mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sean Owen <>
Subject Re: How to make recommendations using ALS
Date Mon, 04 Jun 2012 08:54:30 GMT
Yes, that's how you do it. You just keep the top N.
This is typically quite fast and is parallelizable across cores trivially.

Yes you can also mix in neighborhood based techniques. You can calculate
user and item similarities in feature space, fast. Cosine similarity is
fine in this space. It would help you consider fewer items.

But it comes with related drawbacks. Using neighborhoods to prune the
search space brings back the problem of being 'disconnected' from most good
recommendations due to sparse data. And you still spend time computing

I would suggest starting with the first approach.
This is precisely what Myrrix does, note. Maybe you are already playing
with that.

On Jun 4, 2012 9:35 AM, "Stemmer, Maya" <> wrote:

> Hi,
> I want to know how to use the decomposition of the rating matrix to make
> recommendations.
> If I want to predict a user preference for an item, I simply calculate the
> dot product of the user's row in the user-features matrix and the item's
> column in the features-items matrix.
> But what if I want to recommend N items to a user?
> Should I predict his preference for all items the same way, and just
> return the top N? Will it still be scalable?
> Or maybe there is another way to do this?
> I've read some papers on SVD explaining that it is also possible to use
> the small matrices to obtain a user/ an item neighborhood based on less
> data.
> Is it implemented in Mahout? Which way is better?
> I'd be grateful for some help.
> Thanks,
> Maya
> ---------------------------------------------------------------------
> Intel Electronics Ltd.
> This e-mail and any attachments may contain confidential material for
> the sole use of the intended recipient(s). Any review or distribution
> by others is strictly prohibited. If you are not the intended
> recipient, please contact the sender and delete all copies.

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