results of PMF in collaborative filtering toolkit

[James Jensen]

I'm confused about the results of running the Bayesian probabilistic matrix factorization method. Other methods such as ALS output two files, [name of input file]_U.mm and [name of input file]_V.mm. When I run PMF, neither of these is created, nor is any other file that seems like it's supposed to take their place.

If I add the "--pmf_additional_output=1" option, I get a U.mm and V.mm file for every sampling iteration of the algorithm. I suppose I could average the products of all of these to get the final reconstructed matrix. But I expected this would be done automatically and that the "--pmf_additional_output=1" would be only for looking under the hood. But if so, where is the normal output?

I'm running it like this:

pmf --training=data.mtx --minval=0 --maxval=1 --D=20 --max_iter=100 --quiet=1 --pmf_burn_in=5

This is with the latest version pulled from GitHub.

I apologize if I'm overlooking something painfully obvious here.

[Danny Bickson]

Hi James, 

You assume correctly, you need to average the products of each user and item pair (omitting the first ones to allow for burn in period).

If there are M users and N items there are M*N*t products to average (where t is the number of iterations you run minus the burn in period).

Are you looking for the top K products? or products over a test set? if so, you can use the --test=filename command line argument.

If you are looking for the top K products you will need to compute them youself,

Let us know if this is clearer.

		Danny Bickson Co-Founder US phone: 206-691-8266 Israeli phone: 073-7312889

--
You received this message because you are subscribed to the Google Groups "graphchi-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to graphchi-discu...@googlegroups.com.
To post to this group, send email to graphchi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/graphchi-discuss/f922de03-a28d-40c4-8e95-4e7ce404fd03%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[James Jensen]

Thanks, Danny. That makes sense. And I'm glad you reminded me about the --test=filename option. I'm not using that now but I think I will later on.

Just to clarify: I notice that, if I do 100 iterations with 5 burn-in iterations, it outputs U.mm and V.mm files indexed 0-94 only. Am I right that they are numbered in reverse order, and that 99-95 were from the burn-in? I.e. the user/item pairs from the burn-in period are omitted automatically?

Best,

James

To view this discussion on the web visit https://groups.google.com/d/msgid/graphchi-discuss/CAGKFuX8V5ySHFMEERqcRdEvW9xJ9%3DktYx0S3AmBYntBnApfhig%40mail.gmail.com.