{ "id": 7752, "title": "Something that makes this problem interesting and challenging", "url": "/competitions/acquire-valued-shoppers-challenge/discussion/7752", "author_name": "", "post_date": "2014-04-18T05:19:23.660Z", "votes": 18, "comment_count": 9, "views": 3679, "content": "

Some categories on offers in the training set have a high repeat % (e.g. 2119 and 9909). Other categories have a low repeat % (e.g. 1703 and 4401).

\n

The repeat % of some categories are affected by previous purchases of the same brand (e.g. 1703 & 1726). Others are not (e.g. 5619). (A similar analysis could be done for previous purchases of same company, category, etc.)

\n

The categories on the offers in the training set are different than the categories on the offers in the test set. For example, category 7205 is a large portion of the test set, but it is not in the training set. Does 7205 have a high repeat %, or a low repeat %? Is the repeat % for 7205 affected by previous purchases of the same brand? It is difficult to answer these questions directly from the training set because there are no 7205 offers in the training set.

\n

So the problem could boil down to “Figure out how I will respond to a discount on ice cream – given your response to a discount on deodorant.”

\n

I could create a dataset from transactions that measures how often people buy 7205 and to what extent past purchases of 7205 indicate future purchases of 7205. But would that dataset capture the effect of the discount?

", "messages": [ { "id": "42289", "postDate": "04/18/2014 05:19:23", "content": "

Some categories on offers in the training set have a high repeat % (e.g. 2119 and 9909). Other categories have a low repeat % (e.g. 1703 and 4401).

\n

The repeat % of some categories are affected by previous purchases of the same brand (e.g. 1703 & 1726). Others are not (e.g. 5619). (A similar analysis could be done for previous purchases of same company, category, etc.)

\n

The categories on the offers in the training set are different than the categories on the offers in the test set. For example, category 7205 is a large portion of the test set, but it is not in the training set. Does 7205 have a high repeat %, or a low repeat %? Is the repeat % for 7205 affected by previous purchases of the same brand? It is difficult to answer these questions directly from the training set because there are no 7205 offers in the training set.

\n

So the problem could boil down to “Figure out how I will respond to a discount on ice cream – given your response to a discount on deodorant.”

\n

I could create a dataset from transactions that measures how often people buy 7205 and to what extent past purchases of 7205 indicate future purchases of 7205. But would that dataset capture the effect of the discount?

", "rawMarkdown": "", "votes": null }, { "id": "42324", "postDate": "04/18/2014 15:38:13", "content": "

[quote=BreakfastPirate;42289]

\n

The repeat % of some categories are affected by previous purchases of the same brand (e.g. 1703 & 1726). Others are not (e.g. 5619). (A similar analysis could be done for previous purchases of same company, category, etc.)

\n

[/quote]

\n\n

Could you describe how you did this analysis?

", "rawMarkdown": "", "votes": null }, { "id": "42333", "postDate": "04/18/2014 17:43:59", "content": "

For each row in the training set I:

\n", "rawMarkdown": "", "votes": null }, { "id": "44303", "postDate": "05/09/2014 14:03:30", "content": "

@BreakfastPirate - I am bit confused with a problem statement. Do we have find out the probability of person redeeming the offer OR do we have to find out probability of person redeeming the offer multiple times?

\n

Looking forward for your/anyone's reply ? 

", "rawMarkdown": "", "votes": null }, { "id": "44305", "postDate": "05/09/2014 14:18:34", "content": "

My understanding is that the customers have already redeemed the offer, and we are predicting whether they will buy the product again after the initial incented offer.  This is based on the problem description: "The most valuable customers are those who return after this initial incented purchase."

", "rawMarkdown": "", "votes": null }, { "id": "44346", "postDate": "05/10/2014 16:05:16", "content": "

Thanks for your reply. 

\n

So does that means that ALL customers in trainhistory.csv have redeemed the offer earlier atleast once and the repeater = t indicates the customer has made a repeat purchase after the first purchase. Or what i am saying is that all the customers are those who used the offer once and we are predicting their chances of making same purchase again.

\n

Also, is the initial incented offer is before the offer date in trainhistory.csv file? As there are customers whose repeater = f or repeattrip = 0 which means they have never come for a purchase after the offer date. Please help me understanding this as i am still confused :(

", "rawMarkdown": "", "votes": null }, { "id": "49151", "postDate": "06/15/2014 09:53:00", "content": "

Even my understanding is same as that of Decipher's. 

", "rawMarkdown": "", "votes": null }, { "id": "49161", "postDate": "06/15/2014 18:36:57", "content": "

Snooping the test set with multiple submissions to estimate the repeat percent of 7205 might tell you something. But it's not in the spirit of solving the problem and my efforts in this direction have not helped my score much. So for now I'm just ignoring the offer effect for offers not in the test set and trying to improve in other ways. 

\n\n

(This is not why I have so many submissions. I actually can't explain that except that I've had lots of unproductive ideas).

", "rawMarkdown": "", "votes": null }, { "id": "49164", "postDate": "06/15/2014 19:53:00", "content": "

James,

\n\n

I've had a ton of unproductive ideas that I was positive would pay off but provided negative gain, probably I;m getting bad feature interactions and need to prune my features.  It is always frustrating when you spend a lot of time on a new feature and then it doesn't provide anything of benefit.

