Using 1-5 Ratings for CLIMF and MRR

User 199 | 4/8/2014, 12:41:30 PM

Hi all, it appears it is possible to use ratings between 1-5 for CLIMF, it runs and reports a results. Does GraphChi turn 1-5 ratings into binary (I believe CLIMF operates on binary input).

Second question: which MRR values are good? I see Validation MRR: 0.12635 for instance.. Too high? Too low?

Comments

User 6 | 4/8/2014, 12:48:22 PM

Hi, Regarding CliMF, as you can see here: https://github.com/GraphChi/graphchi-cpp/blob/master/toolkits/collaborativefiltering/climf.hpp#L71 there is a variable called binaryrelevancethresh which you can set during runtime for example using --binaryrelevance_thresh=3 - any value greater than 3 will be a binary 1 and any value lower or equal to 3 will be a zero.

MMR metric along with some examples is explained here: http://bickson.blogspot.co.il/2013/04/climf-algorithm-in-graphchi.html - the optimal value is 1, so 0.126.. does seem a bit low.


User 199 | 4/8/2014, 1:13:33 PM

Thanks Mr Bickson