{ "id": 10848, "title": "Few thoughts about \"device_id\"", "url": "/competitions/avazu-ctr-prediction/discussion/10848", "author_name": "", "post_date": "2014-11-06T00:20:15.997Z", "votes": 1, "comment_count": 8, "views": 2658, "content": "

Hi All,

\n

Just a quick check about my understanding of the field "device_id" . I suppose its a unique identifier of the device from which the website and the ads were viewed.

\n

A device might be a desktop, laptop, tablet, mobile ph. etc.

\n

So if "device_id" is indicative of a device, I'm just wondering how a particular "device_id" can show up a very large number of times in the test dataset, which is just one day's history from the website unless my interpretation of "device_id" is different from what it actually represents.

\n

I would appreciate if contest organizers or other participants can share their views about "device_id".

\n

Thanks,

\n

Avi

", "messages": [ { "id": "57367", "postDate": "11/06/2014 00:20:15", "content": "

Hi All,

\n

Just a quick check about my understanding of the field "device_id" . I suppose its a unique identifier of the device from which the website and the ads were viewed.

\n

A device might be a desktop, laptop, tablet, mobile ph. etc.

\n

So if "device_id" is indicative of a device, I'm just wondering how a particular "device_id" can show up a very large number of times in the test dataset, which is just one day's history from the website unless my interpretation of "device_id" is different from what it actually represents.

\n

I would appreciate if contest organizers or other participants can share their views about "device_id".

\n

Thanks,

\n

Avi

", "rawMarkdown": "", "votes": null }, { "id": "57368", "postDate": "11/06/2014 00:33:53", "content": "

Could be a bot. (i.e., click fraud)

", "rawMarkdown": "", "votes": null }, { "id": "57377", "postDate": "11/06/2014 04:16:25", "content": "

Good theory. If that's the case, then the number of records for a particular device on a particular day might be a good feature.

", "rawMarkdown": "", "votes": null }, { "id": "57382", "postDate": "11/06/2014 05:14:55", "content": "

There's similar patterns in training set as well. I plotted pie and bar charts for each of the variable on a sample (~1M rows) data, and one single device_id makes ~70% of the samples. Considering the 2nd most frequent device_id takes only 0.05%, I suspect deeply that it's indeed a fraud.

\n

If that's true, and it's 70% of the data, I literally don't know what to do next.

\n

I am not sure if I can throw 70% of the data away.

\n

Or maybe I can train two different classifiers for it and the rest 30% separately.

\n

Edit:

\n

I am sorry. It's 80%.

\n

The value counts are sorted in descending order.

\n

\n
d41d8cd9    819422\n5e2208cc       505\n53371cec       472\nrest        179600\ndtype: int64
", "rawMarkdown": "", "votes": null }, { "id": "57387", "postDate": "11/06/2014 06:30:52", "content": "

d41d8cd9 is the hash value of NULL

", "rawMarkdown": "", "votes": null }, { "id": "57388", "postDate": "11/06/2014 06:34:16", "content": "

[quote=Steve Wang;57387]

\n

d41d8cd9 is the hash value of NULL

\n

[/quote]

\n

Thank you, that's very helpful. I mistakenly think null values should be all '-1' like discovered in C23.

", "rawMarkdown": "", "votes": null }, { "id": "58120", "postDate": "11/15/2014 18:09:11", "content": "

between the _id, _ip, and _geo values there should be enough data to mathematically lower the potential fraud% which the industry estimates is between 20-40%.  cd9 may be NULL, but suspect that is where you 20-40% fraud is.  70% seems high.

", "rawMarkdown": "", "votes": null }, { "id": "58124", "postDate": "11/16/2014 01:54:20", "content": "

[quote=iamondialup;58120]

\n

between the _id, _ip, and _geo values there should be enough data to mathematically lower the potential fraud% which the industry estimates is between 20-40%.  cd9 may be NULL, but suspect that is where you 20-40% fraud is.  70% seems high.

\n

[/quote]

\n\n

Anyway, in order to win the Kaggle competition, take into account that click-fraud implies that there is a click, so if your approach is going in the direction of somehow not taking these fraud-clicks into account, my feeling is that you will not succeed in the competition, since these 'fraud-clicks' will also be in the test set.

\n

On the other hand, if your plan is to try to be able to identify which impressions are shown to people/bots with fraudulent intent, I think that pattern would be very useful for the competition, although to my knowledge, many click-fraud-bots are pretty sophisticated and hard to trace.

