search

cmdevries / LMW-tree

Learning M-Way Tree - Web Scale Clustering - EM-tree, K-tree, k-means, TSVQ, repeated k-means, bitwise clustering
http://lmwtree.devries.ninja
BSD 3-Clause "New" or "Revised" License
75 stars 20 forks source link
bit-vectors clustering-algorithm hierarchical-clustering k-means-clustering large-scale-learning locality-sensitive-hashing random-indexing random-projections streaming-algorithms

readme

LMW-tree: learning m-way tree

See the project homepage for the latest news!

LMW-tree is a generic template library written in C++ that implements several algorithms that use the m-way nearest neighbor tree structure to store their data. See the related PhD thesis for more details on m-way nn trees. The algorithms and data structures are generic to support different data representations such as dense real valued and bit vectors, and sparse vectors. Additionally, it can index any object type that can form a prototype representation of a set of objects.

The algorithms are primarily focussed on computationally efficient clustering. Clustering is an unsupervised machine learning process that finds interesting patterns in data. It places similar items into clusters and dissimilar items into different clusters. The data structures and algorithms can also be used for nearest neighbor search, supervised learning and other machine learning applications.

The package includes EM-tree, K-tree, k-means, TSVQ, repeated k-means, clustering, random projections, random indexing, hashing, bit signatures. See the related PhD thesis for more details these algorithms and representations.

LMW-tree is licensed under the BSD license.

See the ClueWeb09 clusters and the ClueWeb12 clusters for examples of clusters produced by the EM-tree algorithm. The ClueWeb09 dataset contains 500 million web pages and was clustered into 700,000 clusters. The ClueWeb12 datasets contains 733 million web pages and was clustered into 600,000 clusters. The document to cluster mappings and other related files area available at SourceForge.

The following people have contributed to the project (sorted lexicographically by last name)

Directory Structure

The LMW-tree project uses several external libraries. It also has several modules contained in namespaces.

Currently we use:

Directory structure:

/

/src - all source contributed by this project where each subdirectory
       is a namespace
/src/lmw - LMW-tree data structures and algorithms
/src/indexer - english language document indexing

/external - all source for external 3rd party libraries
/external/Makefile - GNU Makefile to build external libraries
/external/packages - source packages for external libraries
/external/build - build directory for external libraries
/external/install - installation directory for external libraries

Building and Running

Make dependencies using a GNU Makefile (only tested on Linux)

$ cd external
$ make
$ cd ..

We use CMake for making the main project

$ mkdir build
$ cd build
$ cmake ..
$ make
$ cd ..

Fetch some data to cluster

$ mkdir data
$ cd data
$ wget http://downloads.sourceforge.net/project/ktree/docclust_ir/inex_xml_mining_subset_2010.txt
$ wget http://downloads.sourceforge.net/project/ktree/docclust_ir/wikisignatures.tar.gz
$ tar xzf wikisignatures.tar.gz
$ cd ..

Run the program

$ LD_LIBRARY_PATH=./external/install/lib ./build/emtree