Yes, that was not clear. The simple answer is that the recommendations are
bad in this case. Describe the recommender you are using and the data - we
have to have some information to offer any ideas.

I see, you do not have source. That makes it hard to investigate. If it
really becomes necessary, maybe you can run the Java version on your data
and debug. But first maybe we can think of ideas as to why it might not work
well.

Usually it is a function of not enough data, or just needing to try a
different algorithm

On Sep 2, 2009 11:09 AM, "Claudia Grieco" <[email protected]> wrote:

Maybe you misunderstood me.
I sent you emails in the past about another recommender of mine (I'm testing
3 of them). That time the problem was that I was running the first release,
that contained an older and bugged version of the eval class, so I upgraded
to a newer version (but still pre-Mahout150) and that solved the problem.
Changing recommender type, I encounter a DIFFERENT problem.
The reason I can't debug properly is that the ported version of the code
(done using IKVM that creates dll out of jars) is treated by the c# debugger
as a big black box so I can debug my own classes, but not mahout's.


-----Messaggio originale-----
Da: Sean Owen [mailto:[email protected]]
Inviato: mercoledì 2 settembre 2009 11.21
A: [email protected]
Oggetto: Re: problems evaluating item based recommender

Claudia I believe I've replied to you several times on this mailing
list with possible reasons you are observing this behavior. I am not
clear that you followed up on them, so would point you back to those
first.

You're using a port, of an old version of the code, and we don't have
access to your data, or your code, and you are unable to debug. I am
not sure it is really possible to offer more support then. I strongly
suggest you learn how to run a debugger; it will be extremely useful
here and should point out the cause immediately. It is definitely
possible to debug C#.

It might be useful to compare the port to the current version of these
classes in Java to see if any meaningful changes have occurred that
you might want to re-port.

Mahout 0.2 is probably coming in a month or so. I think this part of
the code is reasonably stable. I don't have any changes planned, and
it's had some time to work out bugs. But I would first encourage you
to get a debugger going rather than update. I don't think it will
'fix' this problem, I think it's a function of your data.

2009/9/2 Claudia Grieco <[email protected]>:
> I'm using an old version of mahout (before the big branch of mahout-150)
and
> the evaluation of my item based recommender done with
> GenericRecommenderIRStatsEvaluator doesn't work very well: it always
returns
> precision 0 and recall 0. I'm using a port of mahout in C# so I can't
debug
> it properly, but I've seen that the program seems to recommend almost the
> same items to everyone and in numerical order (es. 612,613,614.).
>
> What do you think could be the problem?
>
>
>
> (anyway, how much do you think it would take to adjust a program to pass
> from the old version of mahout to the new one? Is there going to be a
stable
> new release in the near future?)
>
>

Reply via email to