Probabilistic QoS Metric Forecasting in Delay-Tolerant Networks Using Conditional Diffusion Models on Latent Dynamics

Basic Information

• Paper ID: 2504.08821
• Title: Probabilistic QoS Metric Forecasting in Delay-Tolerant Networks Using Conditional Diffusion Models on Latent Dynamics
• Authors: Enming Zhang, Zheng Liu, Yu Xiang, Yanwen Qu
• Classification: cs.LG (Machine Learning), stat.ML (Statistics - Machine Learning)
• Publication Date: October 15, 2025 (arXiv v3)
• Paper Link: https://arxiv.org/abs/2504.08821v3

Abstract

This paper proposes a probabilistic forecasting method based on conditional diffusion models for QoS metric prediction in Delay-Tolerant Networks (DTNs). Traditional mean regression methods struggle to adequately capture data complexity, leading to performance degradation in operational tasks such as DTN routing. The paper formalizes QoS metric prediction in DTNs as a probabilistic forecasting problem for multivariate time series, quantifying prediction uncertainty through diffusion models combined with latent temporal dynamics. Experimental results demonstrate that the proposed method achieves optimal performance among popular probabilistic time series forecasting approaches.

Research Background and Motivation

Problem Definition

Delay-Tolerant Networks (DTNs) play a crucial role in communication over extremely long distances or in harsh environments, with widespread applications in vehicular communication, wildlife tracking networks, and rural area connectivity. DTNs frequently face challenges including frequent disconnections, high error rates, and delays that may persist for hours or even days.

Problem Significance

1. Network Performance Optimization: QoS metric prediction enhances network performance in latency, throughput, energy consumption, and reliability
2. Resource Scheduling: Routing protocols can distribute information more efficiently based on network state predictions, reducing energy consumption
3. Traffic Priority Management: Different types of network traffic have varying QoS requirements; prediction facilitates priority allocation

Limitations of Existing Methods

1. Traditional Mean Regression Methods: Based on autoregressive models, focused on providing accurate point predictions by minimizing metrics such as MSE
2. Insufficient Multimodal Data Processing: When data contains multiple modalities, mean regression cannot capture the complete complexity of the data
3. Missing Uncertainty Quantification: Deterministic regression methods cannot quantify prediction uncertainty, which is necessary for operational tasks such as DTN routing

Research Motivation

QoS time series in DTNs exhibit non-stationary and multimodal characteristics, requiring methods that can:

• Capture complex temporal dynamics in sequential data
• Provide a probabilistic forecasting framework that quantifies prediction uncertainty
• Handle non-stationary and multimodal data robustly

Core Contributions

1. Problem Reformulation: Redefines QoS metric prediction in DTNs as a probabilistic time series forecasting problem, enabling quantification of prediction uncertainty
2. Diffusion Model Extension: Extends popular diffusion models to infer samples from prediction distributions, incorporating latent contextual dynamics to improve model adaptability to non-stationary and multimodal time series
3. DiffTCN Framework: Proposes a conditional diffusion model framework combining Temporal Convolutional Networks (TCN) and Transformers, effectively capturing long-term dependencies
4. Experimental Validation: Demonstrates the advantages of the proposed method through extensive experiments, achieving state-of-the-art performance across multiple evaluation metrics

Methodology Details

Task Definition

Given a multivariate time series $X = (x_1, x_2, ..., x_t) \in \mathbb{R}^{m \times t}$, where $m$ is the number of variables and $t$ is the time series length. The objective is to forecast the distribution of the next $p$ time steps $Y = (y_{t+1}, y_{t+2}, ..., y_{t+p}) \in \mathbb{R}^{n \times p}$, where $n \leq m$.

Model Architecture

1. Diffusion Model Fundamentals

The diffusion model comprises two processes:

• Forward Diffusion Process: Progressively adds Gaussian noise to data $q(x^s|x^{s-1}) = \mathcal{N}(x^s; \sqrt{1-\beta_s}x^{s-1}, \beta_s I)$
• Reverse Denoising Process: Recovers data from noise $p_\theta(x^{s-1}|x^s) = \mathcal{N}(x^{s-1}; \mu_\theta(x^s, s), \sigma_\theta(x^s, s)I)$

The loss function simplifies to: $L_{simple} = \mathbb{E}_{s \sim [1,S], x^0, \varepsilon_s}[\|\varepsilon_s - \varepsilon_\theta(\sqrt{\bar{\alpha}_s}x^0 + \sqrt{1-\bar{\alpha}_s}\varepsilon_s, s)\|^2]$

2. Conditional Diffusion Model

To improve prediction accuracy and reliability, conditional information is introduced: $p_\theta(x^{s-1}|x^s, e^j) = \mathcal{N}(x^{s-1}; \mu_\theta(x^s, s, e^j), \sigma_\theta(x^s, s, e^j)I)$

where $e^j$ represents temporal dynamics at time $j$. The conditional loss function is: $L_{simple}^s = \mathbb{E}_{s \sim [1,S], x^0, \varepsilon_s}[\|\varepsilon_s - \varepsilon_\theta(\sqrt{\bar{\alpha}_s}x^0 + \sqrt{1-\bar{\alpha}_s}\varepsilon_s, s), e^j\|^2]$

3. Latent Temporal Dynamics Capture

Temporal Convolutional Networks (TCN) are employed to capture contextual dynamics, with main components including:

• Causal Convolution: Prevents information leakage
• 1D Fully Convolutional Networks: Input-to-output mapping
• Dilated Convolution: Expands receptive field
• Residual Temporal Blocks: Deepens the network

4. Transformer Integration

Transformers replace dilated convolutional layers as the primary computational module:

