PHIDA

(c) 2026 Naoki Masuyama

PHIDA is an online clustering method that first creates IDA-style nodes and then reads out final clusters from node-level 0-dimensional persistent homology.

Requirements

• Python 3.10 or later
• numpy
• scikit-learn
• networkx
• joblib

Install the required packages with your preferred Python environment, for example:

pip install numpy scikit-learn networkx joblib

Main Files

• mainPHIDA.py: example runner for stationary and nonstationary experiments.
• phida.py: core PHIDA model.
• ph_view_builder_phida.py: persistent-homology graph construction utilities.
• continual_metrics.py: average incremental and backward-transfer metrics.

Data

mainPHIDA.py downloads datasets from OpenML through sklearn.datasets.fetch_openml. The default example uses the Iris dataset (OpenML data_id=61). OptDigits (OpenML data_id=28) is also listed in the script as a copy-paste dataset option.

To use another OpenML dataset, change DATASET_NAME in mainPHIDA.py. If the dataset has a known stable OpenML data_id, add it to OPENML_DATASETS; otherwise the runner tries to fetch by name. The example runner expects numeric features and classification labels.

Run

python mainPHIDA.py

Set SETTING = "stationary" to randomly permute all samples and give them to PHIDA in one fit call. Set SETTING = "nonstationary" to split samples by class and give them class-by-class in random class order.

The script prints the number of nodes, number of clusters, runtime, ARI, and AMI. In the nonstationary setting it also prints average incremental scores and backward transfer.

Citation

If you use this code, please cite:

N. Masuyama, Y. Nojima, S. Wermter, Y. Toda, H. Ishibuchi, and C. K. Loo, "PHIDA: Persistence-Guided Node-to-Cluster Mapping for Online Clustering," arXiv preprint arXiv:2605.08673 [cs.LG], 2026.

License

This project is licensed under the MIT License. See LICENSE for details.