  7e4e1dd3 13361
 3:  395f7c7e 13746
 4:  82a24eb2 30382
 5:  c1b8122b 50088
 6:  2b34f027 13532
 7:  51270002 13197
 8:  13402a51 13280
 9:  ad072495 13840
10:  0a60ae81 13571
11:  c7438153 13513
12:  57a7a546 14123
13:  81a912f8 17760
14:  024d7e1e 29139
15:  e8676f9d 13486
16:  ea70f5ed 13264
17:  374dfb81 13349
18:  c4df515c 12991
19:  4fb19f5f 11280

", "rawMarkdown": "", "votes": null }, { "id": "57242", "postDate": "11/03/2014 13:51:28", "content": "

Are you using the up-to-date data? It has been updated twice.

", "rawMarkdown": "", "votes": null }, { "id": "57243", "postDate": "11/03/2014 14:00:25", "content": "

Well, I downloaded it yesterday... so I hope it's up to date :)

\n

Does your data say differently?

", "rawMarkdown": "", "votes": null }, { "id": "57244", "postDate": "11/03/2014 14:04:22", "content": "

No, I didn't do that kind of check at all. I'm just hoping what you found is not an error of data. I'm afraid to do it all over again. I believe admin will give you a good answer. : )

", "rawMarkdown": "", "votes": null }, { "id": "57248", "postDate": "11/03/2014 14:58:45", "content": "

There is also some weird clusters in the dataset, for example:

\n

device_ip, #clicks, N, ratio
395f7c7e 13746 40051 0.3432124042
0a60ae81 13571 39589 0.3427972417
2b34f027 13532 39396 0.3434866484
374dfb81 13349 39303 0.3396432842
c7438153 13513 38899 0.3473868223
51270002 13197 38663 0.3413340920
ea70f5ed 13264 38542 0.3441440506
7e4e1dd3 13361 38519 0.3468677795
04245874 13369 38517 0.3470934912
e8676f9d 13486 38449 0.3507503446
13402a51 13280 38342 0.3463564759
c4df515c 12991 37998 0.3418864151

\n

Could they be click bots or something?

", "rawMarkdown": "", "votes": null }, { "id": "57249", "postDate": "11/03/2014 15:01:44", "content": "

My two cents:

\n

1. data has been subsampled (as mentioned in the Q&A, I seem to recall)

\n

2. for anonymization purposes, identifying features were hashed

\n

I guess the combo of those two facts might explain some data artifacts that are unlikely in the real world.

", "rawMarkdown": "", "votes": null }, { "id": "57250", "postDate": "11/03/2014 15:03:29", "content": "

I think that is bot.

\n

some ip has high click rate.

\n

-----
{:ip=>"c1b8122b", :access_count=>155236, :click_rate=>0.3226571156175114}
{:ip=>"4fb19f5f", :access_count=>130844, :click_rate=>0.08620953196172541}
{:ip=>"036e68bc", :access_count=>106923, :click_rate=>0.05659212704469572}
{:ip=>"c4941be6", :access_count=>106775, :click_rate=>0.05795364083352845}
{:ip=>"2d16bf20", :access_count=>106687, :click_rate=>0.0572234667766457}
{:ip=>"536b37ff", :access_count=>105881, :click_rate=>0.057101840745742864}
{:ip=>"c2863a7a", :access_count=>105133, :click_rate=>0.060941854603216876}
{:ip=>"d5e862d3", :access_count=>104425, :click_rate=>0.06053148192482643}
{:ip=>"ce87aefd", :access_count=>104171, :click_rate=>0.0603910877307504}
{:ip=>"1b1a8268", :access_count=>102762, :click_rate=>0.06217278760631362}
{:ip=>"024d7e1e", :access_count=>94706, :click_rate=>0.30767849977826117}
{:ip=>"81a912f8", :access_count=>65157, :click_rate=>0.272572402044293}
{:ip=>"ad072495", :access_count=>57965, :click_rate=>0.23876477184507894}
{:ip=>"57a7a546", :access_count=>57437, :click_rate=>0.24588679770879399}
{:ip=>"82a24eb2", :access_count=>41712, :click_rate=>0.7283755274261603}

", "rawMarkdown": "", "votes": null }, { "id": "57252", "postDate": "11/03/2014 15:49:58", "content": "

