Container-Level Latency Prediction via Integrated Structure-Aware Graph Modeling and Multi-Scale Temporal Encoding

Abstract

This paper proposes a container-level latency prediction framework that integrates structure-aware graph modeling and multi-scale temporal modeling to address the challenges of frequent latency fluctuations and complex structural dependencies in containerized backend systems. The framework first introduces a structure-aware graph modeling module that dynamically constructs interaction dependency graphs among containers. It uses graph neural networks to capture structural information in key paths and asymmetric communication chains, enhancing the model's ability to represent causal clues in latency modeling. A multi-scale temporal encoding mechanism is then applied to extract short-term variations and long-term trends across multiple temporal granularities. A learnable convolutional kernel fusion strategy is used to improve robustness in modeling non-stationary latency sequences. During feature integration, structural graph embeddings and multi-scale temporal representations are jointly mapped into the latency prediction space. A unified optimization loss function enforces both cross-scale consistency and structural preservation constraints. To further validate the effectiveness of the modeling components, two sets of key sub-experiments are designed to evaluate the impact of multi-scale fusion strategies and time window sliding steps on latency prediction performance. Experiments conducted on real-world container monitoring datasets show that the proposed model significantly outperforms existing methods across multiple evaluation metrics, demonstrating strong generalization and sensitivity to structural dependencies.