timeSVD++ fine tuning

[Ankush Sachdeva]

I was trying to run timeSVD++ on the netflix dataset, but i am not able to decide upon the number of time bins. The koren paper suggests a broader bin size(30) for item bias and a finer bin size for user bias and factor matrix.
Why does graphchi take only one time bin for everything?
And incase anyone has already tried it, what should be the optimal number timebins and other parameters for best RMSE?

Thanks

[Danny Bickson]

Selecting the number of time bins is a matter of trial and error. Typically tends of time bins are fine.

The reason we bin is because if we take the rating time in seconds there will be too many different values. Close times are not grouped together. We group time bins together hoping for finding similar behavior along the same time bins.

There is no binning for the user/items.

Best,

		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/ac80fa90-9d17-4ca3-80e0-bb469a697dba%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Ankush Sachdeva]

Thanks for your reply.

Is it not exactly following the Koren's paper Collaborative Filtering with Temporal Dynamics?, because there we had the bias' also as a functions of time.

Also, since its a very popular dataset, I was hoping if someone could have already tested on it and could help me with the parameter values.

--
You received this message because you are subscribed to a topic in the Google Groups "graphchi-discuss" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/graphchi-discuss/GFeZpt4a1Co/unsubscribe.
To unsubscribe from this group and all its topics, 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/CAGKFuX-7kXF61aVupqXQ%2BthCPPKEewjQHu_6Nt0ok2HJKbyx2A%40mail.gmail.com.

For more options, visit https://groups.google.com/d/optout.

--

Ankush Sachdeva

Junior Engineering Student
Deptt. Of Computer Science

IIT Kanpur

[Danny Bickson]

Hi, 

You are right, we are not using the bias for the time bins. For our dataset we did not find it effective. 

However, it is a change of two lines to include it. 

You will need to change the following two lines:

https://github.com/GraphChi/graphchi-cpp/blob/master/toolkits/collaborative_filtering/timesvdpp.cpp#L161

 for adding the time bias into the prediction and also here:

https://github.com/GraphChi/graphchi-cpp/blob/master/toolkits/collaborative_filtering/timesvdpp.cpp#L305

to update the time bias as done with the other gradients. 

Best,

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

To view this discussion on the web visit https://groups.google.com/d/msgid/graphchi-discuss/CAG45n0Uw9oUEWqeM9q%3DzhTSa8LVRRfELKWH0%3D_o6UpjTFKHQoA%40mail.gmail.com.