Communication-Efficient Machine Learning Architecture for Predicting User Information Needs in Distributed Systems

Abstract

Overwhelming communication overhead, long synchronisation times, and scaling issues are the major issues that machine learning (ML) models deployed in distributed computing settings have to overcome. These are problems that hamper the effectiveness and responsiveness of the intelligent systems, especially in large-scale information networks whereby the user data are spatially distributed. To address these drawbacks, this research suggests a communication-efficient federated learning (FL) architecture optimised in predicting user information requirements to distributed data settings. The architecture will be based on the localised training of models in individual nodes for example institutional repository or digital library servers thus eliminating the necessity of transferring raw data to the central station and adhering to the provisions of privacy. An adaptive communication scheme hierarchically structured into aggregation mechanism is widely applicable in terms of bandwidth consumption minimization and does not undermine performance of the model convergence and predictive accuracy. The results of experimental validation in an experimental topology comprised of five interconnected nodes confirm a 42 percent saving of communication overhead and 15 percent enhancement of training efficiency over traditional centralised learning systems. In addition, the proposed architecture supports scalability of the system, power efficiency and compliance to privacy, which forms a formidable base on big-scale, smart data infrastructures. The study points out the promise of communication-optimized federated learning as one of the enabling factors of secure, adaptive, and resource-sensitive distributed machine learning ecosystems.