Kaggle's Greek Media Monitoring Multilabel Classification (WISE 2014)

This repo holds the MATLAB code I used to make submision to Kaggle's Greek Media Monitoring Multilabel Classification (WISE 2014). The score using this implementation is 0.75342, ranking 24th out of 121 teams. (That entry is placed in ./Submission folder.)

For the MATLAB code, there is also a published html page in the ./Opt_Mean_F1score/html folder.

Method

It implements the following methods to tackle the multilabel classification problem:

• Binary Relevance (BR) and Classifier Chains (CC) methods to transform the multi-label classification problem into binary one [1].

• Linear SVM (from LIBLINEAR package) as the "base classifier". In specific, I implemented the SVM.1 and SCutFBR.1 approach as described in [2] and [3].

Requirement

• LIBLINEAR For instructions of using LIBLINEAR and additional information, see the README file included in the package and the LIBLINEAR FAQ.

• Multi-label classification tool: read_sparse_ml.c

For your convenience, I have already included the necessary pre-complied files of LIBLINEAR and read_sparse_ml.c in the folder ./utils/LIBLINEAR. Please see there for copyright information.

Instruction

• download data from the competition website and put all the data into ./Data dir:

  • ./Data/wise2014-train.libsvm
  • ./Data/wise2014-test.libsvm

• put all the code into ./MATLAB dir:$^1$

  • ./MATLAB/Opt_Mean_F1score/...
  • ./MATLAB/Opt_Macro_F1score/...
  • ./MATLAB/utils/...

• run ./MATLAB/Opt_Mean_F1score/train_WISE.m to create csv submission to Kaggle.

$^1$: If you don't want to do this, you have to specify the correct path to the data and to the utils in the function ./MATLAB/Opt_Mean_F1score/train_WISE.m (see the code in the beginning).

Reference

[1] Jesse Read, Bernhard Pfahringer, Geoff Holmes, and Eibe Frank, "Classifier chains for multi-label classification."

[2] David D. Lewis, Yiming Yang, Tony G. Rose, and Fan Li, "RCV1: A new benchmark collection for text categorization research." Journal of Machine Learning Research, 5:361-397, 2004.

[3] Rong-En Fan and Chih-Jen Lin, "A Study on Threshold Selection for Multi-label Classification."