This repository contains classifiers and inference/evaluation methods for the paper Online Multi-Label Dependency Topic Models for Text Classification. The paper will be published in the Machine Learning Journal.

Included classifiers are:

• BR (1. modified from Meka, 2. a python script using LIBLINEAR for the SVMs)
• SCVB
• SCVB-Dependency
• Dependency-LDA
• Fast-Dependency-LDA

to build the project run mvn package to run a classifier, you need to specify options in the config.txt file, e.g. for Dependency-LDA:

 --dataFilename data/datafile --testDataFilename data/datafile --datasetName name --numIterations 100 --numTopWords 20 --wordThreshold 5 --batch --dependency --alphaSum 0 --testAlphaSum 30 --beta 0.01 --beta_c 0.01 --gamma 10 --testGamma 10 --eta 100 --testEta 100 --innerIterations 1 --t 100 --supervised true --trainingPercentage 1.0 --evaluationInterval 100 --writeDistributions false --print false --run 1

You may then run the program with java -jar target/Multi-Label-Topic-Modeling-1.0-SNAPSHOT.jar linenumber where linenumber is the line in the config.txt file that you want to use

General options are:

• batch run classifier in batch mode with a train and a test set
• incremental run classifier in incremental mode where the classifier is first trained on an initial batch and then alternately tested on the next batch before being trained on the next batch (only available for SCVB and SCVB-Dependency)
  • additional parameter to specify: --batchSize (e.g. batchSize 100 means train on 100 instances then test on the next 100 instances etc.)
• dataFilename the filename with the training data
• testDataFilename the filename with the testing data
• datasetName the name to use for the path to the results file
• writeDistributions true or false, whether or not to write the distributions for the testdataset to file, might be needed for later averaging, takes up a lot of memory depending on the size of the testset and the number of labels
• print true or false, whether or not to write the resulting word clouds to file
  • further options: numTopWords and wordThreshold for the number of words per label and the frequency threshold for writing words
• evaluationInterval after how many iterations the classifier is tested on the testset
• run an integer may be specified that is included in the filename to separate multiple runs with the same configuration
• supervised currently always has to be set to true

The options for the different methods are:

• Dependency-LDA, Fast-Dependency-LDA:
  • dependency for using Dependency-LDA
  • fastdependency for using Fast-Dependency-LDA
  • s2 set to true if you want to use the evaluation strategy S2
  • alphaSum for training
  • testAlphaSum for testing
  • beta
  • beta_c corresponds to beta_y in the paper
  • t number of topics to use
  • gamma for training (not for Fast-Dependency-LDA)
  • testGamma for testing (not for Fast-Dependency-LDA)
  • eta for calculation of alpha' during training (not for Fast-Dependency-LDA)
  • testEta for calculation of alpha' during testing (not for Fast-Dependency-LDA)
  • innerIterations currently not implemented, use a value of 1 (not for Fast-Dependency-LDA)
• SCVB, SCVB-Dependency:
  • SCVB for using SCVB
  • SCVBDependency for using SCVB-Dependency
  • alpha
  • alphaDep not used
  • eta corresponds to beta
  • etaDep corresponds to beta_y
  • rho update parameter (set to 0.9)
  • rhophi update parameter (set to 0.9)

The format of the input file is as follows: Each line is one input document. There are three columns expected, separated by Tabs

1. ID of the document
2. one or more labels, separated by white spaces
3. the text of the document itself