Cold-Start Microservice Workload Prediction by Dynamically Annealed Graph-Regularized Matrix Factorization

In the cold-start phase of microservice systems, workload prediction is highly challenging due to undeveloped global dependencies and sparse, noisy observations. Existing methods often rely on long historical sequences or global modeling, making it difficult to capture short-term yet critical transient couplings.To address this, we propose DAGRMF (Dynamically Annealed Graph-Regularized Matrix Factorization), a two-stage framework for cold-start prediction that unifies global structure recovery with local dependency refinement. The model decomposes the microservice workload matrix into service- and temporal-feature matrices. In Stage I, it extracts low-rank global skeletons via nuclear norm regularization while modeling multi-scale temporal trends through autoregressive constraints. Upon convergence, Stage II activates dynamic graph regularization to reinforce transient local couplings using online observations, with regularization weights smoothly transitioned via a curriculum annealing mechanism. Experiments on synthetic and real-world Alibaba microservice workload data demonstrate that DAGRMF is both robust and accurate in noisy environments and long-term predictions, achieving up to 21.3% lower RMSE than existing matrix factorization methods.