mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sean Owen <sro...@gmail.com>
Subject Re: Custom Item Similarity :datamodel not sure
Date Thu, 27 Sep 2012 18:21:48 GMT
Sure, you don't need a recommender then.  You are just computing
most-similar items using some similarity metric. Forget the
recommender and just loop over all items computing similarity to a
given item and keep the top N. You can look at the mostSimilarItems()
method for a template for how to do it.

I am not sure what item,item input you would construct to feed to a
recommender -- this isn't a recommender problem.

On Thu, Sep 27, 2012 at 7:14 PM, Abhishek Roy <abhishekroy8@gmail.com> wrote:
> Sean Owen <srowen <at> gmail.com> writes:
>
>
>>
>> The input to the recommender remains the same -- user,item,rating.
>> Your similarities are used as weights in a weighted average to make
>> recommendations. This is unrelated -- or rather, not necessarily
>> related at all -- to whatever custom similarity metric you create.
>> Your similarities do not need to be precomputed. You could, but it's
>> not necessary.
>>
>> On Thu, Sep 27, 2012 at 6:48 PM, Abhishek Roy <abhishekroy8 <at> gmail.com>
> wrote:
>> > Sean Owen <srowen <at> gmail.com> writes:
>> >
>> >>
>> >> File for FileDataModel? This does not change. But that input does not
>> >> consist of item item pairs. Are you talking about something else?
>> >> On Sep 27, 2012 5:10 PM, "Abhishek Roy" <abhishekroy8 <at> gmail.com>
> wrote:
>> >>
>> >> > Hi Sean,
>> >> > For using a custom ItemSimilarity what should my data model file(item
> id1,
>> >> > item
>> >> > id2) include ?
>> >> >
>> >> > Please advise.
>> >> >
>> >> > Thanks,
>> >> > Roy
>> >> >
>> >> >
>> >>
>> >
>> > Thanks for the quick response Sean.
>> > My end goal (short term) is to show "related / similar" items for my site
> when
>> > the user(any user, including unregistered user) is looking at a particular
> item.
>> > Basically I am looking at (rather created) a custom ItemSimilarity using
> domain
>> > specific attributes that computes a similarity score between a pair of
> items. I
>> > am using a GenericItemBasedRecommender and then calling n mostSimilarItems()
> to
>> > get my recommendations. The problem is, and I didn't see anything on that in
> the
>> > book as well as the forum, that I am not sure about the data model to feed
> to
>> > the GenericItemBasedRecommender. I did a brute force, computed
> nC2(combinations)
>> > of {item id, item id} pairs and fed that as the data model. Works, but
>> > definitely not scalable and sensible. What data model does this kind of a
> system
>> > need ? I am not having preference data(very little), and since this is
> content
>> > based recommendation, am puzzled about the data to be encapsulated by the
>> > datamodel. I hope I am clear..
>> > Please suggest...
>> >
>> >
>> >
>>
>>
>
> Sean, let me clarify. I am in a way, trying to recommend "similar" items to a
> particular item. 100% Content based.
> And I don't have the user,item,rating data. No user angle at all. No
> "preference" angle at all.
>
> All I have is the set of all items in the system, and their
> attributes(genre,title,description etc). I have read and realized that the
> user,item,rating data can as well be : item,item data ...(rating/preference)
> absent. Hence the confusion. So, in this case, what data do I give as an input ?
> Do I compute item,item entries based on a certain criteria ? What is the least
> data I can give the system as an input to get my n most similar items based on
> my custom ItemSimilarity ?
>
>
>

Mime
View raw message