Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

The code implements the Label Filter algorithm for speeding up prediction in Extreme Classification problems (i.e. multi-label classification problems with an extremely large labels set). Label Filters are a computationally efficient technique for pre-selecting a small set of candidate labels for each test example before applying a more expensive multi-label classifier. For details please see the following paper:

Alexandru Niculescu-Mizil, Ehsan Abbasnejad: "Label Filters for Large Scale Multilabel Classification" - Proceedings of the 20th International Conference on AI and Statistics (AISTATS `17).



The code requires the following libraries:

  • boost for Dynamic Bitset and Program Options libraries
  • gperftools for tcmalloc and profiler
  • Eigen (included as a submodule)
  • CRoaring (included as a submodule)


First modify src/Makefile to point to the correct path for the boost libraries then execute:

cd src
make CRoaring #if using the included submodule

This will compile the label filter library and put the result into the bin/ directory

The compilation will produce the following files:

  • -- label filter library, for dynamic linking
  • libmcfilter.a -- label filter library, for static linking
  • mcsolve -- example program to learn a filter
  • mcpoj -- example program to apply the filter with linear classifiers.
  • convert_data -- program to convert data form XML to binary dense/sparse format
  • convert_linearmodel -- program to convert model files from text to binary format

Using the code

cd Mediamill

gunzip train_split1_Mediamill_data.txt.gz test_split1_Mediamill_data.txt.gz

# learn the label filters
../bin/mcsolve -o mediamill_C1_0.1_C2_0.1.filter --C1=0.1 --C2=0.1 --nfilters=2 -x train_split1_Mediamill_data.txt --maxiter=1000000

# classify using an SVM model
numactl --interleave=all -- ../bin/mcproj --nProj={0,2} -f mediamill_C1_0.1_C2_0.1.filter -x test_split1_Mediamill_data.txt --modelFiles svm_C10_split1_Mediamil.svmmodel

# list all options
../bin/mcsolve --help
../bin/mcproj --help

File formats

Data file format

The data file uses the XML format.

First line is an optional header line:

nExamples nFeatures nClasses

Subsequently the file should contain one data point per line, in the following sparse format:

label1,label2,label3  feature:value feature:value ... feature:value  #comment

Labels are consecutive integers from 0 to nClasses-1. Features are consecutive integers from 0 to nFeatures-1.

The code also supports a binary format with both dense and sparse storage. Use convert_data to convert from text to binary format.

Label filter file format

The label filter file is composed of three matrices:

  1. The filter directions as a nFilters x nFeatures matrix
  2. The lower bounds as a nFilters x nClass matrix
  3. The upper bounds as a nFilters x nClass matrix

Each matrix starts with a header:

nRows nCols

followed by nRows rows of nCols space separated floats.

Linear model file format

The linear classifiers are stored in an SVMLight like format.

The first row in the file is an optional header:

nClasses nFeatures

Next the weights of each classifier are stored, one classifier per line, in the order of the labels (i.e. classifier corresponding to label '0' is first, then the classifier corresponding to label '1' and so on. The format is:

intercept feature:weight feature:weight ... feature:weight

Features are consecutive integers from 0 to nFeatures-1.

The intercept is optional. If not present, it is treated as 0.

The code also supports a binary format with both dense and sparse storage. Binary model files are smaller and much faster to load. Use convert_linearmodel to convert from text to binary format.

Important classes

  • MClearnFilter -- main class for training the label filters from data.

  • MClinearClass -- main class for testing using a linear model and optional label filters. Handles prediction and evaluation.

  • MCxyDataArgs -- defines parameters and command line options for handling the data

  • MCsolveArgs -- defines parameters and command line options for learning the label filter

  • MCprojectorArgs -- defines parameters and command line options for applying the label filters

  • MCclassifierArgs -- defines parameters and command line options for applying the classifier

  • MCxyData -- encapsulates the data for training/testing. Handles both dense and sparse data. Manages loading and saving data in various formats.

  • MCsoln -- Encapsulates the label filter parameters. Manages loading and saving label filters in text or binary formats.

  • PredictionSet -- encapsulates the predictions of the model.

  • linearModel -- encapsulates a linear classifier. Handles loading, saving and prediction (not training)


Label Filters for Extreme Classification







No releases published


No packages published