🤖 machine learning

DSTFView: Multi-View Cloud-Edge Workload Forecasting with Dual-Input Spatio-Temporal-Frequency Modeling

The paper proposes DSTFView, a dual-input spatio-temporal-frequency multi-view framework that effectively balances feature modeling and forecasting efficiency to predict collaborative cloud-edge workloads by jointly capturing closeness, period, and frequency-domain dependencies while adaptively fusing views to handle abrupt changes.

Original authors: Qingzhong Li, Hui Ma, Yajun Zhang, Qingchang Ma, Zhou Long

Published 2026-07-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Qingzhong Li, Hui Ma, Yajun Zhang, Qingchang Ma, Zhou Long

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the internet as a giant, bustling city. In the old days, all the heavy lifting—like processing your photos or running complex games—happened in a massive, distant skyscraper called "The Cloud." But as we started asking our devices to do more things instantly, like self-driving cars or augmented reality, that long trip to the skyscraper became too slow. So, we built smaller, local "edge" stations right in the neighborhoods to handle the quick stuff. This is the world of Cloud-Edge Collaboration: a team effort where the big brain (Cloud) and the local helpers (Edge) work together.

However, these edge stations are chaotic. They are like busy coffee shops where the number of customers changes wildly from minute to minute. Sometimes it's a quiet Tuesday, and sometimes it's a sudden rush of orders. If the shop owner (the computer system) doesn't guess the rush correctly, they either run out of coffee (crash) or waste money keeping too many baristas on standby (wasted energy). The big challenge for scientists is figuring out how to predict these crazy rushes accurately, especially when the patterns are different for every shop and change over time.

Enter DSTFView, a new "crystal ball" designed by researchers to solve this prediction puzzle. Think of it as a super-smart weather forecaster for computer traffic. Instead of just looking at the past few minutes, this system looks at three different "lenses" at once: the immediate past (closeness), the repeating daily habits (period), and the hidden rhythms in the data (frequency). It also acts like a two-way street between the neighborhood stations and the central skyscraper. The local stations send a quick summary of what's happening to the Cloud, which figures out the big picture and sends advice back down. The local stations then mix that big-picture advice with their own detailed observations to make a final, highly accurate guess about what will happen next.

The researchers tested this idea on two real-world datasets: one tracking how hard computer processors were working (CPU) and another measuring network traffic speed (TP). They compared their new method against many other popular forecasting tools. The results showed that DSTFView was consistently better at guessing the future workload than the other methods. It didn't just get the average right; it was particularly good at spotting sudden spikes and drops in traffic, which are the hardest parts to predict. By combining the local view with the global view and using those three different lenses to understand time and patterns, DSTFView suggests a more reliable way to keep our digital cities running smoothly without wasting resources or crashing under pressure.

Technical Summary of DSTFView: Multi-View Cloud-Edge Workload Forecasting with Dual-Input Spatio-Temporal-Frequency Modeling

1. Problem Statement

The proliferation of edge-side AI inference has created a demand for edge platforms to support latency-sensitive, highly concurrent, and reliability-critical applications. Effective resource management in these environments relies heavily on accurate workload forecasting. However, existing methods face significant challenges in collaborative cloud-edge settings:

Heterogeneity and Complexity: Edge workloads exhibit complex features including spatial dependencies across sites, temporal dynamics (burstiness, trend shifts), and frequency-domain regularities (service cycles). Current methods often focus on only one perspective (temporal, frequency, or spatio-temporal) or fail to uniformly integrate these multi-source, multi-pattern features.
Uncertainty in Dependencies: Relationships in edge environments (geographic proximity, service affinity, shared-resource coupling) are dynamic, partially observable, and inherently uncertain due to user mobility and dynamic deployment.
Efficiency vs. Accuracy Trade-off: Cloud-centric solutions introduce communication latency and costs, while edge-only solutions overlook cross-site dependencies. Furthermore, existing collaborative frameworks struggle with multi-step forecasting, often failing to capture multi-scale periodic behaviors or handling unreliable connections that degrade performance over extended horizons.

2. Methodology: DSTFView Framework

The authors propose DSTFView, a dual-input spatio-temporal-frequency (STF) multi-view workload forecasting framework designed for collaborative cloud-edge architectures. The system operates via a two-stage pipeline:

A. Framework Overview
The framework consists of an Edge-side Aggregation Layer, a Cloud-side Global Stage, and an Edge-side Local Stage.

