The input to the MinHashDriver are tf-vectors, e.g.:
hdfs -text /..../tf-vectors/part-r-00000
10020926
{96037:1.0,114292:1.0,119745:1.0,64091:1.0,84488:1.0,116350:1.0,135598:1.0,53382:1.0}
10020927
{53382:1.0,1076:1.0,116350:1.0,36446:1.0,135598:1.0,96037:1.0,84488:1.0}
10020928
{53382:1.0,58740:1.0,13358:1.0,116350:1.0,126929:1.0,135598:1.0,136284:1.0}
10020929
{53382:1.0,13358:1.0,116350:1.0,126929:1.0,135598:1.0,114292:1.0,76732:1.0}

So the vectors are sparse and indexed by wordId. Clearly, this example
doesn't fit into current implementation.

On Mon, Jul 30, 2012 at 2:19 PM, Sean Owen <[email protected]> wrote:

> Hmm yeah what's the output format of that job? At first glance it looks
> like it makes a dictionary and fills in sequentially with term frequencies.
> If so, applying min-hash this way isn't wildly wrong but I still don't
> think it works.
>
> Can anyone who knows this code at all weigh in? it's not my area...
>
> On Mon, Jul 30, 2012 at 1:11 PM, Elena Smirnova <[email protected]
> >wrote:
>
> > If vectors are treated as dense, then we have to modify the example given
> > for this class, which clearly talks about documents and words:
> > https://cwiki.apache.org/confluence/display/MAHOUT/Minhash+Clustering
> >
> > On Mon, Jul 30, 2012 at 2:02 PM, Sean Owen <[email protected]> wrote:
> >
> > > Yes I know what you mean. In my understanding you typically apply
> minhash
> > > to a large sparse vector that acts like a bit set, where the index is
> > > really the set member. There you want to hash the index, and doing so
> by
> > > considering all indices would be completely wrong.
> > >
> > > Here I think the set elements are the values. and the vectors seem to
> be
> > > treated as a list, really. So I'm not surprised they're treated as
> > dense. I
> > > still think it's a good idea to iterate over non-default items, since
> I'm
> > > not clear whether the implementation is guaranteed to accept only dense
> > > input vectors, where all dimensions have a value -- in which case it
> > > doesn't matter and the current implementation is OK.
> > >
> > > Ankur are you still around to answer? I think that's a good guess as to
> > the
> > > original intent.
> > >
> > > On Mon, Jul 30, 2012 at 12:51 PM, Elena Smirnova <[email protected]
> > > >wrote:
> > >
> > > > I agree about performance effect of iterating over zeros. But the
> > > > correctness effect comes due to hashing values of the element and not
> > its
> > > > index (at least in documents and words example).
> > > >
> > > > Do you agree?
> > >
> >
>

Reply via email to