SHARP-QoS: Unified Multi-QoS Prediction

• SHARP-QoS is a unified neural architecture that jointly predicts multiple QoS parameters in service-oriented computing, reducing model redundancy.
• It leverages hyperbolic hierarchical convolutions and sparsely-gated routing to extract and fuse both task-specific and contextual features efficiently.
• An EMA-based dynamic loss balancing strategy stabilizes training, leading to significant improvements in prediction accuracy and robustness.

SHARP-QoS is a unified neural architecture for joint prediction of multiple Quality of Service (QoS) parameters in service-oriented computing. It is designed to address the challenges of extreme QoS data sparsity, noise, inherent hierarchical dependencies, and negative transfer in multi-task settings. Unlike conventional methods that individually predict each QoS parameter—resulting in redundant models and poor generalization—SHARP-QoS leverages shared hyperbolic graph representations, adaptive feature-sharing through sparsely-gated routing, and a novel exponential moving average (EMA)-based loss balancing strategy. These innovations enable efficient, robust, and scalable joint prediction across all QoS parameters in a single inference pass (Kumar et al., 19 Dec 2025).

1. Overall Model Structure and Prediction Pipeline

SHARP-QoS predicts $P$ QoS parameters (e.g., response time RT, throughput TP, reliability RE, cost CT) for any user–service pair $(u_i, s_j)$, using a four-stage pipeline:

1. Preprocessing:
   • Constructs $P$ bipartite QoS invocation graphs $\mathcal{G}^p$ (one per parameter), and two context graphs: region ($\mathcal{G}^r$) and autonomous system ($\mathcal{G}^a$).
   • Generates initial node features through nonnegative matrix factorization (NMF) on $\mathcal{Q}^p$, and context representations via autoencoders.
2. Hierarchical Feature Extraction Block (HFEB):
   • Extracts per-task structural features using a Hyperbolic Hypergraph Graph Convolution Network (HHGCN) on $\mathcal{G}^p$.
   • Extracts contextual features using a Hyperbolic GCN (HyGCN) on $\mathcal{G}^r$ and $\mathcal{G}^a$.
   • Both operate in the Poincaré ball $\mathbb{B}^{d, c}$ (hyperbolic space).
3. Feature Sharing and Fusion Block (FSFB):
   • Enables adaptive cross-task feature sharing through sparsely-gated Sub-Network Routing (SNR) for context and Cross-SNR for structural representations.
   • Uses a learned gate to fuse each task’s structural and shared features.
4. Joint QoS Prediction Module (JQPM):
   • For each $p$, a small feed-forward network produces user/service embeddings; their inner product gives the predicted QoS score $\hat{\mathcal{Q}}^p$.
   • All task-specific loss terms are combined via the EMA-based strategy.

Joint prediction is achieved by parameter sharing across all $P$ tasks for the HFEB and FSFB, unified loss balancing, and simultaneous output of all predicted QoS matrices in a single model evaluation.

2. Hierarchical Hyperbolic Feature Extraction

The model operates on graphs embedded in a constant-curvature hyperbolic space $$\mathbb{B}^{d, c} = \{x \in \mathbb{R}^d \mid c\|x\|^2 < 1\}$$ with learnable curvature $c$. Key hyperbolic algebraic operations include:

• Exponential and logarithmic maps:

$\exp_{<0>}^{c}(v) = \tanh\bigl(\sqrt{c}\|v\|\bigr)\,\frac{v}{\sqrt{c}\|v\|}, \ \log_{<0>}^{c}(x) = \frac{1}{\sqrt{c}} \artanh(\sqrt{c}\|x\|) \frac{x}{\|x\|}$

• Möbius addition:

$$x \oplus_c y = \frac{(1+2c\langle x,y\rangle + c\|y\|^2) x + (1-c\|x\|^2) y}{1+2c\langle x,y\rangle + c^2\|x\|^2\|y\|^2}$$

