Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust Root Cause Diagnosis using In-Distribution Interventions

Lokesh Nagalapatti, Ashutosh Srivastava, Sunita Sarawagi, Amit Sharma

TL;DR

Root-cause diagnosis in complex systems is challenged by anomalies that lie outside training distributions. The paper introduces In-Distribution Interventions (IDI), which diagnose root causes by intervening on suspected nodes within the in-distribution regime, avoiding the unreliable abduction steps of counterfactuals. The authors provide theoretical bounds showing interventional errors depend on exogenous noise variance rather than distribution shift, and they validate IDI across toy, PetShop, and synthetic SCMs, where IDI consistently outperforms state-of-the-art baselines. The work demonstrates practical robustness for RCD in cloud deployments and releases code for broader use.

Abstract

Diagnosing the root cause of an anomaly in a complex interconnected system is a pressing problem in today's cloud services and industrial operations. We propose In-Distribution Interventions (IDI), a novel algorithm that predicts root cause as nodes that meet two criteria: 1) **Anomaly:** root cause nodes should take on anomalous values; 2) **Fix:** had the root cause nodes assumed usual values, the target node would not have been anomalous. Prior methods of assessing the fix condition rely on counterfactuals inferred from a Structural Causal Model (SCM) trained on historical data. But since anomalies are rare and fall outside the training distribution, the fitted SCMs yield unreliable counterfactual estimates. IDI overcomes this by relying on interventional estimates obtained by solely probing the fitted SCM at in-distribution inputs. We present a theoretical analysis comparing and bounding the errors in assessing the fix condition using interventional and counterfactual estimates. We then conduct experiments by systematically varying the SCM's complexity to demonstrate the cases where IDI's interventional approach outperforms the counterfactual approach and vice versa. Experiments on both synthetic and PetShop RCD benchmark datasets demonstrate that \our\ consistently identifies true root causes more accurately and robustly than nine existing state-of-the-art RCD baselines. Code is released at https://github.com/nlokeshiisc/IDI_release.

Robust Root Cause Diagnosis using In-Distribution Interventions

TL;DR

Root-cause diagnosis in complex systems is challenged by anomalies that lie outside training distributions. The paper introduces In-Distribution Interventions (IDI), which diagnose root causes by intervening on suspected nodes within the in-distribution regime, avoiding the unreliable abduction steps of counterfactuals. The authors provide theoretical bounds showing interventional errors depend on exogenous noise variance rather than distribution shift, and they validate IDI across toy, PetShop, and synthetic SCMs, where IDI consistently outperforms state-of-the-art baselines. The work demonstrates practical robustness for RCD in cloud deployments and releases code for broader use.

Abstract

Diagnosing the root cause of an anomaly in a complex interconnected system is a pressing problem in today's cloud services and industrial operations. We propose In-Distribution Interventions (IDI), a novel algorithm that predicts root cause as nodes that meet two criteria: 1) **Anomaly:** root cause nodes should take on anomalous values; 2) **Fix:** had the root cause nodes assumed usual values, the target node would not have been anomalous. Prior methods of assessing the fix condition rely on counterfactuals inferred from a Structural Causal Model (SCM) trained on historical data. But since anomalies are rare and fall outside the training distribution, the fitted SCMs yield unreliable counterfactual estimates. IDI overcomes this by relying on interventional estimates obtained by solely probing the fitted SCM at in-distribution inputs. We present a theoretical analysis comparing and bounding the errors in assessing the fix condition using interventional and counterfactual estimates. We then conduct experiments by systematically varying the SCM's complexity to demonstrate the cases where IDI's interventional approach outperforms the counterfactual approach and vice versa. Experiments on both synthetic and PetShop RCD benchmark datasets demonstrate that \our\ consistently identifies true root causes more accurately and robustly than nine existing state-of-the-art RCD baselines. Code is released at https://github.com/nlokeshiisc/IDI_release.
Paper Structure (29 sections, 4 theorems, 37 equations, 12 figures, 9 tables, 1 algorithm)

This paper contains 29 sections, 4 theorems, 37 equations, 12 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem Formulation
  4. Training Setup
  5. Qualifying Criteria for Root Cause
  6. Interventions vs. Counterfactuals for RCD
  7. Our Approach: IDI
  8. Experiments
  9. Toy Experiments
  10. Comparison with Baselines
  11. Experiments on PetShop: RCD in a Deployed Cloud System
  12. Experiments on Synthetic SCMs
  13. Experiment on IDI vs. IDI (using CFs)
  14. Limitations and Conclusion
  15. Acknowledgements
  16. ...and 14 more sections

Key Result

Theorem 5

Suppose the oracle SCM $\mathcal{S}$ is an additive noise model over a chain graph $\mathcal{G} = X_1 \rightarrow \cdots \rightarrow X_n$, with structural equations of the form $f_i(x_{i-1}) + \epsilon_i$, where each $\epsilon_i$ has bounded variance $\sigma^2$, and each function $f_i$ is $K$-Lipsch

Figures (12)

  • Figure 1: This figure demonstrates how abduction errors in counterfactuals impact RCD performance. In panel (a), an instance ${\bm{x}}$ shows an anomaly at the purple target node $x_4$, with the root cause being the gold node $x_1$, which is affected by an abnormal intervention. Hence $x_1$ takes OOD values, and influences its downstream nodes also to take on OOD values. Latent exogenous nodes are shown in grey. Panel (b) illustrates the true CF ${\bm{x}}^{\text{CF}(1)}$ obtained by applying the fix. Panel (c) presents the estimated CF for the fix, using exogenous estimates$\hat{\epsilon}$ -- involving abduction that requires SCM evaluations in OOD regions. Finally, panel (d) shows the interventional estimate, which uses sampled$\tilde{\epsilon}$, yielding a resulting $\widehat{x}^{\text{int}(1)}_4$ that conforms to the usual regime. Our theory in Sec \ref{['sec:theory']} captures this rigorously.
  • Figure 2: IDI's pipeline: During training, we use samples $D_{\text{trn}}$, a causal graph, and anomaly thresholds $\tau_i$ to learn anomaly detectors $\phi_i$, and structural equations $f_i$ as part of SCM training. During inference, we are given a root cause test case. The example illustrates two nodes $X_1, X_4$ with abnormal interventions. The edge weights indicate the strength of parent's influence (e.g., $\hat{f_5}(x_2, x_4) = x_2 + 0.01 x_4$). While applying the anomaly condition, IDI discards $X_2$ in $\mathcal{R}_{\text{cand}}$ because of its anomalous parent $X_1$. Finally, the fix conditions excludes $X_4$ because it is insufficient to restore $X_5$ to its usual value. IDI declares $X_1$ as the root cause.
  • Figure 3: A chain graph with more than one root cause in a simple path. Dotted lines denote a directed path.
  • Figure 4: An example graph with more than one root cause.
  • Figure 5: Assessing the impact of variance of $\epsilon_i$ using a four variable additive noise toy dataset.
  • ...and 7 more figures

Theorems & Definitions (9)

  • Definition 1
  • Definition 2
  • Definition 3
  • Definition 4
  • Theorem 5
  • Theorem 6
  • Corollary 7
  • Definition 8: dabook
  • Lemma 9