• Captures long-term dependencies in time series
• Dynamically weights feature values across different time steps
• Captures interdependencies between different variables through attention mechanisms

Technical Innovations

1. Probabilistic Forecasting Paradigm: Transitions from deterministic to probabilistic prediction, enabling uncertainty quantification
2. Conditional Diffusion Design: Incorporates contextual temporal dynamics as conditioning information in the diffusion process
3. TCN-Transformer Fusion: Combines TCN's parallelizability with Transformer's long-term dependency capture capabilities
4. Iterative Prediction Strategy: Employs step-by-step forecasting applicable to arbitrary prediction horizons

Experimental Setup

Datasets

Public sensor datasets containing QoS metrics such as internet latency:

• Data Scale: Over 100 days of data, sampled at one-minute intervals
• Data Partition:
  • D1: Approximately 80,000 timestamps
  • D2: Approximately 40,000 timestamps with different value ranges and distributions
• Forecasting Setting: Uses the most recent 120 timestamps to predict latency values for the next 10 time steps

Evaluation Metrics

1. MAE (Mean Absolute Error): Average absolute error
2. MSE (Mean Squared Error): Mean squared error
3. CRPS (Continuous Ranked Probability Score): Continuous ranked probability score $CRPS(F,x) = \int_{-\infty}^{\infty} (F(y) - \mathbf{1}(x \leq y))^2 dy$

Baseline Methods

• DeepAR: Deep autoregressive method based on recurrent networks
• DeepFactor: Hybrid model combining classical time series models and deep neural networks
• TimeGrad: Autoregressive denoising method based on diffusion models
• Diffusion: Vanilla diffusion model

Implementation Details

• Optimizer: Adam with learning rate 0.001
• Diffusion Steps: 50
• Noise Levels: β₁ = 0.0001, βₜ = 0.5
• Batch Size: 64
• Inference Samples: 100 samples to approximate distribution
• Hardware: Single Nvidia 3090 GPU (24GB)

Experimental Results

Main Results

MAE and MSE Results

CRPS Results

Experimental Findings

1. Consistent Advantages: DiffTCN achieves optimal performance across all forecasting horizons and datasets
2. Uncertainty Quantification: CRPS results demonstrate that DiffTCN significantly outperforms baseline methods in probabilistic forecasting quality
3. Stability: DiffTCN exhibits minimal performance variability, indicating method robustness
4. TCN Advantages: Compared to TimeGrad's RNN encoder, TCN's parallelizability yields significant accuracy improvements

Related Work

QoS Prediction in DTNs

• Neural Network Methods: Abdellah et al. use NARX-RNN for network latency prediction
• LSTM Applications: Nagai et al. construct wireless sensor network testbeds using LSTM for latency prediction
• Matrix Factorization: Ghandi et al. formalize network latency prediction as non-negative matrix factorization

Time Series Forecasting

• Traditional Methods: Prediction methods based on statistical models such as ARMA
• Deep Learning: Recurrent neural network approaches including RNN, LSTM, and GRU
• Probabilistic Forecasting: Probabilistic forecasting frameworks such as DeepAR and DeepFactor

Diffusion Model Applications

• Image Generation: Successful applications of diffusion models in image generation tasks
• Time Series: Application of diffusion models to time series forecasting such as TimeGrad

Conclusions and Discussion

Main Conclusions

1. Probabilistic Forecasting Advantages: Reformulating QoS prediction as probabilistic forecasting effectively quantifies uncertainty
2. Diffusion Model Effectiveness: Conditional diffusion models effectively handle non-stationary and multimodal time series data
3. Reasonable Architecture Design: The combination of TCN and Transformer leverages the strengths of both approaches

Limitations

1. Computational Complexity: Diffusion models require multi-step sampling, incurring significant computational overhead
2. Dataset Scope: Experiments are validated only on a single type of network latency data
3. Practical Deployment: Lacks deployment and validation in real DTN environments

Future Directions

The paper identifies the primary future research direction as extending the framework to handle out-of-distribution scenarios, i.e., inference on data with distributions different from training data.

In-Depth Evaluation

Strengths

1. Clear Problem Definition: Accurately identifies limitations of existing methods and reasonably reformulates the problem as probabilistic forecasting
2. Method Innovation: Cleverly combines diffusion models with temporal dynamics modeling; the technical approach is novel
3. Comprehensive Experiments: Compares against multiple strong baselines using multiple evaluation metrics with convincing results
4. Engineering Implementation: Provides detailed implementation details facilitating reproducibility

Weaknesses

1. Missing Theoretical Analysis: Lacks theoretical analysis explaining why the method is effective and convergence proofs
2. Limited Dataset Diversity: Validation only on network latency data; lacks verification on other types of QoS metrics
3. Computational Efficiency: Insufficient discussion of diffusion model computational overhead and real-time performance
4. Insufficient Ablation Studies: Lacks detailed analysis of individual component contributions

Impact

1. Academic Contribution: Provides new technical approaches for QoS prediction in DTN domains
2. Practical Value: Probabilistic forecasting capability has significant value for actual network operations
3. Method Generalizability: The proposed framework can be extended to other time series forecasting tasks

Applicable Scenarios

1. DTN Network Optimization: Routing decisions, resource allocation, load balancing
2. Network Operations: Fault prediction, performance monitoring, capacity planning
3. Other Time Series Forecasting: Time series prediction tasks requiring uncertainty quantification

References

The paper cites 51 relevant references covering multiple domains including DTNs, time series forecasting, and diffusion models, providing a solid theoretical foundation for the research.

Overall Assessment: This paper addresses practical QoS prediction requirements in DTNs with a technically innovative and experimentally well-validated solution. Despite certain limitations in theoretical analysis and experimental breadth, it represents high-quality research work with important reference value for related fields.