Thanks to you both, Peng/Pat.

Sure I can map them out. Not a problem. The issue is the output returns all the 
matches from within the same set. So while I am only interested in the 
inter-set comparison (A vs B), the output I get has only intra-set comparisons 
(A vs A and B vs B). Upon filtering, there is no entry left for the A vs B.

Sent from my iPad

> On Jul 10, 2014, at 7:44 PM, Pat Ferrel <[email protected]> wrote:
> 
> Right, as Peng says IDs are the row and column numbers in the input matrix 
> 0..n, 0..m You assign them by incrementing for each unique application 
> specific id. Use a hashmap or databased to maintain your own mapping 
> dictionary from your IDs below into and out of Mahout IDs. The matrix is 
> users = rows, items = columns.
> 
> So for a particular item in the A set you want to recommend something similar 
> from the B set. You do this by creating an input set that meats the above 
> needs for IDs, the preference can be your rating number but often translating 
> that into a boolean is as good if not better. Either way is OK but no 0 or 
> below number for preference. The ID will be for all users and for all books 
> from either set.
> 
> Then you want to ask the ItemBasesRecommender for
> 
>  List<RecommendedItem> mostSimilarItems(long itemID, int howMany) throws 
> TasteException;
> 
> Filter out any items you don’t want from the list and return the ones with 
> the highest strength. Notice the itemID is the Mahout column number for the 
> item in question. So if You were on an Amazon page for a book it would be the 
> Mahout ID for the book the user was looking at. The results will be from A 
> and B so filter the ones you don’t want.
> 
> 
> 
> 
> On Jul 10, 2014, at 7:11 PM, Peng Zhang <[email protected]> wrote:
> 
> Jiten,
> 
> User and Book ids should be 0,1,2 etc. Plz don't use negative integers. 
> 
> Since I don't understand +ve/-ve you mentioned, but if you know which books 
> belong to class A/B, you can filter the output correlation matrix yourself to 
> get what you want. 
> 
> Peng Zhang
> 
> --
> Sent from my iPhone
> 
>> On Jul 11, 2014, at 9:49 AM, "Jiten Gore" <[email protected]> wrote:
>> 
>> Thanks for your reply, Pat. Here's an illustration:
>> 
>> The input set looks like this:
>> User    Book    Preference
>> 1    10    7
>> 2    10    4
>> 3    10    5
>> 1    15    3
>> 2    15    8
>> 3    15    4
>> 1    20    9
>> 2    20    2
>> 3    20    3
>> 1    -1    8
>> 2    -1    8
>> 3    -1    2
>> 1    -2    5
>> 2    -2    6
>> 3    -2    4
>> 1    -3    7
>> 2    -3    5
>> 3    -4    3
>> ...
>> 
>> For this set for the books (+ve set) I am interested in only the
>> recommendations for the Books (-ve set) 
>> 
>> The desired output (the values are made up):
>> Book1    Book2    ρ-value
>> 10    -2    0.75
>> 10    -3    0.66
>> 10    -1    0.41
>> 15    -1    0.80
>> 15    -2    0.55
>> 15    -3    0.21
>> 20    -2    0.91
>> 20    -1    0.65
>> 20    -3    0.47
>> 
>> Thus, I would *not* like to see the ones like these two:
>> 10    20    0.61
>> -1    -2    0.73
>> 
>> Does this help in clarifying? And yes I am interested in Pearson correlation
>> and I am using Hadoop.
>> 
>> Best,
>> Jiten
>> 
>> -----Original Message-----
>> From: Pat Ferrel [mailto:[email protected]] 
>> Sent: Thursday, July 10, 2014 1:42 PM
>> To: [email protected]
>> Subject: Re: Recommender question
>> 
>> Can you rephrase this "I would like to see recommendations for A class books
>> only out of those from class B."
>> 
>> 
>> On Jul 10, 2014, at 11:25 AM, Jiten Gore <[email protected]> wrote:
>> 
>> Dear Mahout community,
>> 
>> I am running an item based recommender on Mahout and using the Amazon
>> paradigm, consider that there are books - two classes of books, one (A)
>> coded with +ve ID values and the other (B) with -ve values for ID.
>> 
>> I would like to see recommendations for A class books only out of those from
>> class B. The user ratings structure is the same for both classes.
>> 
>> I will appreciate any pointers you might have.
>> 
>> Best,
>> Jiten
>> 
>> Sent from my iPhone
> 
> 

Reply via email to