$p$-value Adjustment for Monotonous, Unbiased, and Fast Clustering Comparison
Authors: Kai Klede, Thomas Altstidl, Dario Zanca, Bjoern Eskofier
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate its unbiased clustering selection, approximation quality, and runtime efficiency on synthetic benchmarks. In experiments on image and social network datasets, we show how the PMI2 can help practitioners choose better clustering and community detection algorithms. |
| Researcher Affiliation | Academia | 1Machine Learning and Data Analytics (Ma D) Lab Friedrich-Alexander Universität Erlangen-Nürnberg 2Translational Digital Health Group Institute of AI for Health, Helmholtz Zentrum München |
| Pseudocode | No | No pseudocode or algorithm blocks are present in the paper. |
| Open Source Code | Yes | The code for all experiments is available at https://github.com/mad-lab-fau/pmi-experiments. |
| Open Datasets | Yes | k-means clustering [21] with varying k to a) the UCI handwritten digit dataset [6] and b) the Olivetti faces dataset [27] and select the solution with the highest similarity to the ground truth. We simulate this procedure on a network of email conversations between European research institutions, where each institution is a ground truth cluster [20]. |
| Dataset Splits | No | For several numbers of clusters k, we apply k-means clustering [21] with 1000 different random seeds. We repeat this experiment 1000 times with different random seeds and plot the selection probability under the RI, AMI2, and normal approximation of the PMI2. This process is repeated for 100 subsets per dataset, and the resulting probabilities are shown in Figure 4c. |
| Hardware Specification | Yes | The experiments were executed on an AMD Ryzen 9 5950X. |
| Software Dependencies | No | We compared five community detection algorithms implemented in networkit [32]. |
| Experiment Setup | Yes | For several numbers of clusters k, we apply k-means clustering [21] with 1000 different random seeds. We then select the most similar algorithm to the ground truth using RI, AMI2, and PMI2 Φ(SMI2). This process is repeated for 100 subsets per dataset, and the resulting probabilities are shown in Figure 4c. Table 4: Parameter configurations for the experiment on the email EU core dataset. For Leiden, we used all combinations of γ and randomize. |