[quote=rcarson;57244]

\n

No, I didn't do that kind of check at all. I'm just hoping what you found is not an error of data. I'm afraid to do it all over again. I believe admin will give you a good answer. : )

\n

[/quote]

\n

there is no error, just keep going

", "rawMarkdown": "", "votes": null }, { "id": "57306", "postDate": "11/04/2014 17:29:34", "content": "

Right, there is no error

\n

devide_ ip  count_of_id sum_of_click ratio

\n

c1b8122b 155236 50088 32.265711562
4fb19f5f 130844 11280 8.6209531962
036e68bc 106923 6051 5.6592127045
c4941be6 106775 6188 5.7953640834
2d16bf20 106687 6105 5.7223466777
536b37ff 105881 6046 5.7101840746
c2863a7a 105133 6407 6.0941854603
d5e862d3 104425 6321 6.0531481925
ce87aefd 104171 6291 6.0391087731
1b1a8268 102762 6389 6.2172787606
024d7e1e 94706 29139 30.767849978
81a912f8 65157 17760 27.257240204
ad072495 57965 13840 23.876477185
57a7a546 57437 14123 24.588679771
82a24eb2 41712 30382 72.837552743
395f7c7e 40051 13746 34.321240418
0a60ae81 39589 13571 34.279724166
2b34f027 39396 13532 34.348664839
374dfb81 39303 13349 33.964328423

", "rawMarkdown": "", "votes": null }, { "id": "57559", "postDate": "11/09/2014 16:58:42", "content": "

I applied row-based filtering given the list by nagadomi, assuming these IPs indeed belong to spammers or click bots and are basically noise. I reject train set samples which contain an IP from this list. For dealing with test set samples there are two approaches as we cannot throw out rows: 

\n

1) predict as if it were random noise or

\n

2) replace the prediction with the mean CTR for that IP.

\n

This approach covers 1.6M train samples (3%) and 260K test samples (5%).

\n

Option 1 would have my preference but that degrades overall leaderboard score. Some overfitting is necessary to win based on leaderboard score.
Option 2 provided a slightly better but most likely insignificant improvement compared to not filtering at all. I expect this is because filtering essentially moves overfitting from a classifier to a pre and postprocessing step.

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 57242, "author_name": "jiweiliu", "author_url": "", "post_date": "11/03/2014 13:51:28", "content": "

Are you using the up-to-date data? It has been updated twice.

", "votes": null, "replies": [] }, { "id": 57243, "author_name": "maineiac", "author_url": "", "post_date": "11/03/2014 14:00:25", "content": "

Well, I downloaded it yesterday... so I hope it's up to date :)

\n

Does your data say differently?

", "votes": null, "replies": [] }, { "id": 57244, "author_name": "jiweiliu", "author_url": "", "post_date": "11/03/2014 14:04:22", "content": "

No, I didn't do that kind of check at all. I'm just hoping what you found is not an error of data. I'm afraid to do it all over again. I believe admin will give you a good answer. : )

", "votes": null, "replies": [] }, { "id": 57248, "author_name": "herrahuu", "author_url": "", "post_date": "11/03/2014 14:58:45", "content": "

There is also some weird clusters in the dataset, for example:

\n

device_ip, #clicks, N, ratio
395f7c7e 13746 40051 0.3432124042
0a60ae81 13571 39589 0.3427972417
2b34f027 13532 39396 0.3434866484
374dfb81 13349 39303 0.3396432842
c7438153 13513 38899 0.3473868223
51270002 13197 38663 0.3413340920
ea70f5ed 13264 38542 0.3441440506
7e4e1dd3 13361 38519 0.3468677795
04245874 13369 38517 0.3470934912
e8676f9d 13486 38449 0.3507503446
13402a51 13280 38342 0.3463564759
c4df515c 12991 37998 0.3418864151

\n

Could they be click bots or something?

", "votes": null, "replies": [] }, { "id": 57249, "author_name": "konradb", "author_url": "", "post_date": "11/03/2014 15:01:44", "content": "

My two cents:

\n

1. data has been subsampled (as mentioned in the Q&A, I seem to recall)

\n

2. for anonymization purposes, identifying features were hashed

\n

I guess the combo of those two facts might explain some data artifacts that are unlikely in the real world.