• Möbius matrix–vector multiplication:

$$W \otimes_c x = \exp_{<0>}^{c}\left(W \log_{<0>}^{c}(x)\right)$$

• Wrapped activation:

$$\sigma^{\otimes_c}(x) = \exp_{<0>}^{c}(\sigma(\log_{<0>}^{c}(x)))$$

Each HyGCN layer performs message passing in the tangent space, Möbius-linear transformation, activation, and returns to Euclidean space with

$$\begin{aligned} X_{l+1} &= \sigma^{\otimes_c}\left(\exp_{<0>}^c(\hat A\,\log_{<0>}^c(X_l) W_{l1})\,\oplus_c\,\exp_{<0>}^c(b_{l1})\right) \ X_{l+2} &= \sigma^{\otimes_c}\left(\exp_{<0>}^c(\log_{<0>}^c(X_{l+1}) W_{l2})\,\oplus_c\,\exp_{<0>}^c(b_{l2})\right) \end{aligned}$$

For each QoS graph $\mathcal{G}^p$, two associated hypergraphs (user-based and service-based, incorporating second-hop hyperedges) feed into parallel HHGCNs, after which features are concatenated and passed through a final linear layer to yield the structural feature $\mathcal{Y}^p \in \mathbb{R}^{N \times d}$. Contextual features $\mathcal{Y}^r, \mathcal{Y}^a$ are similarly extracted for region and AS context graphs.

3. Adaptive Feature Sharing and Gated Fusion

In the FSFB, SHARP-QoS employs a dual mechanism:

• Sparsely-Gated Sub-Network Routing (SNR):
  • Fused context features $\mathcal{Y}^{ra} = \mathcal{Y}^r + \mathcal{Y}^a$ are processed by $K_1$ Dense blocks $\phi_k$, routed per-task via hard-concrete gates $c_k^p \in \{0,1\}$.
  • Each task's sparsely-shared feature representation is:

  $$\mathcal{Y}_s^p = \frac{1}{\sum_k c_k^p}\sum_{k=1}^{K_1} c_k^p\,\tilde\phi_k^p(\mathcal{Y}^{ra})$$ - Gates are sampled with a temperature-controlled relaxation during training, thresholded during inference.

• Cross-Task SNR:

  • Applies a similar routing scheme (with $K_2$ blocks) to other tasks' structural features, enabling $$\mathcal{Y}^{p}_{cs} = \sum_{j \neq p} \Theta(\mathcal{Y}^j)$$ and thus $$\mathcal{Y}^{p}_{scs} = \mathcal{Y}^p_s + \mathcal{Y}^{p}_{cs}$$.
• Fusion by Learned Gate:
  • Each task's structural feature $\mathcal{Y}^p$ and aggregated context/structural shared feature $\mathcal{Y}^p_{scs}$ are fused by a sigmoid gate:

  $$g^p = \sigma\left([\mathcal{Y}^p || \mathcal{Y}^p_{scs}] W^p_g\right)$$

  $$\mathcal{Z}^p = g^p \odot \mathcal{Y}^p + (1 - g^p) \odot \mathcal{Y}^p_{scs}$$ - $\mathcal{Z}^p$ serves as the final embedding for each task.

4. Dynamic Loss Balancing with EMA

To mitigate negative transfer due to loss scale mismatches across tasks, SHARP-QoS incorporates an EMA-based loss balancing:

• Per-task Loss:

$$\mathcal{L}^p = \frac{1}{|\mathrm{TD}|}\sum_{(i,j) \in \mathrm{TD}} |\mathcal{Q}^p_{ij} - \hat{\mathcal{Q}}^p_{ij}|$$

• EMA Calculation:

$$\tilde{\mathcal{L}}^p_i = \beta\,\tilde{\mathcal{L}}^p_{i-1} + (1-\beta)\,\mathcal{L}^p_i$$

• Dynamic Weights:

$$w^p = \frac{(\tilde{\mathcal{L}}^p + \epsilon)^{-1}}{\sum_{q=1}^P (\tilde{\mathcal{L}}^q + \epsilon)^{-1}}$$

• Joint Objective:

$$\mathcal{L}_{\mathrm{EMA}} = \sum_{p=1}^P w^p\,\mathcal{L}^p$$

• Final Loss with Sparsity Regularization:

$$\mathcal{L}_{\mathrm{final}} = \mathcal{L}_{\mathrm{EMA}} + \lambda \bigl(\mathcal{L}_0^{\mathrm{snr}} + \mathcal{L}_0^{\mathrm{cross\text{-}snr}}\bigr)$$

The $\ell_0$ penalty encourages sparse routing, further promoting effective specialization. A plausible implication is that this mechanism stabilizes training under widely varying QoS parameter scales.

