kgraph | A library for k-nearest neighbor search | Machine Learning library

KGraph depends on a recent version of GCC with C++11 support, cmake and the Boost library. The package can be built and installed with. A Makefile.plain is also provided in case cmake is not available. The Python API can be installed with.
Both index.build and index.search supports a number of optional keywords arguments to fine tune the performance. The default values should work reasonably well for many datasets. One exception is that reverse=-1 should be added if the purpose of building index is to speedup search, which is the typical case, rather than to obtain the k-NN graph itself. Two precautions should be taken: * Although matrices of both float32 and float64 are supported, the latter is not optimized. It is recommened that matrices be converted to float32 before being passed into kgraph. * The dimension (columns of matrices) should be a multiple of 4. If not, zeros must be padded. For performance considerations, the Python API does not support user-defined similarity function, as the callback function is invoked in such a high frequency that, if written in Python, speedup will inevitably be brought down. For the full generality, the C++ API should be used.
The KGraph C++ API is based on two central concepts: the index oracle and the search oracle. (Oracle is a fancy way of calling a user-defined callback function that behaves like a black box.) KGraph works solely with object IDs from 0 to N-1, and relies on the oracles to map the IDs to actual data objects and to compute the similarity. To use KGraph, the user has to extend the following two abstract classes. The similarity values computed by the oracles must satisfy the following two conditions: * The more similar the objects are, the smaller the similarity value (0.1 < 10, -10 < 1). * Similarity must be symmetric, i.e. f(a, b) = f(b, a). KGraph’s heuristic algorithm does not make assumption about properties such as triangle-inequality. If the similarity is ill-defined, the worst it can do is to lower the accuracy and to slow down computation.