dobraczka / kiez Goto Github PK

A Python library for hubness reduced nearest neighbor search for the task of entity alignment with knowledge graph embeddings. The term kiez is a german word that refers to a city neighborhood.

Hubness is a phenomenon that arises in high-dimensional data and describes the fact that a couple of entities are nearest neighbors (NN) of many other entities, while a lot of entities are NN to no one. For entity alignment with knowledge graph embeddings we rely on NN search. Hubness therefore is detrimental to our matching results. This library is intended to make hubness reduction techniques available to data integration projects that rely on (knowledge graph) embeddings in their alignment process. Furthermore kiez incorporates several approximate nearest neighbor (ANN) libraries, to pair the speed advantage of approximate neighbor search with increased accuracy of hubness reduction.

You can install kiez via pip:

pip install kiez

If you have a GPU you can make kiez faster by installing faiss (if you do not already have it in your environment):

conda env create -n kiez-faiss python=3.10
conda activate kiez-faiss
conda install -c pytorch -c nvidia faiss-gpu=1.7.4 mkl=2021 blas=1.0=mkl
pip install kiez

For more information see their installation instructions.

You can also get other specific libraries with e.g.:

  pip install kiez[nmslib]

Simple nearest neighbor search for source entities in target space:

from kiez import Kiez
import numpy as np
# create example data
rng = np.random.RandomState(0)
source = rng.rand(100,50)
target = rng.rand(100,50)
# fit and get neighbors
k_inst = Kiez()
k_inst.fit(source, target)
nn_dist, nn_ind = k_inst.kneighbors(5)

Using (A)NN libraries and hubness reduction methods:

from kiez import Kiez
import numpy as np
# create example data
rng = np.random.RandomState(0)
source = rng.rand(100,50)
target = rng.rand(100,50)
# prepare algorithm and hubness reduction
k_inst = Kiez(n_candidates=10, algorithm="Faiss", hubness="CSLS")
# fit and get neighbors
k_inst.fit(source, target)
nn_dist, nn_ind = k_inst.kneighbors(5)

Beginning with version 0.5.0 torch can be used, when using Faiss as NN library:

    from kiez import Kiez
    import torch
    source = torch.randn((100,10))
    target = torch.randn((200,10))
    k_inst = Kiez(algorithm="Faiss", hubness="CSLS")
    k_inst.fit(source, target)
    nn_dist, nn_ind = k_inst.kneighbors()

You can also utilize tensor on the GPU:

    k_inst = Kiez(algorithm="Faiss", algorithm_kwargs={"use_gpu":True}, hubness="CSLS")
    k_inst.fit(source.cuda(), target.cuda())
    nn_dist, nn_ind = k_inst.kneighbors()

You can find more documentation on readthedocs

The results and configurations of our experiments can be found in a seperate benchmarking repository

If you find this work useful you can use the following citation:

@article{obraczka2022fast,
  title={Fast Hubness-Reduced Nearest Neighbor Search for Entity Alignment in Knowledge Graphs},
  author={Obraczka, Daniel and Rahm, Erhard},
  journal={SN Computer Science},
  volume={3},
  number={6},
  pages={1--19},
  year={2022},
  publisher={Springer},
  url={https://link.springer.com/article/10.1007/s42979-022-01417-1},
  doi={10.1007/s42979-022-01417-1},
}

PRs and enhancement ideas are always welcome. If you want to build kiez locally use:

git clone [email protected]:dobraczka/kiez.git
cd kiez
poetry install

To run the tests (given you are in the kiez folder):

poetry run pytest tests

Or install nox and run:

nox

which checks all the linting as well.

kiez is licensed under the terms of the BSD-3-Clause license. Several files were modified from scikit-hubness, distributed under the same license. The respective files contain the following tag instead of the full license text.

    SPDX-License-Identifier: BSD-3-Clause

This enables machine processing of license information based on the SPDX License Identifiers that are here available: https://spdx.org/licenses/