viethhuynh/mc2svi

Scalable nonparametric Bayesian multilevel clustering

Source codes for stochastic variational inference for nonparametric Bayesian multilevel clustering models (MC2SVI) [Java/ Apache Spark]

This package implements the stochastic variational inference for nonparametric Bayesian multilevel clustering models (MC2SVI) described in the following paper:

Huynh, Viet, Phung, Dinh, Venkatesh, Svetha, Nguyen, Xuan Long, Hoffman, Matt and Bui, Hung Hai 2016, Scalable nonparametric Bayesian multilevel clustering, in UAI 2016: Proceedings of the 32nd Conference on Uncertainty in Artificial Intelligence, AUAI Press, Corvallis, Or., pp. 289-298.

Disclaimer: We have made our best effort in ensuring fairness in acknowledging existing codes and any materials we used. However, if you have any question/concern, please write to us.

Using the code

Data

Each dataset includes three data files: content, context and meta data. In data folder, a sample of dataset, NIPS is included:

• content_nips.txt: the content file which contains spare vector in libsvm format
• context_nips.txt: the context file spare vector in libsvm format
• meta_nips.txt: describe the dimensions of content and context data

Configuration file: config.properties

• mc2.trunM=150 % truncation level for number of topics

• mc2.trunK=80 % truncation level for number of clusters

• mc2.trunT=100 % truncation level for number of topics for each cluster (restaurant)

• mc2.aa = 10 % concentration for cluster proportion

• mc2.ee = 10 % concentration for topic proportion at restaurant level

• mc2.vv = 10 % concentration for topic proportion d

• mc2.batchSize=100 %mini-batch size

• mc2.numIter=1 % number of running epochs

• mc2.varrho = 1 % learning rate

• mc2.iota = 0.8 % learning rate

• mc2.contentDirichletSym=0.001 % prior parameter for content

• mc2.contextDirichletSym=0.1 % prior parameter for context

• mc2.contextType=Multinomial % context distribution type

• mc2.metaPath=meta_nips.txt % path to meta data file

• mc2.contentPath=content_nips.txt % path to content data file

• mc2.contextPath=context_nips.txt % path to context data file

• mc2.outFolderPath=out % path to output folder

Install Apache Spark on the local machine

• Java 7
• Installation: Download Spark 1.5.1 from http://spark.apache.org/downloads.html (spark-1.5.1-bin-hadoop2.6.tgz) unzip to folder spark-1.5.1-bin-hadoop2.6
• Set PATH to the folder spark-1.5.1-bin-hadoop2.6

Running

• Open command line ( terminal)
• Change to code folder
• Run: spark-submit --master local[8] BNPStat.jar config.properties
• Output will be store in mc2.outFolderPath

Output (in matlab file format)

The variables stored in each matlab file after running each mini-batch

• pp: the content atoms
• qq: the context atoms
• qcc: corresponding to μ^c
• qzzs: corresponding to μ^z
• rhos, varphis, zetas: stick breaking hyperparamters (corresponding to λ^β,λ^ϵ,λ^τ)