Github user debasish83 commented on the pull request:
https://github.com/apache/spark/pull/1290#issuecomment-62190595
For matrix factorization we have user x product sparse matrix...You can
think of this sparse matrix as the feature matrix for ANN...Now consider two
matrices H1 and H2 of size feature x rank...where rank is the number of hidden
layers...With this the problem is minimize || X - f(H1'X)H2 || + lambdaL1(H1) +
lambdaL2(H2)
The major difference is can H1'X breaks the way matrix factorization breaks
? If it can then we should be able to use ALS design...or an extension of ALS
design...
But say the hidden layer grows from 1 to 10 (Latest Google paper mentioned
22 layers)...then I don't think this idea works...we have to formulate the
problem on graphx where the model is distributed over workers and not built on
Master
@witgo you think we can break f(H1'X) in ALS way? I have not thought more
on it !
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
