SSSE

Scalable Semi-supervised clustering via structural entropy with different constraints.

The SSSE framework. We illustrate SSSE on a small dataset. (I) Two graphs G and G′ are constructed from input data and prior information in the formats of pairwise constraints and label constraints, respectively. (II) Semi-supervised clustering is performed on the sampled graphs. (III) Semi-supervised clustering is performed on the remaining graphs, in which obtained clusters are taken as new data points. For semi-supervised partitioning clustering, clustering in (II) and (III) is achieved by two operators merging and moving, while for semi-supervised hierarchical clustering, clustering is achieved by stretching and compressing.

Installation

Install the required packages listed in the file requirement.txt. The code is tested on Python 3.10.0.

Datasets used are in the data branch.

Usage

In the root directory of this project:

python main.py [-h][--method METHOD][--dataset DATASET]
               [--constraint_ratio RATIO][--constraint_weight WEIGHT]
               [--sigmasq SIGMASQ][--exp_repeats REPEATS]
               [--knn_constant KNN_CONSTANT][--hie_knn_k HIE_KNN_K]
               [--sampling SAMPLING][--sampling_size SAMPLING_SIZE]

example: python main.py --method SSSE_partitioning_pairwise --dataset Yale --constraint_ratio 0.2

required arguments:
  --method METHOD    running different components of SSE. Choices are SSSE_partitioning_pairwise, SSSE_partitioning_bio, and so on.
  --dataset DATASET    dataset to run. They should be stored in directory ./datasets.
  --constraint_ratio   constraint ratio. Recommend setting 0.2 for pairwise constraints and 0.1 for label constraints.

optional arguments:
  --constraint_weight     weight for penalty term. (default 1).
  --sigmasq SIGMASQ       square of Gaussian kernel band width, i.e., sigma^2.
  --exp_repeats REPEATS   number of experiment repeats. (default 10).
  --knn_constant          a constant for graph construction in partitioning clustering. (default 20).
  --hie_knn_k             a constant for graph construction in hierarchical clustering. (default 5).
  --sampling              sampling with neighborhood-preserving graph sampling strategy (neighbor) or random sampling (random). (default neighbor).
  --sampling_size         sampling size for graph sampling-based scalability strategy. (default 1000).