Personalized Federated Learning Models for Data Heterogeneity: A Systematic Literature Review and Comparative Taxonomy

Federated learning (FL) enables multiple clients to collaboratively train a shared model without exchanging raw data, but its performance degrades sharply when client data are non-independent and identically distributed (non-IID). Personalized federated learning (PFL) has emerged as the principal remedy, replacing the single global model objective with client-specific models that retain the privacy and communication advantages of FL while adapting to local data distributions.
This paper presents a systematic literature review and comparative taxonomy of PFL techniques published between 2017 and 2026. Following a structured search and screening protocol across major digital libraries, we synthesize the literature into six methodological families —data-based,
model-based (parameter-decoupling), similarity/clustering-based, meta learning and optimization-based, knowledge-distillation and transfer-based, and mixture/hypernetwork-based approaches. For each family we describe the underlying mechanism, representative algorithms, and the specific type of heterogeneity it is best suited to address. We further provide a comparative analysis of representative algorithms across personalization degree, communication efficiency,
computational overhead, robustness, and scalability, and we catalog the benchmark datasets and non-IID partitioning protocols used to evaluate them. The review concludes by identifying open challenges and by outlining promising directions for future research.