Taste review

[Grant Ingersoll <gs...@apache.org>]

Saw this on Twitter the other day in regards to Taste:  http://www.iletken-project.com/index.php?option=com_content&view=category&layout=blog&id=34&Itemid=59&lang=en

Would be useful to learn from it where possible (there is also some  
likely bias inherent in, since the vendor sells their own solution).   
I suspect with some of the work Sean is doing in terms of refactoring  
right now, that a lot of this will change in the near future.

-Grant

Re: Taste review

[Grant Ingersoll <gs...@apache.org>]

On Aug 3, 2009, at 4:33 PM, Sean Owen wrote:

> Yes I think that is the meaning, and I don't intend to 'expand' along
> these lines (by myself -- sounds like a nice sibling project). I see
> it as fairly separate question, what the inputs means and how they are
> constructed. It's not like supplying a fork but no spoon. So that's
> why I didn't quite get why that's somehow a disadvantage. Kind of like
> criticizing Tomcat for not including Wordpress or something.

Yeah, I wouldn't worry about that kind of stuff.  We get it all the  
time in Lucene as people (especially analysts) somehow don't get that  
it is just an API built in Java.  These kind of technologies always  
need several layers.  Taste is aimed at being a kernel, while a full  
application would need more.

>
> On Mon, Aug 3, 2009 at 6:15 PM, Otis
> Gospodnetic<ot...@yahoo.com> wrote:
>> Regarding "only doing CF" - maybe they mean to imply there are no  
>> other inputs other than co-visitation data.  For example, no  
>> watching of user behaviour on the page (how long did they stay?   
>> did they scroll?  did they page?  did they highlight a piece of  
>> text while reading?) or consideration of document similarity.
>>
>> Of course, Taste is not at all web-specific, so that user behaviour  
>> is something one would/could build on top of Taste.  But of course  
>> they didn't email mahout-user to discuss this first -- it's not in  
>> their interest. :)
>>

Re: Taste review

[Sean Owen <sr...@gmail.com>]

Yes I think that is the meaning, and I don't intend to 'expand' along
these lines (by myself -- sounds like a nice sibling project). I see
it as fairly separate question, what the inputs means and how they are
constructed. It's not like supplying a fork but no spoon. So that's
why I didn't quite get why that's somehow a disadvantage. Kind of like
criticizing Tomcat for not including Wordpress or something.

On Mon, Aug 3, 2009 at 6:15 PM, Otis
Gospodnetic<ot...@yahoo.com> wrote:
> Regarding "only doing CF" - maybe they mean to imply there are no other inputs other than co-visitation data.  For example, no watching of user behaviour on the page (how long did they stay?  did they scroll?  did they page?  did they highlight a piece of text while reading?) or consideration of document similarity.
>
> Of course, Taste is not at all web-specific, so that user behaviour is something one would/could build on top of Taste.  But of course they didn't email mahout-user to discuss this first -- it's not in their interest. :)
>

Re: Taste review

[Otis Gospodnetic <ot...@yahoo.com>]

Regarding "only doing CF" - maybe they mean to imply there are no other inputs other than co-visitation data.  For example, no watching of user behaviour on the page (how long did they stay?  did they scroll?  did they page?  did they highlight a piece of text while reading?) or consideration of document similarity.

Of course, Taste is not at all web-specific, so that user behaviour is something one would/could build on top of Taste.  But of course they didn't email mahout-user to discuss this first -- it's not in their interest. :)

 Otis
--
Sematext is hiring -- http://sematext.com/about/jobs.html?mls
Lucene, Solr, Nutch, Katta, Hadoop, HBase, UIMA, NLP, NER, IR



----- Original Message ----
> From: Sean Owen <sr...@gmail.com>
> To: mahout-dev@lucene.apache.org
> Sent: Monday, August 3, 2009 10:29:36 AM
> Subject: Re: Taste review
> 
> Agree, of course there's some bias. I think the problem here is he's
> judging the entire framework based on a one-off demo I put together
> for the Netflix data set, which is hardly a good example. The review
> even notes this. The review then goes on to judge the result accuracy
> based on performance of one algo on one data set. Hmm.
> 
> The parts about memory usage are dead on, and that's what I have tried
> to fix recently -- I have driven it down about 40% at least.
> The comment about not having a truly distributed recommender is right
> on too. I don't know what algo they are using but have some sense of
> it -- makes some sense.
> 
> The big part they are overlooking is their approach is completely
> offline. This rather misses the point of any interesting CF
> application in my view, which all involve something far more
> real-time.
> 
> I don't really understand the criticism about a CF framework "only doing CF".
> 
> On Mon, Aug 3, 2009 at 3:20 PM, Grant Ingersollwrote:
> > Saw this on Twitter the other day in regards to Taste:
> > 
>  http://www.iletken-project.com/index.php?option=com_content&view=category&layout=blog&id=34&Itemid=59〈=en
> >
> > Would be useful to learn from it where possible (there is also some likely
> > bias inherent in, since the vendor sells their own solution).  I suspect
> > with some of the work Sean is doing in terms of refactoring right now, that
> > a lot of this will change in the near future.
> >
> > -Grant

Re: Taste review

[Sean Owen <sr...@gmail.com>]

Agree, of course there's some bias. I think the problem here is he's
judging the entire framework based on a one-off demo I put together
for the Netflix data set, which is hardly a good example. The review
even notes this. The review then goes on to judge the result accuracy
based on performance of one algo on one data set. Hmm.

The parts about memory usage are dead on, and that's what I have tried
to fix recently -- I have driven it down about 40% at least.
The comment about not having a truly distributed recommender is right
on too. I don't know what algo they are using but have some sense of
it -- makes some sense.

The big part they are overlooking is their approach is completely
offline. This rather misses the point of any interesting CF
application in my view, which all involve something far more
real-time.

I don't really understand the criticism about a CF framework "only doing CF".

On Mon, Aug 3, 2009 at 3:20 PM, Grant Ingersoll<gs...@apache.org> wrote:
> Saw this on Twitter the other day in regards to Taste:
>  http://www.iletken-project.com/index.php?option=com_content&view=category&layout=blog&id=34&Itemid=59&lang=en
>
> Would be useful to learn from it where possible (there is also some likely
> bias inherent in, since the vendor sells their own solution).  I suspect
> with some of the work Sean is doing in terms of refactoring right now, that
> a lot of this will change in the near future.
>
> -Grant