Hypergraph and Latent ODE Learning for Multimodal Root Cause Localization in Microservices

Xin Liu*, Yuhang He, Sichen Zhao, Kejian Tong, Xingyu Zhang · 2026

Read Hypergraph and Latent ODE Learning for Multimodal Root Cause Localization in Microservices by Xin Liu*, Yuhang He, Sichen Zhao, Kejian Tong, Xingyu Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-00351.txt.

Introduction

Xin Liu and 4 co-authors build a system that finds the root cause of failures in cloud microservice platforms. A modern service is split into many small parts, so one failure can ripple through tangled dependencies while logs, traces, and metrics pile up in different forms. Their framework, HyperODE, attacks the dependency, timing, and data variety problems together.

Hypergraph attention learning captures interactions that involve several services at once, beyond simple pairs. A latent differential equation encoder follows how an anomaly evolves through irregularly timed observations. Cross-attention then fuses the five kinds of evidence, with routing that weighs each source by context. Added safeguards, including an information bottleneck and causal regularization, aim at robustness. On a public operations benchmark, the system beat strong baselines on ranking and classification while keeping its reasoning inspectable.

Catalog record

CREATORXin Liu*, Yuhang He, Sichen Zhao, Kejian Tong, Xingyu Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE20 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.00351

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Hypergraph and Latent ODE Learning for Multimodal Root Cause Localization in Microservices Xin Liu*, Yuhang He, Sichen Zhao, Kejian Tong, Xingyu Zhang Abstract Root cause localization in cloud native microservice systems requires modeling complex service dependencies, irregular temporal dynamics, and heterogeneous observability data. We present HyperODE RCA, a unified framework that combines hypergraph attention learning, latent ordinary differential equations, and multimodal cross attention fusion for fine grained root cause analysis. The method learns higher order service interactions through differentiable hyperedge construction, captures continuous anomaly evolution from irregular observations with an ODE RNN encoder, and adaptively fuses logs, traces, metrics, entities, and events using context aware modality routing. We further improve robustness with a variational information bottleneck, temporal causal regularization, and invariant risk constraints. Experiments on the Tianchi AIOps benchmark show clear gains over strong baselines in ranking and classification performance, while preserving interpretability through learned hypergraph attention. I Introduction Cloud native applications are commonly deployed as large microservice systems, where failures propagate across loosely coupled components and produce heterogeneous evidence in logs, traces, metrics, and events. Advances in continuous-time modeling, higher-order relational learning, and multimodal representation learning create new opportunities for more faithful diagnosis of such systems [ 7 , 3 , 9 ] . Existing root cause localization pipelines still have three major limitations. First, pairwise service graphs can overlook coordinated failures involving multiple services. Second, discrete sequence models handle irregular observations poorly and do not explicitly capture propagation speed. Third, simple fusion strategies underuse complementary observability signals, reducing causal attribution under noisy incidents [ 7 , 3 ] . We propose HyperODE RCA, a unified framework for multimodal root cause localization in microservice environments. It learns data-driven hypergraphs for higher-order service interactions, uses a latent ODE encoder to model continuous anomaly evolution, and applies cross-attention fusion with adaptive modality routing. Temporal causal regularization and a variational bottleneck with invariant constraints further improve diagnostic accuracy and interpretability across diverse fault scenarios [ 9 ] . II Related Work Early lines of recent work on system diagnosis increasingly connect root cause analysis with temporal dependency discovery and causal structure learning. Xue et al. [ 14 ] proposed a risk-aware dynamic routing framework that combines a GCN-GRU spatiotemporal graph model with dynamic edge-weight adjustment, enabling prediction-guided adaptation to time-varying network risk. For HyperODE-RCA, this idea is technically relevant because it suggests modeling service dependencies as dynamically reweighted graph relations conditioned on predicted anomaly evolution, which can strengthen fault-propagation reasoning in microservice systems. Methods for learning directed relations from time series provide a useful foundation because they formalize how events in one component may precede changes in another, yet they typically assume relatively simple variable interactions and are not designed for heterogeneous observability streams in microservice systems [ 10 ] [ 2 ] .Wang et al. [ 13 ] introduced SEAM, a defense that couples optimization on benign and harmful data, together with adversarial gradient ascent, so that models retain legitimate capabilities while becoming intrinsically resistant to harmful fine-tuning. This robustness-oriented training perspective is relevant to HyperODE-RCA because it motivates objectives that preserve diagnostic utility while reducing brittle adaptation to spurious or adversarial observability patterns. A second line of work studies stronger representation learning and fusion architectures