ClusterFuG: Clustering Fully connected Graphs by Multicut (arxiv)

Solve multicut problem on complete graph alleviating need for graph structure design. Similarity between a pair of nodes is computed by inner product of input node features.

Requirements

We use GCC 10. Other combinations might also work but not tested. CMake is required for compilation.

Installation

C++ solver:

git clone [email protected]:aabbas90/cluster-fug.git
cd cluster-fug
git submodule update --init --recursive
mkdir build
cd build
cmake ..
make -j 4

Python bindings:

We also provide python bindings using pybind. Simply run the following command:

python -m pip install git+https://github.com/aabbas90/cluster-fug.git

Usage

C++ solver:

We require multicut instance stored in a (.txt) file in the following format:

FACTORIZED COMPLETE MULTICUT
a
n d
feature_node_1_dim_1 ... feature_node_1_dim_d
...
feature_node_n_dim_1 ... feature_node_n_dim_d

which corresponds to a graph with n nodes, and d-dimensional node features. The value of a controls affinity strength where increasing the value helps in creating more clusters and vice versa. Afterwards run (for our DAppLAEC algorithm):

./src/dense_multicut_text_input <PATH_TO_PROBLEM_INSTANCE_TXT> HNSW laec_bf_later

Other algorithms mentioned in the paper can be run as:

./src/dense_multicut_text_input <PATH_TO_PROBLEM_INSTANCE_TXT> <INDEX_TYPE> <CONTRACTION_TYPE>

where INDEX_TYPE and CONTRACTION_TYPE can be choosen as:

Algorithm	INDEX_TYPE	CONTRACTION_TYPE
GAEC	brute_force	gaec
DGAEC	faiss_brute_force	dense_gaec
DGAECInc	faiss_brute_force	dense_gaec_inc
DLAEC	faiss_brute_force	dense_laec
DAppLAEC	HNSW	dense_laec_bf_later

For example to run GAEC: bash ./src/dense_multicut_text_input <PATH_TO_PROBLEM_INSTANCE_TXT> brute_force gaec For more information run:

./src/dense_multicut_text_input --help
	```

### Python solver:
An example to compute multicut on a set of features:
```python
import dense_multicut_py
import numpy as np

num_nodes = 100
dim = 16
affinity_strength = 5.0
k = 5
k_cap = k
# DAppLAEC:
INDEX_TYPE, CONTRACTION_TYPE = "HNSW", "dense_laec_bf_later"
features = np.random.rand(num_nodes, dim).astype(np.float32)
node_labels = dense_multicut_py.dense_multicut(features.flatten(), num_nodes, dim, affinity_strength, k, INDEX_TYPE, CONTRACTION_TYPE, k_cap)

Benchmark instances:

Instances used in the paper can be obtained from structured-prediction-prob-archive.

References

If you use this work please cite as

@article{abbas2023clusterfug,
  title={ClusterFuG: Clustering Fully connected Graphs by Multicut},
  author={Abbas, Ahmed and Swoboda, Paul},
  journal={arXiv preprint arXiv:2301.12159},
  year={2023}
}