Refinement of Clustering (CluReAL.v2)

TU Wien, CN group

FIV, May 2021

Dependencies

Check python dependencies (requirements.txt).

    $ python3 install-dep.py

CluReAL.v2

CluReAL refinement and SK ideograms quick example.

    $ python3 toy_example.py

Material for experiment replicability

Experiments consist on comparing the performance of (a) the best clustering obtained from parameter search around ideal parameters (according to the GT), with (b) CluReAL refining a suboptimal clustering obtained with arbitrary parameters.

Datasets

Datasets used in the experiments that were not generated with MDCGen (https://github.com/CN-TU/mdcgen-matlab) must be downloaded from the author's website:

http://cs.uef.fi/sipu/datasets/

We provide here only the corresponding GT (labels). Please download the datasets and prepare the files as comma-separated-values without headers, e.g., (for a 2D-dataset):

    53920,42968,8
    52019,42206,8
    52570,42476,8
    ...

Later, place them in the corresponding folder before running the scripts. In the following table missing datasets are listed with their corresponding folder, GT_file name, expected final name and the direct link to the original source.

| Folder ------- | GT file name ------- | Final name ------- | Link to the orig. source ------------------------- | |-:-:-------- |-:-:---------------- | -:-:------------ | -:-:--------------------------------------------| | data2d/ | a2_GT.csv | a2.csv | http://cs.uef.fi/sipu/datasets/a2.txt | | data2d/ | a3_GT.csv | a3.csv | http://cs.uef.fi/sipu/datasets/a3.txt | | data2d/ | s1_GT.csv | s1.csv | http://cs.uef.fi/sipu/datasets/s1.txt | | data2d/ | s2_GT.csv | s2.csv | http://cs.uef.fi/sipu/datasets/s2.txt | | data2d/ | s3_GT.csv | s3.csv | http://cs.uef.fi/sipu/datasets/s3.txt | | data2d/ | unbalance_GT.csv | unbalance.csv | http://cs.uef.fi/sipu/datasets/unbalance.txt | | dataMd/ | multidim_0002_GT | multidim_0002 | http://cs.uef.fi/sipu/datasets/data_dim_txt.zip | | dataMd/ | multidim_0003_GT | multidim_0003 | http://cs.uef.fi/sipu/datasets/data_dim_txt.zip | | dataMd/ | multidim_0005_GT | multidim_0005 | http://cs.uef.fi/sipu/datasets/data_dim_txt.zip | | dataMd/ | multidim_0010_GT | multidim_0010 | http://cs.uef.fi/sipu/datasets/data_dim_txt.zip | | dataMd/ | multidim_0015_GT | multidim_0015 | http://cs.uef.fi/sipu/datasets/data_dim_txt.zip | | dataMd/ | multidim_0032_GT | multidim_0032 | http://cs.uef.fi/sipu/datasets/dim032.txt | | dataMd/ | multidim_0064_GT | multidim_0064 | http://cs.uef.fi/sipu/datasets/dim064.txt | | dataMd/ | multidim_0256_GT | multidim_0256 | http://cs.uef.fi/sipu/datasets/dim256.txt | | dataMd/ | multidim_0512_GT | multidim_0512 | http://cs.uef.fi/sipu/datasets/dim512.txt | | dataMd/ | multidim_1024_GT | multidim_1024 | http://cs.uef.fi/sipu/datasets/dim1024.txt |

Alternatively, you can remove these datasets from the scripts.

Finally, real datasets (i.e., real_1, real_2, real_3, real_4) are obtained from the scikit.learn package and transformed with tSNE. The script to extract such datasets are in the [extra] folder:

    $ python3 extract_real.py 

Folders

Create two folders called [plots] and [results] before running experiments.

    $ mkdir plots 

    $ mkdir results 

Experiments

• CluReAL vs k-Sweep (2D, k-means). Comparison of clustering optimization methods with 2D-data and k-means algorithm. Plots are saved in the [plots] folder.

    $ python3 2d_comparison_kmk.py
  

• CluReAL vs k-Sweep (multiD, partitional clust.). Comparison of clustering optimization methods with Multi-dimensional data and partitional algorithms. Results are saved in the [results] folder.

    $ python3 Md_comparison_k.py
  

• CluReAL vs Random Search (multiD, density-based clust.). Comparison of clustering optimization methods with Multi-dimensional data and density-based algorithms. Results are saved in the [results] folder.

    $ python3 Md_comparison_n.py
  

• CluReAL.v1 vs CluReAL.v2 (multiD, k-means). Comparison of CluReAL.v1 and CluReAL.v2 for refining suboptimal k-means clustering with Multi-dimensional data. Results are saved in the [results] folder.

    $ python3 clureal_v1vs2_comparison.py
  

• Sensitivity analysis (number of samples), CluReAL vs k-Sweep. Comparison of runtime requirements of clustering optimization methods with Multi-dimensional datasets of different sizes (n={500,1000,2500,5000,10000,25000}) and partitional algorithms. Results are saved in the [results] folder.

    $ python3 runtime_comparison_k.py
  

• Sensitivity analysis (number of samples), CluReAL vs Random Search. Comparison of runtime requirements of clustering optimization methods with Multi-dimensional datasets of different sizes (n={500,1000,2500,5000,10000,25000}) and density-based algorithms. Results are saved in the [results] folder.

    $ python3 runtime_comparison_n.py
  

• CluReAL options vs high-overlap. Study of CluReAL alternatives to deal with high overlap: (a) more strict kinship-based edge-pruning during refinement, (b) using coresets. Plots are saved in the [plots] folder.

    $ python3 high-overlap.py