for complex inputs. Gao et al. [ 4 ] introduced an enhanced retrieval-augmented generation framework that couples ScaNN-based nearest-neighbor search with the Gemma language model, improving retrieval efficiency, scalability, and context-aware response generation. For our setting, this line of work suggests a practical way to strengthen the retrieval and contextual grounding of heterogeneous observability evidence before multimodal fusion, thereby complementing HyperODE-RCA’s downstream diagnosis pipeline. Multimodal transformers improve information exchange across heterogeneous signals, while scalable latent architectures support flexible processing of structured inputs. These advances motivate RCA models that move beyond handcrafted aggregation and allow incident dependent weighting of different evidence sources [ 6 ] . Tong et al. [ 11 ] proposed an integrated machine-learning and deep-learning framework that combines neural networks with ensemble base learners and uses SMOTE-based preprocessing to improve robustness under class imbalance. For HyperODE-RCA, this hybrid classification strategy is a relevant complement to the final root-cause scoring stage, suggesting a practical way to improve decision reliability when fault classes are unevenly distributed. Our method is also closely related to research on irregular time series modeling and hypergraph learning. Neural controlled differential models show the value of continuous dynamics for uneven observations, and recent hypergraph networks demonstrate how higher order relations can be learned more effectively than pairwise graphs alone. In addition, information bottleneck based robust learning offers a principled way to reduce spurious correlations during prediction [ 5 ] [ 1 ] . Wang et al. [ 12 ] showed that large reasoning models can rely on retrieval shortcuts that compete with explicit reasoning, and introduced FARL to suppress such shortcut behavior through memory unlearning combined with reinforcement learning. For our setting, this perspective further motivates mechanisms that reduce reliance on spurious observability cues and strengthen diagnosis based on causally informative multimodal evidence. Long et al. [ 8 ] proposed a Transformer-based matching framework trained with InfoNCE loss and knowledge distillation, which improves robustness to noisy related samples and increases matching accuracy and efficiency. In our setting, these ideas can strengthen noise-resistant alignment among heterogeneous observability signals, particularly by complementing contrastive log template clustering and cross-modal representation learning. III Methodology We propose HyperODE-RCA for root cause localization in cloud-native microservices. Unlike prior methods based on pairwise dependencies and discrete-time assumptions, it models higher-order fault propagation and continuous anomaly evolution. The framework combines HyperGAT, which learns multi-service dependencies through differentiable hyperedge generation with Gumbel-Softmax relaxation, and a Latent ODE encoder, which captures continuous-time dynamics, supports interpolation at arbitrary times, and derives anomaly propagation velocity. Logs, traces, metrics, and events are fused by Transformer-based cross-attention with learned modality routing, where asymmetric inter-modal attention improves causal inference. The model further incorporates a variational information bottleneck, invariant-risk-minimization-based causal sufficiency constraints, temporal causal masking, contrastive log template clustering, and median-absolute-deviation normalization. On the Tianchi AIOps benchmark, HyperODE-RCA outperforms MicroRCA, CloudRanger, and recent causal discovery methods, while providing interpretable explanations through attention visualization over the learned hypergraph. The overall architecture is shown in Fig. 1 . Fig. 1: Multi-modal observability signals (logs, traces, metrics, entities, events) are encoded by modality-specific encoders, then processed through three core modules. IV Algorithm and Model IV-A HyperGAT: Hypergraph Attention Network Pairwise graph models are insufficient for microservice failures caused by coordinated multi-service interactions. HyperGAT addresses this with learned hyperedges and causal message passing, as illustrated in Fig. 2 . Fig. 2: HyperGAT module. (Center) Learned hypergraph with overlapping hyperedges capturing multi-service failure patterns. (Top-left) Differentiable hyperedge learning via Gumbel-Softmax with temperature annealing. (Top-right) Two-stage message passing: vertex-to-hyperedge aggregation followed by attention-weighted hyperedge-to-vertex distribution. (Bottom) Temporal causal regularization penalizing backward-in-time attention weights. IV-A 1 Hypergraph Representation We model the service topology as a hypergraph \mathcal{H}=(\mathcal{V},\mathcal{E}_{h}) with incidence matrix \mathbf{H}\in\{0,1\}^{|\mathcal{V}|\times|\mathcal{E}_{h}|} : H_{v,e}=\begin{cases}1&\text{if }v\in e\\ 0&\text{otherwise}\end{cases} (1) Hyperedges are inferred from data rather than predefined. IV-A 2 Differentiable Hyperedge Learning To enable gradient-based learning, hyperedge assignment is relaxed by Gumbel-Softmax. For candidate hyperedge e_{k} over service subset S_{k} , \ell_{k}=\mathbf{w}_{e}^{T}\cdot\text{MLP}_{\phi}\left(\frac{1}{|S_{k}|}\sum_{v_{i}\in S_{k}}\mathbf{h}_{v_{i}}\right) (2) and the soft incidence value is \tilde{H}_{v,e_{k}}=\frac{\exp((\ell_{k}+g_{1})/\tau)}{\exp((\ell_{k}+g_{1})/\tau)+\exp((-\ell_{k}+g_{2})/\tau)} (3) where g_{1},g_{2}\sim\text{Gumbel}(0,1) and \tau>0

Read the complete text