", "rawMarkdown": "", "votes": null }, { "id": "58126", "postDate": "11/16/2014 03:22:11", "content": "

[quote=F_Constantino;58124]

\n

Anyway, in order to win the Kaggle competition, take into account that click-fraud implies that there is a click, so if your approach is going in the direction of somehow not taking these fraud-clicks into account, my feeling is that you will not succeed in the competition, since these 'fraud-clicks' will also be in the test set.

\n

[/quote]

\n

That's a great point. In the Acquire Valued Shoppers Challenge, figuring out which "shoppers" were actually just the default store card swipes (i.e., for the shoppers without a loyalty card) placed me 34 out of 952. 

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 57368, "author_name": "inversion", "author_url": "", "post_date": "11/06/2014 00:33:53", "content": "

Could be a bot. (i.e., click fraud)

", "votes": null, "replies": [] }, { "id": 57377, "author_name": "alkhwarizmi", "author_url": "", "post_date": "11/06/2014 04:16:25", "content": "

Good theory. If that's the case, then the number of records for a particular device on a particular day might be a good feature.

", "votes": null, "replies": [] }, { "id": 57382, "author_name": "", "author_url": "", "post_date": "11/06/2014 05:14:55", "content": "

There's similar patterns in training set as well. I plotted pie and bar charts for each of the variable on a sample (~1M rows) data, and one single device_id makes ~70% of the samples. Considering the 2nd most frequent device_id takes only 0.05%, I suspect deeply that it's indeed a fraud.

\n

If that's true, and it's 70% of the data, I literally don't know what to do next.

\n

I am not sure if I can throw 70% of the data away.

\n

Or maybe I can train two different classifiers for it and the rest 30% separately.

\n

Edit:

\n

I am sorry. It's 80%.

\n

The value counts are sorted in descending order.

\n

\n
d41d8cd9    819422\n5e2208cc       505\n53371cec       472\nrest        179600\ndtype: int64
", "votes": null, "replies": [] }, { "id": 57387, "author_name": "stevewang0", "author_url": "", "post_date": "11/06/2014 06:30:52", "content": "

d41d8cd9 is the hash value of NULL

", "votes": null, "replies": [] }, { "id": 57388, "author_name": "", "author_url": "", "post_date": "11/06/2014 06:34:16", "content": "

[quote=Steve Wang;57387]

\n

d41d8cd9 is the hash value of NULL

\n

[/quote]

\n

Thank you, that's very helpful. I mistakenly think null values should be all '-1' like discovered in C23.

", "votes": null, "replies": [] }, { "id": 58120, "author_name": "iamondialup", "author_url": "", "post_date": "11/15/2014 18:09:11", "content": "

between the _id, _ip, and _geo values there should be enough data to mathematically lower the potential fraud% which the industry estimates is between 20-40%.  cd9 may be NULL, but suspect that is where you 20-40% fraud is.  70% seems high.

", "votes": null, "replies": [] }, { "id": 58124, "author_name": "fernandoconstantino", "author_url": "", "post_date": "11/16/2014 01:54:20", "content": "

[quote=iamondialup;58120]

\n

between the _id, _ip, and _geo values there should be enough data to mathematically lower the potential fraud% which the industry estimates is between 20-40%.  cd9 may be NULL, but suspect that is where you 20-40% fraud is.  70% seems high.

\n

[/quote]

\n\n

Anyway, in order to win the Kaggle competition, take into account that click-fraud implies that there is a click, so if your approach is going in the direction of somehow not taking these fraud-clicks into account, my feeling is that you will not succeed in the competition, since these 'fraud-clicks' will also be in the test set.

\n

On the other hand, if your plan is to try to be able to identify which impressions are shown to people/bots with fraudulent intent, I think that pattern would be very useful for the competition, although to my knowledge, many click-fraud-bots are pretty sophisticated and hard to trace.

", "votes": null, "replies": [] }, { "id": 58126, "author_name": "inversion", "author_url": "", "post_date": "11/16/2014 03:22:11", "content": "

[quote=F_Constantino;58124]

\n

Anyway, in order to win the Kaggle competition, take into account that click-fraud implies that there is a click, so if your approach is going in the direction of somehow not taking these fraud-clicks into account, my feeling is that you will not succeed in the competition, since these 'fraud-clicks' will also be in the test set.

\n

[/quote]

\n

That's a great point. In the Acquire Valued Shoppers Challenge, figuring out which "shoppers" were actually just the default store card swipes (i.e., for the shoppers without a loyalty card) placed me 34 out of 952. 

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "57367": "", "57368": "", "57377": "", "57382": "", "57387": "", "57388": "", "58120": "", "58124": "", "58126": "" }, "source": "meta" }