Article: Personalised news recommendation via dynamic-threshold federated reinforcement learning Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.53 pp.88 - 108 Abstract: News interests shift quickly, and collecting fine-grained reading logs in one place is increasingly risky, so privacy-preserving personalisation must handle heterogeneous clients and unstable feedback. This paper proposes a dynamic-threshold federated reinforcement learning scheme for personalised news delivery. In the scheme, first, each device learns a sequential policy from local interactions to optimise long-horizon utility. Then, each round estimates update reliability and adjusts a participation cutoff to filter noisy client contributions. Finally, the server aggregates selected shared updates while keeping lightweight personalisation on device. Experimental results show that the proposed scheme raises NDCG at ten from 0.401 to 0.423, improves diversity from 0.287 to 0.319, increases cumulative reward from 1.866 to 2.034, and reduces communication per round from 10.9 to 7.8 megabytes, achieving a stronger balance of utility, diversity, and efficiency. Inderscience Publishers - linking academia, business and industry through research

Title: Personalised news recommendation via dynamic-threshold federated reinforcement learning

Authors: Ya Liu; Yingxi Zhang

Addresses: College of Literature, Journalism and Communication, Sanjiang University, Nanjing 210000, China; Cheongju University, Cheongju 363170, South Korea ' Cheongju University, Cheongju 363170, Korea

Abstract: News interests shift quickly, and collecting fine-grained reading logs in one place is increasingly risky, so privacy-preserving personalisation must handle heterogeneous clients and unstable feedback. This paper proposes a dynamic-threshold federated reinforcement learning scheme for personalised news delivery. In the scheme, first, each device learns a sequential policy from local interactions to optimise long-horizon utility. Then, each round estimates update reliability and adjusts a participation cutoff to filter noisy client contributions. Finally, the server aggregates selected shared updates while keeping lightweight personalisation on device. Experimental results show that the proposed scheme raises NDCG at ten from 0.401 to 0.423, improves diversity from 0.287 to 0.319, increases cumulative reward from 1.866 to 2.034, and reduces communication per round from 10.9 to 7.8 megabytes, achieving a stronger balance of utility, diversity, and efficiency.

Keywords: dynamic threshold; federated reinforcement learning; personalised news recommendation; client heterogeneity; communication efficiency.

DOI: 10.1504/IJICT.2026.153716

International Journal of Information and Communication Technology, 2026 Vol.27 No.53, pp.88 - 108

Received: 13 Feb 2026
Accepted: 25 Mar 2026
Published online: 21 May 2026 *

Title: Personalised news recommendation via dynamic-threshold federated reinforcement learning

Keep up-to-date