Aggregation: Edge sites compress historical VM-level workload traces into lightweight site-level sequences using a learnable aggregation layer to minimize communication overhead.
Dual-Input Mechanism: Both global and local stages utilize two distinct input branches:
1. Closeness Input: Captures recent dynamics.
2. Period Input: Models recurring temporal patterns.

B. Global Stage (Cloud-side)

Input: Aggregated site-level sequences from all participating edge sites.
Modeling: The cloud performs global STF modeling over multi-view site graphs.
- Frequency Extraction: Fast Fourier Transform (FFT) is applied to extract frequency-domain features from both input branches.
- Temporal Modeling: Gated Temporal Convolutional Networks (TCN) capture temporal dependencies.
- Spatial Modeling: Graph Convolutional Networks (GCN) operate over three inter-site graph views: a Spatial Graph (geographic), a Temporal Graph, and a Self-adaptive Graph.
- Adaptive Multi-view Fusion (AMVF): An adaptive mechanism dynamically reweights the representations from the three graph views based on learned importance scores, allowing the model to focus on the most relevant relational structures.
Output: Coarse-grained, multi-step site-level forecasts are generated and returned to edge sites as global guidance.

C. Local Stage (Edge-side)

Input: Fine-grained VM-level workload sequences and the global guidance from the cloud.
Modeling: The local predictor mirrors the global STF architecture but operates at the VM granularity using intra-site multi-view graphs:
- Graph Views: Physical Graph, Logical Graph, and Self-adaptive Graph.
- Fusion: The AMVF module adaptively fuses these views to refine dependency modeling.
Disaggregation and Final Fusion: The site-level global forecast is mapped back to the VM level via a disaggregation layer. The final VM-level prediction is obtained by adaptively fusing the local prediction and the disaggregated global guidance using a learnable gating tensor ( $\lambda_m$ ).

3. Key Contributions

The paper outlines three primary contributions:

Novel Architecture: A dual-input spatio-temporal-frequency architecture specifically designed for collaborative cloud-edge multi-step forecasting, which jointly models closeness and period dependencies while extracting spatial, temporal, and frequency-domain features.
Adaptive Fusion Strategy: A multi-view graph fusion strategy that dynamically integrates different relational views (spatial, temporal, physical, logical, and self-adaptive) to adapt to varying workload patterns.
Collaborative Pipeline and Validation: The construction of a global-local collaborative forecasting pipeline and extensive experimental validation on real-world datasets (CPU and TP), demonstrating consistent superiority over baseline methods.

4. Experimental Results

The authors evaluated DSTFView on two datasets:

CPU Dataset: CPU utilization traces from 1,728 VMs across 77 edge sites (Alibaba ENS).
TP Dataset: Network traffic traces (uplink/downlink throughput) from the Network Edge Platform (Alibaba NEP).

Performance Metrics:

Metrics: Mean Absolute Error (MAE), Symmetric Mean Absolute Percentage Error (SMAPE), and $R^2$ .
Horizons: Predictions were tested at horizons $H=1, 4, 8, 12$ .
Baselines: The model was compared against temporal baselines (e.g., PatchTST, TimesNet, iTransformer), spatio-temporal baselines (e.g., STGCN, DCRNN, GWNET), and collaborative cloud-edge baselines (ELASTIC, XELASTIC).

Findings:

DSTFView consistently outperformed all representative baselines across both datasets and all forecasting horizons.
Ablation Studies:
- Removing the multi-view fusion or relying on single-view graphs (spatial-only, temporal-only) resulted in degraded performance, confirming the necessity of joint modeling.
- The full AMVF configuration achieved the best results, validating the effectiveness of adaptively fusing multiple relational views.
Visualization: Qualitative analysis showed DSTFView tracked ground-truth series more closely than the strongest baseline (XELASTIC), particularly during peaks, valleys, and local trend reversals.

5. Significance and Claims

The paper claims that DSTFView addresses the critical gap in balancing multidimensional feature modeling with forecasting efficiency in collaborative cloud-edge environments. By unifying the modeling of recent dynamics, periodic patterns, and heterogeneous spatial dependencies through a dual-input STF approach, the framework achieves higher accuracy in multi-step forecasting.

The authors emphasize that the adaptive fusion mechanism allows the model to capture abrupt changes and handle the uncertainty inherent in edge dependencies. The results suggest that this approach is superior to existing methods that either ignore cross-site dependencies or fail to effectively integrate frequency-domain information and multi-view graph structures.

Future Work:
The authors note that future research will focus on dynamic graph construction, continual learning, and integrating the forecasting framework with resource management tasks such as auto-scaling and scheduling to enhance practical adaptability in real-world edge environments.