Machine-Learning-Based Selection of Data Serialisation Formats for Web Applications Under Variable Network Conditions

Abstract

The heterogeneity of modern network environments requires adaptive mechanisms for improving data transmission efficiency under changing bandwidth, latency, connection stability, and computational constraints. This study aimed to develop an intelligent system for selecting the optimal data serialisation format according to current network conditions. The proposed architecture combined real-time network monitoring, a multicriteria utility function, and LSTM, Random Forest, and Deep Q-Learning algorithms. Six formats – JSON, XML, Protocol Buffers, Avro, MessagePack, and BSON – were evaluated in 3G, 4G LTE, 5G, fast and slow Wi-Fi, and edge-computing environments. The models were trained using approximately 4.8 million telemetry and event-log records, and each experimental scenario was repeated 30 times. LSTM achieved 94% prediction accuracy, while Random Forest reached 95% with comparatively low computational requirements. Deep Q-Learning provided rapid adaptation and an accuracy of 92-95%. Compared with static format selection, the proposed system reduced latency by 35% and improved transmission efficiency by up to 87.5% under bandwidth-constrained conditions. The average internal response time was approximately 45 ms. Protocol Buffers demonstrated the most stable performance for large data volumes and constrained networks, whereas MessagePack provided a balanced solution for small and medium payloads. The findings confirm that context-sensitive format selection can improve transmission performance, resource utilisation, and system reliability.