mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gökhan Çapan <gkhn...@gmail.com>
Subject Re: Predicting Successor Item
Date Thu, 17 Jun 2010 07:48:09 GMT
Sean, Ted,
Thanks for sharing your ideas.
I will try things we have talked about, compare with current recommendation
techniques in terms of performance and accuracy, and share the results with
all Mahouts.

On Wed, Jun 16, 2010 at 7:05 PM, Ted Dunning <ted.dunning@gmail.com> wrote:

> I would follow Sean's suggestion and try simpler methods first.  My guess
> is
> that the important structure of the HMM may be much easier to learn by
> sparsification techniques.
>
> Sequence aware methods also have potential for harm in that they may just
> be
> reverse-engineer your current link structure.  Reverse links might well be
> just as useful to the customer.
>
> I generally subtract the users history from recommendations in any case.
>  That gives most of the benefit of sequence based recommenders (in the
> areas
> of most historical interest for me ... music and videos).  Even without
> direct sequence information, I found  that a simple item based recommender
> would discern the episodic structure of video series without any problem.
>  The top link would be the next episode and the second best link would
> typically be the previous episode, but if the person had just come from the
> previous episode, it wouldn't appear.
>
> On Wed, Jun 16, 2010 at 12:44 AM, Gökhan Çapan <gkhncpn@gmail.com> wrote:
>
> > My first HMM ideas came
> > after this, but Sean said that it may be overkill. What do you think
> about
> > HMM, is it worth to try?
> >
>



-- 
Gökhan Çapan

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message