Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Are GNNs Actually Effective for Multimodal Fault Diagnosis in Microservice Systems?

Fei Gao, Ruyue Xin, Xiaocui Li, Yaqiang Zhang

TL;DR

The paper tackles whether graph neural networks are truly beneficial for fault diagnosis in microservice systems by introducing DiagMLP, a topology-agnostic baseline that preserves multimodal fusion while removing explicit graph modeling. Through ablations across five datasets, DiagMLP matches or surpasses state-of-the-art GNN-based methods on fault detection, localization, and classification, suggesting that preprocessing and embedding pipelines already encode critical dependency information. The authors visualize embeddings to show that topology information adds little discriminative power beyond the rich multimodal features, challenging the assumed value of graph structures. They advocate rigorous baselines, standardized preprocessing, and larger, more explicit datasets to accurately evaluate architectural innovations in fault-diagnosis research.

Abstract

Graph Neural Networks (GNNs) are widely adopted for fault diagnosis in microservice systems, premised on their ability to model service dependencies. However, the necessity of explicit graph structures remains underexamined, as existing evaluations conflate preprocessing with architectural contributions. To isolate the true value of GNNs, we propose DiagMLP, a deliberately minimal, topology-agnostic baseline that retains multimodal fusion capabilities while excluding graph modeling. Through ablation experiments across five datasets, DiagMLP achieves performance parity with state-of-the-art GNN-based methods in fault detection, localization, and classification. These findings challenge the prevailing assumption that graph structures are indispensable, revealing that: (i) preprocessing pipelines already encode critical dependency information, and (ii) GNN modules contribute marginally beyond multimodality fusion. Our work advocates for systematic re-evaluation of architectural complexity and highlights the need for standardized baseline protocols to validate model innovations.

Are GNNs Actually Effective for Multimodal Fault Diagnosis in Microservice Systems?

TL;DR

The paper tackles whether graph neural networks are truly beneficial for fault diagnosis in microservice systems by introducing DiagMLP, a topology-agnostic baseline that preserves multimodal fusion while removing explicit graph modeling. Through ablations across five datasets, DiagMLP matches or surpasses state-of-the-art GNN-based methods on fault detection, localization, and classification, suggesting that preprocessing and embedding pipelines already encode critical dependency information. The authors visualize embeddings to show that topology information adds little discriminative power beyond the rich multimodal features, challenging the assumed value of graph structures. They advocate rigorous baselines, standardized preprocessing, and larger, more explicit datasets to accurately evaluate architectural innovations in fault-diagnosis research.

Abstract

Graph Neural Networks (GNNs) are widely adopted for fault diagnosis in microservice systems, premised on their ability to model service dependencies. However, the necessity of explicit graph structures remains underexamined, as existing evaluations conflate preprocessing with architectural contributions. To isolate the true value of GNNs, we propose DiagMLP, a deliberately minimal, topology-agnostic baseline that retains multimodal fusion capabilities while excluding graph modeling. Through ablation experiments across five datasets, DiagMLP achieves performance parity with state-of-the-art GNN-based methods in fault detection, localization, and classification. These findings challenge the prevailing assumption that graph structures are indispensable, revealing that: (i) preprocessing pipelines already encode critical dependency information, and (ii) GNN modules contribute marginally beyond multimodality fusion. Our work advocates for systematic re-evaluation of architectural complexity and highlights the need for standardized baseline protocols to validate model innovations.
Paper Structure (24 sections, 3 equations, 5 figures, 4 tables)

This paper contains 24 sections, 3 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. GNN-Based Fault Diagnosis Methods
  3. Preprocessing
  4. Embedding
  5. GNN-based Modeling
  6. Observations and Motivation
  7. A Topology-Agnostic Baseline
  8. Problem Formulation
  9. A Motivating Example
  10. Topology-agnostic DiagMLP
  11. Design Choices of DiagMLP
  12. Learnable Position Embeddings
  13. Single-Layer MLP
  14. Fusion by Concatenation
  15. Experiments
  16. ...and 9 more sections

Figures (5)

  • Figure 1: Overview of GNNs ablation methods. (a) shows fault diagnosis using GNNs as the backbone to model the service dependency graphs. (b) shows the model with GNNs replaced by a topology-agnostic MLPs while keeping the other components.
  • Figure 2: Pipeline of Existing GNN-Based Multimodal Fault Diagnosis Models. (a) Preprocessing transforms raw multimodal data (metrics, logs, traces) into standardized formats, such as time series, text, and graphs; (b) Embedding modules, often pretrained, encode these features into dense representations; (c) GNN modules, varying across methods, integrate multimodal features with dependency graphs for diagnosis.
  • Figure 3: Motivating Example. UMAP visualization of preprocessed multimodal features from the SN and TT datasets Eadro. Each point represents a fault window, colored by its root cause, showing clear clustering without topology modeling.
  • Figure 4: DiagMLP Architecture. DiagMLP (➁+➂) replaces GNN modules, positioned between any embedding module (➀) and downstream task module (➃), processing multimodal features without topology modeling.
  • Figure 5: UMAP Visualization of GAIA Dataset. Fault windows are depicted using node-wise Original Features (from FastText), MLP Embeddings (from DiagMLP), and GNN Embeddings (from TVDiag).