In personalised federated learning, the server must decide which clients are similar enough to share a model. Most current methods derive this signal from training losses or model updates, which means the grouping has to be recomputed during training. The new paper, posted on arXiv, proposes PTOPOFL, which instead builds client descriptors using persistent homology, a topological technique that summarises shape-like structure in data across scales.
Because these descriptors come from the data itself rather than from model behaviour, they may offer a more stable basis for clustering clients. The server can use the topology-guided descriptors to decide which clients should share a model without relying on training dynamics.
The abstract is truncated mid-sentence, so the full method and experimental results are not available in the source. No other sources were provided, so there is no independent result to compare against.