Bayesian Personalized Ranking from Implicit Feedback in Java

Bayesian Personalized Ranking (BPR) is a per-user learning algorithm for personalized ranking that optimizes Area Under the Curve for each user.

This implementation of BPR is written Java and implemented in Apache Giraph, which is a scalable graph processing framework built on top of Apache Hadoop.

####Input

TXT file where each line is represented as
idI idU positive, where

• idI identifies an item,
• idU identifies a user,
• positive equals 1 if the user idU has an implicit feedback on the item idI, 0 otherwise.

####Output

TXT file where each line is represented as
id factors, where

• id identifies a user or an item
• factors represent a vector of latent factors for predicting feedback.

####Running

hadoop jar $GIRAPH_HOME org.apache.giraph.GiraphRunner
org.apache.giraph.examples.BPR
-eif org.apache.giraph.examples.FactorIdTextEdgeInputFormat -eip $INPUT -op $OUTPUT -vof org.apache.giraph.io.formats.IdWithValueTextOutputFormat -ca numFactors=$NUMFACTORS
-ca learningRate=$LEARNINGRATE
-ca regularization=$REGULARIZATION
-ca iterations=$ITERATIONS
-w $WORKERS

Where
$GIRAPH_HOME is directory for Giraph jar file,
$INPUT is directory for input,
$OUTPUT is directory for output,
$NUMFACTORS is number of factors in vectors of latent factors,
$LEARNINGRATE, $REGULARIZATION are learning rate and regularization for gradient descent, $ITERATIONS is number of iterations to execute

####Literature

Steffen Rendle , Christoph Freudenthaler , Zeno Gantner , Lars Schmidt-Thieme, BPR: Bayesian personalized ranking from implicit feedback, Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, p.452-461, June 18-21, 2009, Montreal, Quebec, Canada.