OK, got your points, thanks Ferrel and Peng.

On Sun, Sep 21, 2014 at 11:39 PM, Pat Ferrel <[email protected]> wrote:

> @li
>
> Mahout down-samples the input data based on how “important” the
> cooccurrence of interactions seems to be. I’d use SIMILARITY_LOGLIKELIHOOD
> for the best measure of this. You will still get no recs for some users if
> they don’t have enough interactions or they are not “important”
> interactions.
>
> On Sep 15, 2014, at 7:38 PM, Peng Zhang <[email protected]> wrote:
>
> As far as I know, mahout would not do any post processing to remove
> records from recs. I assume you refer to recs as recommendation candidates.
> If you find some user is not in recs (the output), please make sure he/she
> has sufficient interactions with the items.
>
> On Sep 16, 2014, at 10:17 AM, Wei Li <[email protected]> wrote:
>
> > Thanks Peng.
> >
> > Yes, I agree your points, there may not be enough interactions between
> > users and items to do the recommendations, but do our Mahout code does
> some
> > extra filtering to remove the recommendation candidates?
> >
> > On Mon, Sep 15, 2014 at 3:35 PM, Peng Zhang <[email protected]>
> wrote:
> >
> >> Mahout does not guarantee specified recs for each user. There are many
> >> reasons, e.g, there might not be enough similar users or items for a
> user.
> >>
> >> Peng Zhang
> >>
> >> --
> >> Sent from my iPhone
> >>
> >>> On Sep 15, 2014, at 3:15 PM, Wei Li <[email protected]> wrote:
> >>>
> >>> Hi Mahout Users:
> >>>
> >>>  We are using the RecommderJob to perform the item-based
> >>> recommendations, the following settings are used:
> >>>
> >>> similairtyClassname=SIMILARITY_COOCCURRENCE
> >>> numRecommendations=20
> >>> other parameters are set to default values
> >>>
> >>> while we see that the size of the recommendation results for some users
> >> is
> >>> less than 20, only 1 or 2. Since we have no time to dive into the
> source
> >>> code now, we do know if we see the right parameters. Does any one can
> >> help
> >>> us on this issue? many thanks :)
> >>>
> >>> Best
> >>> Wei
> >>
>
>
>

Reply via email to