Decentralized Personalized Online Federated Learning

Renzhi Wu, Saayan Mitra, Xiang Chen, Anup B. Rao

Introduction

In the evolving landscape of distributed machine learning, Federated Learning (FL) has emerged as a cornerstone for training models without the need to centralize sensitive user data. However, traditional "vanilla" FL frameworks often struggle with the practical constraints of enterprise edge computing. In these environments, data arrives in continuous streams, regional preferences vary significantly, and central servers represent both a performance bottleneck and a regulatory risk.

To address these challenges, researchers from Georgia Tech and Adobe Research have introduced AlphaEdge, a framework designed to navigate the "trifecta" of modern edge requirements: decentralization, personalization, and online learning. This paper formalizes the Decentralized Personalized Online Federated Learning (DP-OFL) paradigm, offering a robust solution for non-convex optimization problems in dynamic network environments.

The DP-OFL Paradigm

Standard Federated Learning typically operates under three assumptions that are frequently violated in real-world enterprise settings:

  1. Static Data (Offline Training): Traditional FL assumes data is static and training occurs in discrete rounds. In reality, edge devices generate a continuous stream of data (e.g., clickstreams or sensor readings) that requires real-time adaptation.
  2. Global Consensus: Vanilla FL seeks to converge on a single global model. However, a recommendation engine optimized for one demographic may perform poorly for another due to cultural or regional heterogeneity.
  3. Centralization: Reliance on a central parameter server creates a single point of failure and complicates compliance with data sovereignty laws (such as GDPR or CCPA), which may prohibit regional servers from communicating with a central hub in a different jurisdiction.

AlphaEdge integrates these three dimensions into a single framework, allowing edge servers to maintain high-performing, personalized models that adapt to real-time data shifts without a central coordinator.

Methodology: The AlphaEdge Framework

AlphaEdge introduces two primary innovations to manage the complexities of DP-OFL: Learned Aggregation and Greedy Peer Selection.

1. Learned Aggregation

In traditional FL, models are typically averaged using fixed rules, such as weighting by local data volume. AlphaEdge transforms this into a dynamic optimization problem. When an edge server receives model updates from its neighbors, it utilizes its current batch of online data to evaluate which neighbor's model provides the most utility.

By applying gradient descent to learn the optimal aggregation weights ($\alpha$), the system automatically prioritizes neighbors with similar data distributions and ignores those providing "noisy" or irrelevant updates. This mechanism ensures that the resulting model is inherently personalized to the local node's specific data context.

2. Greedy Peer Selection

Maintaining a fully connected network is computationally and communicationally expensive, scaling at $O(n^2)$. To ensure scalability, AlphaEdge limits each node to communicating with a subset of $K$ neighbors. The framework employs a "greedy" strategy to optimize this neighbor list over time. Each node monitors its "two-hop" neighbors (the neighbors of its neighbors) and calculates their potential contribution based on existing weights. If a current neighbor provides diminishing returns, it is swapped for a more promising candidate from the two-hop pool, allowing the network topology to evolve alongside the models.

Experimental Evaluation

The researchers evaluated AlphaEdge across four diverse datasets to demonstrate its versatility:

  • E-commerce Datasets (A & B): Two large-scale private datasets (40M and 20M records) involving user-item interactions (views vs. purchases).
  • Movielens-1m: A classic movie rating dataset adapted into a Click-Through Rate (CTR) task, with injected label noise to simulate regional data shifts.
  • AirQuality: A regression dataset from 35 weather stations in Beijing, used to predict PM2.5 levels, representing a non-classification edge use case.

Key Results and Insights

AlphaEdge consistently outperformed established baselines, including NoAgg, FedWeight, and UniWeight, across all test scenarios. Key findings include:

  • Superior Accuracy: AlphaEdge achieved the highest AUC scores in recommendation tasks and significantly outperformed other decentralized methods in the AirQuality regression task.
  • Robustness to Network Volatility: The system demonstrated resilience against feedback delays and network instability. In stress tests where 25% of nodes were randomly taken offline, the framework recovered quickly without significant performance degradation.
  • Adversarial Defense: By learning specific weights for neighbors, AlphaEdge provides an inherent defense against malicious actors. When 25% of nodes were configured as "malicious" (flipping labels), the learned weights for those nodes naturally dropped to zero, whereas standard FL methods experienced catastrophic performance loss.

Conclusion

AlphaEdge represents a significant step forward for edge intelligence by successfully merging decentralization, personalization, and online learning. By utilizing the immediate feedback loop of online learning to "train the network topology," the framework ensures that edge models remain relevant and robust.

While AlphaEdge addresses structural challenges, future work remains in the realm of privacy. Sharing model parameters can still lead to indirect data leakage. Integrating Differential Privacy (DP) or Secure Multi-Party Computation (SMPC) into the AlphaEdge framework, alongside optimizing for physical network latency in P2P meshes, are the logical next steps for production-ready deployment.

Summary updated

Abstract

Vanilla federated learning does not support learning in an online environment, learning a personalized model on each client, and learning in a decentralized setting. There are existing methods extending federated learning in each of the three aspects. However, some important applications on enterprise edge servers (e.g. online item recommendation at global scale) involve the three aspects at the same time. Therefore, we propose a new learning setting Decentralized Personalized Online Federated Learning that considers all the three aspects at the same time.In this new setting for learning, the first technical challenge is how to aggregate the shared model parameters from neighboring clients to obtain a personalized local model with good performance on each client. We propose to directly learn an aggregation by optimizing the performance of the local model with respect to the aggregation weights. This not only improves personalization of each local model but also helps the local model adapting to potential data shift by intelligently incorporating the right amount of information from its neighbors. The second challenge is how to select the neighbors for each client. We propose a peer selection method based on the learned aggregation weights enabling each client to select the most helpful neighbors and reduce communication cost at the same time. We verify the effectiveness and robustness of our proposed method on three real-world item recommendation datasets and one air quality prediction dataset.
BibTeX
@inproceedings{wu2023decentralized,
  author = {Renzhi Wu and Saayan Mitra and Xiang Chen and Anup B. Rao},
  title = {{Decentralized Personalized Online Federated Learning}},
  booktitle = {IEEE Big Data},
  year = {2023},
  url = {https://openreview.net/forum?id=1fsZxZaemj},
}
Figure