Adding Jianping Wang to the thread, since he contributed the SVDPlusPlus implementaiton.
Jianping, Can you take a look at this message? Thanks. On Fri, Apr 3, 2015 at 8:41 AM, Michael Malak < michaelma...@yahoo.com.invalid> wrote: > I believe that in the initialization portion of GraphX SVDPlusPluS, the > initialization of biases is incorrect. Specifically, in line > > https://github.com/apache/spark/blob/master/graphx/src/main/scala/org/apache/spark/graphx/lib/SVDPlusPlus.scala#L96 > instead of > (vd._1, vd._2, msg.get._2 / msg.get._1, 1.0 / scala.math.sqrt(msg.get._1)) > it should be > (vd._1, vd._2, msg.get._2 / msg.get._1 - u, 1.0 / > scala.math.sqrt(msg.get._1)) > > That is, the biases bu and bi (both represented as the third component of > the Tuple4[] above, depending on whether the vertex is a user or an item), > described in equation (1) of the Koren paper, are supposed to be small > offsets to the mean (represented by the variable u, signifying the Greek > letter mu) to account for peculiarities of individual users and items. > > Initializing these biases to wrong values should theoretically not matter > given enough iterations of the algorithm, but some quick empirical testing > shows it has trouble converging at all, even after many orders of magnitude > additional iterations. > > This perhaps could be the source of previously reported trouble with > SVDPlusPlus. > > http://apache-spark-user-list.1001560.n3.nabble.com/GraphX-SVDPlusPlus-problem-td12885.html > > If after a day, no one tells me I'm crazy here, I'll go ahead and create a > Jira ticket. > > --------------------------------------------------------------------- > To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org > For additional commands, e-mail: dev-h...@spark.apache.org > >
