Research map / Graph learning / Knowledge and temporal graphs
Spatial-Temporal Heterogenous Graph Contrastive Learning for Microservice Workload Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STEAM, a spatio-temporal heterogeneous graph contrastive learning model that combines non-stationary decomposition self-attention, system-state encoding, and contrastive spatial encoding over similarity graphs to predict microservice CPU-core workloads.
Problem
Predicting microservice CPU workloads is hard because workload series are non-stationary, multi-periodic, and affected by dynamic system states, and microservices lack physical topology, so similarity graphs built with fixed thresholds introduce prior interference.
Approach
Builds a workload similarity graph as backbone, creates two heterogeneous augmentation views by adding response-time or request similarity edges, and applies cross-view graph contrastive learning to extract robust spatial features, combined with de-stationary decomposition self-attention for long-term temporal patterns and CNN for short-term patterns.
Key results
- On the MS dataset with input length L=48, STEAM achieves MAE 0.0738 and RMSE 0.1013, the best among all baselines in Table 1.
- On the Ali dataset with L=48, STEAM reaches MAE 0.1348 and RMSE 0.1747, outperforming PatchTST (MAE 0.1612) and MTGNN.
- Ablations on MS show removing the non-stationary module, system-state augmentation, or short-term CNN each degrades performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the non-stationary decomposition mechanism instead of a Transformer encoder (ablation) | STEAM · MS | workload prediction performance | Helped |
| add system-states graph augmentation and contrastive learning (ablation) | STEAM · MS | workload prediction performance | Helped |
| add CNN-based short-term temporal feature encoding (ablation) | STEAM · MS | workload prediction performance | Helped |
| set λ to 0.5 rather than 0.9 (ablation) | STEAM · MS (L=12) | MAE: 0.0846 -> 0.0789 | Helped |
| set λ to 0.7 rather than 0.9 (ablation) | STEAM · MS (L=24) | MAE: 0.0767 -> 0.0715 | Helped |
| set λ to 0.5 rather than 0.9 (ablation) | STEAM · MS (L=36) | MAE: 0.0770 -> 0.0891 | Hurt |
| set λ to 0.7 rather than 0.9 (ablation) | STEAM · MS (L=36) | MAE: 0.0770 -> 0.0852 | Hurt |
| set λ to 0.5 rather than 0.9 (ablation) | STEAM · MS (L=48) | MAE: 0.0738 -> 0.0990 | Hurt |
| set λ to 0.7 rather than 0.9 (ablation) | STEAM · MS (L=48) | MAE: 0.0738 -> 0.0979 | Hurt |
Surprising findings
- Statistical and some baselines like Prophet and ARIMA achieve competitive results at specific input lengths, and PatchTST is strong on MS, showing baseline performance varies non-monotonically with input length.
Limitations
- Evaluated only on two workload datasets, one of which is proprietary and not released.
- Similarity thresholds are fixed to the top 10% Pearson coefficients, which the paper itself argues can introduce prior interference.
- Loss weight lambda must be tuned per input length, with no single setting best across all L.
Experimental setup
Datasets: MS (Ant Group microservice workload, 46 microservices, 10 days) Ali (Alibaba Cluster Trace 2018, 148 machines, 8 days)
Benchmarks: MS Ali
Metrics: MAE RMSE
Baselines: ARIMA Prophet STGCN MTGNN ST-SSL DLinear FEDformer N-Transformer PatchTST Pyraformer
Builds on: Non-stationary Transformers Autoformer STGCN ST-SSL STAMP
Closest papers by meaning
- FastPERT: Towards Fast Microservice Application Latency Prediction via Structural Inductive Bias over PERT Networks (AAAI 2025 · similarity 0.85)
- Optimizing IT FinOps and Sustainability through Unsupervised Workload Characterization (AAAI 2024 · similarity 0.81)
- Root Cause Analysis in Microservice Using Neural Granger Causal Discovery (AAAI 2024 · similarity 0.81)
- Personalized Federated Learning for Spatio-Temporal Forecasting: A Dual Semantic Alignment-Based Contrastive Approach (AAAI 2025 · similarity 0.81)
- GraphLand: Evaluating Graph Machine Learning Models on Diverse Industrial Data (NeurIPS 2025 · similarity 0.80)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.