BPRMF in cold start conditions

[evge...@is.haifa.ac.il]

My research is focused on the cold-start scenario with item recommendation from positive-only feedback option. Some users in our dataset only have a known (positive) feedback for a single item. It may happen that no other user provided feedback for the same item. In such cases, one expects recommendation performance to be zero (i.e., have no recommendations generated). But, when experimenting with BPRMF, MyMediaLite gave relatively high performance.
I tried different options to use MML – from the code, from the command line. For example, using the following command line

item_recommendation --training-file=mf_train.dat --test-file=mf_test.dat --recommender=BPRMF --measures=MAP

where the test file includes one user, the training file includes only one feedback for the tested user and no other user provided feedback for the same item (the both training and test file are attached), I obtain MAP around 0.12.
The question is, does MyMediaLite apply backoff policy in such cases? If so - what is that policy (back off to popularity, perhaps?).

[Zeno Gantner]

Hello Evgenia,

Which if the 5 options for item selection (see output of --help) did you choose for evaluation?

I'd suggest you run the evaluation for all options except the file option and then see whether you still have questions.

Best regards,

    Z.

--
You received this message because you are subscribed to the Google Groups "MyMediaLite" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mymedialite...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.