A C++ and (optional) Python tool for partitional clustering. Optimised implementations of K-Medoids and K-Means, for various data types. More information is in our paper at arXiv 1609.04723.
Given N elements x(1)...x(N), select K elements indexed by c(1)...c(K), to minimise sum(i=1...N) min(k=1...K) E(distance (x(i), x(c(k)))) where distance is a valid distance and E is a non-decreasing function with E(0) = 0.
distance options are
- for sparse and dense vectors : l-0, l-1, l-2, l-infinity
- for sequence data : Levenshtein and Normalised Levenshtein.
Energy E options are
- identity, quadratic, cubic, square-potential, exponential, and logarithmic.
- minimise sum of squares of l2 distances to cluster mean
- minimise sum of l1 distances to cluster dimension-wise median
- CMake
- for the Python library: Cython and Python
Create a build directory:
mkdir build; cd build;
If you do NOT want the Python library,
cmake -DBUILD_PYTHON_LIB=NO ..
If you do want the Python library,
cmake ..
The library can be built, from the build directory
make -j5
The shared library should now be in ./build/zentas (libzentas.so in Linux) and the Python shared library in ./build/python (pyzentas.so in Linux). These can be moved/copied elsewhere manually, there is currently no install option for zentas.
Example use cases of the C++ library and headers are in testsexamples, with the corresponding executables in build/testsexamples. There is an example of clustering dense vectors (exdense.cpp), sparse vectors (exsparse.cpp), and sequences (exwords.cpp).
To use the Python library, make sure pyzentas.so is on PYTHONPATH, for example you can use sys.path.append(/path/to/pyzentas.so). Examples using pyzentas are in python/examples.py. More information can be obtained from the doc strings, try
import pyzentas
help(pyzentas)
Please raise an issue in the zentas repository