", "votes": null, "replies": [] }, { "id": 57250, "author_name": "nagadomi", "author_url": "", "post_date": "11/03/2014 15:03:29", "content": "

I think that is bot.

\n

some ip has high click rate.

\n

-----
{:ip=>"c1b8122b", :access_count=>155236, :click_rate=>0.3226571156175114}
{:ip=>"4fb19f5f", :access_count=>130844, :click_rate=>0.08620953196172541}
{:ip=>"036e68bc", :access_count=>106923, :click_rate=>0.05659212704469572}
{:ip=>"c4941be6", :access_count=>106775, :click_rate=>0.05795364083352845}
{:ip=>"2d16bf20", :access_count=>106687, :click_rate=>0.0572234667766457}
{:ip=>"536b37ff", :access_count=>105881, :click_rate=>0.057101840745742864}
{:ip=>"c2863a7a", :access_count=>105133, :click_rate=>0.060941854603216876}
{:ip=>"d5e862d3", :access_count=>104425, :click_rate=>0.06053148192482643}
{:ip=>"ce87aefd", :access_count=>104171, :click_rate=>0.0603910877307504}
{:ip=>"1b1a8268", :access_count=>102762, :click_rate=>0.06217278760631362}
{:ip=>"024d7e1e", :access_count=>94706, :click_rate=>0.30767849977826117}
{:ip=>"81a912f8", :access_count=>65157, :click_rate=>0.272572402044293}
{:ip=>"ad072495", :access_count=>57965, :click_rate=>0.23876477184507894}
{:ip=>"57a7a546", :access_count=>57437, :click_rate=>0.24588679770879399}
{:ip=>"82a24eb2", :access_count=>41712, :click_rate=>0.7283755274261603}

", "votes": null, "replies": [] }, { "id": 57252, "author_name": "stevewang0", "author_url": "", "post_date": "11/03/2014 15:49:58", "content": "

[quote=rcarson;57244]

\n

No, I didn't do that kind of check at all. I'm just hoping what you found is not an error of data. I'm afraid to do it all over again. I believe admin will give you a good answer. : )

\n

[/quote]

\n

there is no error, just keep going

", "votes": null, "replies": [] }, { "id": 57306, "author_name": "mike64", "author_url": "", "post_date": "11/04/2014 17:29:34", "content": "

Right, there is no error

\n

devide_ ip  count_of_id sum_of_click ratio

\n

c1b8122b 155236 50088 32.265711562
4fb19f5f 130844 11280 8.6209531962
036e68bc 106923 6051 5.6592127045
c4941be6 106775 6188 5.7953640834
2d16bf20 106687 6105 5.7223466777
536b37ff 105881 6046 5.7101840746
c2863a7a 105133 6407 6.0941854603
d5e862d3 104425 6321 6.0531481925
ce87aefd 104171 6291 6.0391087731
1b1a8268 102762 6389 6.2172787606
024d7e1e 94706 29139 30.767849978
81a912f8 65157 17760 27.257240204
ad072495 57965 13840 23.876477185
57a7a546 57437 14123 24.588679771
82a24eb2 41712 30382 72.837552743
395f7c7e 40051 13746 34.321240418
0a60ae81 39589 13571 34.279724166
2b34f027 39396 13532 34.348664839
374dfb81 39303 13349 33.964328423

", "votes": null, "replies": [] }, { "id": 57559, "author_name": "tomderuijter", "author_url": "", "post_date": "11/09/2014 16:58:42", "content": "

I applied row-based filtering given the list by nagadomi, assuming these IPs indeed belong to spammers or click bots and are basically noise. I reject train set samples which contain an IP from this list. For dealing with test set samples there are two approaches as we cannot throw out rows: 

\n

1) predict as if it were random noise or

\n

2) replace the prediction with the mean CTR for that IP.

\n

This approach covers 1.6M train samples (3%) and 260K test samples (5%).

\n

Option 1 would have my preference but that degrades overall leaderboard score. Some overfitting is necessary to win based on leaderboard score.
Option 2 provided a slightly better but most likely insignificant improvement compared to not filtering at all. I expect this is because filtering essentially moves overfitting from a classifier to a pre and postprocessing step.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "57232": "", "57242": "", "57243": "", "57244": "", "57248": "", "57249": "", "57250": "", "57252": "", "57306": "", "57559": "" }, "source": "meta" }