COMFORTage at the International Conference on Service-Oriented Computing 20th International Workshop on Engineering Service-Oriented Applications and Cloud Services.
Nemania Borovits representing JADS presented Enhancing Federated Learning with SOA: An Approach to Tackle Non-IID Data Challenges.
In science, solving a problem often involves drawing insights from a conceptually related domain. We turned to Service-Oriented Architecture (SOA) literature to address the non-IID data challenge in Federated Learning (FL). Inspired by SOA’s aggregator pattern, we proposed in-client clustering, which organizes local data into homogeneous subsets, reducing variability in updates sent to the global model and improving overall model convergence.
Testing our approach on a Named Entity Recognition (NER) task using two medical datasets (ADE and CADEC) and the CoNLL dataset for generalizability, we observed an improvement of 3% on average in the weighted F1 score. While this may seem modest, it roughly translates to more accurate predictions for 1,000 patients and 12,900 more accurately classified tokens, based on the size of the datasets we used, which is a notable achievement for applications in domains like healthcare.The real-world impact of this research will be evaluated within the context of the COMFORTage | Horizon EU Project. The project aims to develop an AI platform to assist medical professionals and patients in managing frailty and dementia, with the support of 13 clinical partners across 8 EU member states providing the necessary clinical context to ensure its relevance and effectiveness in addressing these medical challenges.
More info:
Do you want to see pictures and find out everything about what happened that day? Click on the following links:
https://x.com/comfortage_eu
https://www.facebook.com/COMFORTage
https://www.instagram.com/comfortage_eu/
https://www.linkedin.com/company/comfortage/
