performance of bpr

[jackyhawk]

Hi all,

I just run bpr ,
and my input data is like this:
5 million users
680,000 items
5 billion clicks

and my machine is as the details at bottom.

and it would take me about 3.5 hours to train (convergence) with 20 epoch.

and is there any way to accelerate the speed of the training?

and my machine is like this:

lsb_release -a
LSB Version: :core-4.1-amd64:core-4.1-noarch:cxx-4.1-amd64:cxx-4.1-noarch:desktop-4.1-amd64:desktop-4.1-noarch:languages-4.1-amd64:languages-4.1-noarch:printing-4.1-amd64:printing-4.1-noarch
Distributor ID: CentOS
Description: CentOS Linux release 7.3.1611 (Core)
Release: 7.3.1611
Codename: Core

cat /proc/cpuinfo| grep "physical id"| sort| uniq| wc -l
2

cat /proc/cpuinfo| grep "cpu cores"| uniq
cpu cores : 8

cat /proc/cpuinfo| grep "processor"| wc -l
32

Thanks very much

[jackyhawk]

Any suggestions for this? Thanks very much

[jackyhawk]

And would als be a lot fater that bpr?

Thanks very much

[zenogantner]

In my experience ALS is often faster. Just try it out!

[jackyhawk]

Thanks very much.

And I wonder why bpr is so slow?

[zenogantner]

There may be implementation details that contribute to it, but in general:

ALS (WRMF etc.) is linear in the number of observations and cubic in the number of factors (there may be better speed-ups nowadays).
BPR is also linear in the number of observations and linear in the number of factors, but needs more epochs to converge.

[jackyhawk]

it seems that als is slower than bpr,
with the same amount of data( 0.5 billion)

[jackyhawk]

it would take about 18 hours to train bpr and 40 hours to train als for about 0.5 billion training data