NeSy-Edge: Neuro-Symbolic Trustworthy Self-Healing in the Computing Continuum
Modern AI services increasingly run across the computing continuum—from cloud to edge devices—yet fault management remains challenging due to resource constraints, noisy telemetry, and cascading failures. This paper proposes NeSy-Edge, a three-layer neuro-symbolic framework that performs local log parsing, causal graph construction, and root-cause analysis on edge nodes, invoking cloud LLMs only when local evidence is insufficient. The core idea is to combine lightweight symbolic caching and prior-constrained causal discovery with selective neural inference, trading off autonomy against accuracy under strict memory budgets ($\sim$1500 MB).
The paper presents a coherent, layered architecture for edge-resident self-healing that improves upon task-specific baselines on standard Loghub benchmarks. The neuro-symbolic design is principled: symbolic caches and prior masks keep computation local, while a fallback to small and large language models handles novelty. The evaluation shows consistent gains at the perception, reasoning, and action layers under synthetic semantic noise. However, the work is limited by small-scale evaluation (190 reasoning cases, 144 action windows), emulation rather than deployment on real edge hardware, and reliance on synthetic noise injections that may not reflect production log drift. The claimed end-to-end accuracy (65% at best) and RCA accuracy (76% at best) under maximum noise are modest, suggesting the problem remains challenging.
The three-tier perception router (exact cache → semantic retrieval → local LLM) is a practical engineering choice that demonstrably reduces latency versus direct SLM parsing while maintaining higher parsing accuracy (PA 0.87–0.99) than Drain under noise. The reasoning layer's prior-constrained extension of DYNOTEARS is well-motivated: the penalty masks $W_{mask}$ and $A_{mask}$ successfully bias the optimizer toward historically supported edges, yielding sparser graphs (33–41 edges) with better root-cause ranking (AvgRank 1.00–1.74) than prior-free baselines. The selective cloud invocation in the action layer—bypassing the LLM when local evidence converges—is a credible strategy for edge-cloud trade-offs.
First, the evaluation relies on synthetic semantic noise (token swapping and variable perturbation) rather than organic deployment data; robustness to controlled synonym injection does not guarantee resilience to real-world semantic drift, schema changes, or adversarial logs. Second, the experimental scale is small—only 144 complete incident windows for end-to-end evaluation—limiting claims about generalization across heterogeneous continuum environments. Third, the 'edge-first' claim is weakened by the necessity of cloud LLM (DeepSeek-V3.2) for ambiguous cases, and the hardware evaluation emulates an edge node on a 24 GB MacBook Pro M4 rather than measuring on actual constrained IoT or gateway hardware. Fourth, the lack of end-to-end baselines is acknowledged but unresolved: comparing against vanilla LLM and RAG-only on isolated layers does not establish that the full NeSy-Edge pipeline outperforms a tightly integrated neural baseline in throughput or latency.
The layer-wise comparisons are fair and show consistent gains: NeSy-Edge outperforms Drain and Direct SLM in parsing accuracy, and beats DYNOTEARS, PC, and Pearson in causal graph quality (Table II). However, the action-layer comparison against 'Vanilla LLM' and 'RAG-only' is underspecified—it is unclear whether these baselines received identical prompt engineering or context windows—and the 65% end-to-end success rate suggests significant room for improvement. The claim that symbolic priors suppress 'implausible reverse-direction relations' is supported by the AvgRank metric, though the metric itself (rank of ground-truth edge) rewards sparse graphs by penalizing misses with |E|+1, which may favor the method's sparsity-inducing masks.
Reproducibility is partially obstructed. The authors provide an anonymous GitHub repository (https://anonymous.4open.science/r/NeSy-Edge-D794), which suggests the code will be available upon acceptance, but the current link is not persistent. Critical hyperparameters—such as the similarity threshold $\delta_{Sim}$, penalty values $c^{prior}$, $c^{rev}$, $c^{bg}$, and the pruning threshold $\theta_{prune}$—are described qualitatively but their exact values are not tabulated, making replication difficult. The reliance on a commercial cloud API (DeepSeek-V3.2) for the action layer introduces non-determinism and cost barriers, though the temperature is set to 0.0. The local model (Qwen3-0.6B) and datasets (Loghub) are public, which aids partial replication, but the exact preprocessing pipelines and embedding models for the retrieval stage are not fully specified.
The computational demands of modern AI services are increasingly shifting execution beyond centralized clouds toward a computing continuum spanning edge and end devices. However, the scale, heterogeneity, and cross-layer dependencies of these environments make resilience difficult to maintain. Existing fault-management methods are often too static, fragmented, or heavy to support timely self-healing, especially under noisy logs and edge resource constraints. To address these limitations, this paper presents NeSy-Edge, a neuro-symbolic framework for trustworthy self-healing in the computing continuum. The framework follows an edge-first design, where a resource-constrained edge node performs local perception and reasoning, while a cloud model is invoked only at the final diagnosis stage. Specifically, NeSy-Edge converts raw runtime logs into structured event representations, builds a prior-constrained sparse symbolic causal graph, and integrates causal evidence with historical troubleshooting knowledge for root-cause analysis and recovery recommendation. We evaluate our work on representative Loghub datasets under multiple levels of semantic noise, considering parsing quality, causal reasoning, end-to-end diagnosis, and edge-side resource usage. The results show that NeSy-Edge remains robust even at the highest noise level, achieving up to 75% root-cause analysis accuracy and 65% end-to-end accuracy while operating within about 1500 MB of local memory.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.