\n\n

I can't wait for this competition to be over so I can reveal some of my features and see if others can get them to work.  I have a suspicion that with a properly tuned algorithm they will give a big boost.

", "rawMarkdown": "", "votes": null }, { "id": "49172", "postDate": "06/16/2014 03:02:21", "content": "

[quote=LetterRip;49164]

\n

It is always frustrating when you spend a lot of time on a new feature and then it doesn't provide anything of benefit.

\n

[/quote]

\n\n

Agreed. I confess to being completely bamboozled by this competition; one I thought I would be able to do well on. Breakfast Pirate's comments seem to be a the heart of it; most of the test data is from offers not on the training set, so deducing something about how they behave from the data we do have is crucial. I see some evidence that the compay/brand/category buying behavior for the out of sample offers is different from the offers we have in the test set.

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 42324, "author_name": "zachmayer", "author_url": "", "post_date": "04/18/2014 15:38:13", "content": "

[quote=BreakfastPirate;42289]

\n

The repeat % of some categories are affected by previous purchases of the same brand (e.g. 1703 & 1726). Others are not (e.g. 5619). (A similar analysis could be done for previous purchases of same company, category, etc.)

\n

[/quote]

\n\n

Could you describe how you did this analysis?

", "votes": null, "replies": [] }, { "id": 42333, "author_name": "breakfastpirate", "author_url": "", "post_date": "04/18/2014 17:43:59", "content": "

For each row in the training set I:

\n", "votes": null, "replies": [] }, { "id": 44303, "author_name": "decipher", "author_url": "", "post_date": "05/09/2014 14:03:30", "content": "

@BreakfastPirate - I am bit confused with a problem statement. Do we have find out the probability of person redeeming the offer OR do we have to find out probability of person redeeming the offer multiple times?

\n

Looking forward for your/anyone's reply ? 

", "votes": null, "replies": [] }, { "id": 44305, "author_name": "breakfastpirate", "author_url": "", "post_date": "05/09/2014 14:18:34", "content": "

My understanding is that the customers have already redeemed the offer, and we are predicting whether they will buy the product again after the initial incented offer.  This is based on the problem description: "The most valuable customers are those who return after this initial incented purchase."

", "votes": null, "replies": [] }, { "id": 44346, "author_name": "decipher", "author_url": "", "post_date": "05/10/2014 16:05:16", "content": "

Thanks for your reply. 

\n

So does that means that ALL customers in trainhistory.csv have redeemed the offer earlier atleast once and the repeater = t indicates the customer has made a repeat purchase after the first purchase. Or what i am saying is that all the customers are those who used the offer once and we are predicting their chances of making same purchase again.

\n

Also, is the initial incented offer is before the offer date in trainhistory.csv file? As there are customers whose repeater = f or repeattrip = 0 which means they have never come for a purchase after the offer date. Please help me understanding this as i am still confused :(

", "votes": null, "replies": [] }, { "id": 49151, "author_name": "anshul", "author_url": "", "post_date": "06/15/2014 09:53:00", "content": "

Even my understanding is same as that of Decipher's. 

", "votes": null, "replies": [] }, { "id": 49161, "author_name": "jfkingiii", "author_url": "", "post_date": "06/15/2014 18:36:57", "content": "

Snooping the test set with multiple submissions to estimate the repeat percent of 7205 might tell you something. But it's not in the spirit of solving the problem and my efforts in this direction have not helped my score much. So for now I'm just ignoring the offer effect for offers not in the test set and trying to improve in other ways. 

\n\n

(This is not why I have so many submissions. I actually can't explain that except that I've had lots of unproductive ideas).

", "votes": null, "replies": [] }, { "id": 49164, "author_name": "letterrip", "author_url": "", "post_date": "06/15/2014 19:53:00", "content": "

James,

\n\n

I've had a ton of unproductive ideas that I was positive would pay off but provided negative gain, probably I;m getting bad feature interactions and need to prune my features.  It is always frustrating when you spend a lot of time on a new feature and then it doesn't provide anything of benefit.

\n\n

I can't wait for this competition to be over so I can reveal some of my features and see if others can get them to work.  I have a suspicion that with a properly tuned algorithm they will give a big boost.

", "votes": null, "replies": [] }, { "id": 49172, "author_name": "jfkingiii", "author_url": "", "post_date": "06/16/2014 03:02:21", "content": "

[quote=LetterRip;49164]

\n

It is always frustrating when you spend a lot of time on a new feature and then it doesn't provide anything of benefit.

\n

[/quote]

\n\n

Agreed. I confess to being completely bamboozled by this competition; one I thought I would be able to do well on. Breakfast Pirate's comments seem to be a the heart of it; most of the test data is from offers not on the training set, so deducing something about how they behave from the data we do have is crucial. I see some evidence that the compay/brand/category buying behavior for the out of sample offers is different from the offers we have in the test set.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "42289": "", "42324": "", "42333": "", "44303": "", "44305": "", "44346": "", "49151": "", "49161": "", "49164": "", "49172": "" }, "source": "meta" }