5. Experimental Protocol and Baselines

Experiments were conducted on three datasets, each incorporating region and AS context:

Dataset	Users × Services	QoS Parameters
WSDREAM-2T	339 × 5825	RT, TP
Small-3T	112 × 36	RT, TP, RE
gRPC-4T	57 × 150	RE, CT, LT, PW

• Feature preprocessing: Each $\mathcal{Q}^p$ processed with NMF (output dim $d_1 = 128$); region/AS one-hot encoding followed by autoencoder (dim $d_2 = 128$). Features are concatenated ($\mathcal{F}^i \in \mathbb{R}^{(n + m) \times 256}$).
• Hyperparameters: 2 HyGCN/HHGCN layers ($L=2$), hidden dim $d=128$, 4 SNR/Cross-SNR blocks ($K_1=K_2=4$), block output $d_o=64$, gate threshold $\delta=0.5$, EMA $\beta=0.99$, $\lambda=10^{-5}$; training with AdamW (lr$=10^{-3}$, decay $10^{-4}$), 10k epochs, patience=400.
• Evaluation metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), relative improvement $I(\%)$.
• Baselines: Multi-task (DNM, JQSP, MGEN, WAMTL); Single-task (PMF, CMF, DCALF, QoSGNN, llmQoS).

6. Quantitative and Qualitative Performance

Multi-Task and Single-Task Results

• Multi-task: SHARP-QoS achieves $\sim$19.5% lower average MAE and $\sim$19.3% lower RMSE versus the best alternative multi-task baseline across all datasets and observed density splits. Relative improvements increase with the number of predicted QoS parameters.
• Single-task: On WSDREAM-2T, MAE matches or improves upon all state-of-the-art single-task baselines, despite SHARP-QoS jointly predicting multiple tasks.

This suggests that joint modeling with adaptive feature sharing does not compromise single-task accuracy.

Computational Efficiency and Robustness

• Model size and runtime: 1.62M parameters, 133.4G FLOPs per pass, fastest inference (4 × $10^{-8}$ s).
• Ablation studies: Removing HHGCN increases error by 1.6–7.9%; removing SNR or Cross-SNR each substantially degrades metrics; omitting any two key modules degrades performance by up to 25% on TP.
• Loss balancing: EMA strategy reduces MAE by up to 6.8% and RMSE by up to 3.9% over previous weighting methods; module ablations confirm each novel contribution is critical.
• Robustness to outliers: Removing the top 2–10% of outliers yields up to 81.8% MAE and over 70% RMSE improvement on RT.
• Cold-start: When 20% of users/services are fully absent from training, MAE increases moderately (e.g., RT MAE from 0.293 to 0.458), indicating graceful degradation due to context-graph sharing.
• Context graph impact: Isolated inclusion of region or AS context improves RT/TP MAE by 1.3–3.5%; using disconnected HHGCNs for QoS, RG, and AS outperforms graph merging by $\sim$1%.

These results demonstrate that SHARP-QoS delivers improved accuracy, computational efficiency, and robustness for joint multi-QoS prediction under challenging real-world conditions (Kumar et al., 19 Dec 2025).

7. Methodological Significance

SHARP-QoS introduces three methodological innovations established by ablation and comparative study:

• Hyperbolic hierarchical convolution enables expressive, scalable representation of graph-structured and hierarchical relationships in sparse heterogenous QoS data.
• Sparsely-gated adaptive multi-context routing facilitates efficient feature sharing and specialization across tasks without negative transfer.
• EMA-based loss balancing directly addresses scale mismatches between diverse QoS objectives, stabilizing multi-task training.

These innovations collectively allow SHARP-QoS to outperform established single-task and multi-task models, specifically under data sparsity, outlier, and cold-start regimes, while maintaining practical computational overhead. The architecture establishes a new direction for graph-based multi-task models in service-oriented computing by integrating hyperbolic geometry, adaptive neural routing, and dynamic objective weighting (Kumar et al., 19 Dec 2025).