🤖 machine learning

Root Cause Analysis of Measurement and Mechanistic Anomalies

This paper proposes a causal model and an efficient inference procedure to distinguish between measurement errors and mechanistic shifts in anomaly detection, enabling accurate root cause localization and anomaly type classification with state-of-the-art performance.

Original authors: Hendrik Suhr, David Kaltenpoth, Jilles Vreeken

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hendrik Suhr, David Kaltenpoth, Jilles Vreeken

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery in a factory. Something has gone wrong, and the data looks weird. But there's a catch: not all "weird" is created equal.

Sometimes, the machine is working perfectly, but the thermometer is broken and giving a fake reading. Other times, the thermometer is fine, but the machine itself has actually changed its behavior (maybe a gear slipped, or the fuel mix is wrong).

The paper by Suhr, Kaltenpoth, and Vreeken introduces a new detective tool called CALI (Causal Anomaly Localization and Inspection). Its superpower is figuring out which of these two scenarios is happening and where exactly the problem started.

Here is the breakdown of their work in simple terms:

1. The Two Types of "Weird"

The authors point out that most existing tools just say, "Hey, this number is weird!" They don't tell you why. The paper splits "weirdness" into two distinct categories:

Measurement Anomalies (The Broken Ruler): Imagine you are measuring a table. The table is perfectly normal, but your ruler is bent, so you write down the wrong number. The "truth" hasn't changed; only the recording of it did. In a factory, this is like a sensor glitch.
Mechanistic Anomalies (The Broken Machine): Imagine the table itself has warped because the wood was wet. The ruler is perfect, but the object being measured has genuinely changed. In a factory, this is a real process failure, like a valve sticking or a chemical reaction going off.

Why does this matter? If you have a broken ruler, you just fix the sensor. If you have a warped table, you need to stop the machine and fix the process. Mixing them up leads to wasted time and money.

2. The Detective's Map (The Causal DAG)

To solve this, CALI uses a "map" of how things influence each other. Think of this as a flowchart of cause-and-effect.

Example: Rain $\rightarrow$ Wet Grass $\rightarrow$ Slippery Road.

If the Road is slippery, is it because:

The sensor measuring the road is broken (Measurement)?
It actually rained (Mechanistic)?

CALI looks at the whole map. If the Grass is dry but the Road says it's slippery, CALI knows the road sensor is lying (Measurement). If the Grass is wet and the Road is slippery, the rain caused it (Mechanistic).

3. How CALI Works (The "What If" Game)

The tool doesn't just guess; it plays a game of "What If?" using a concept called Interventions.

The Setup: It learns how the system usually behaves (the "clean" process).
The Test: When it sees a weird data point, it asks: "If I pretend this specific sensor is broken, does the rest of the data make sense?" OR "If I pretend the machine process changed here, does the rest of the data make sense?"
The Winner: It picks the explanation that requires the fewest assumptions (a principle called "Sparse Mechanism Shift"). It prefers the idea that "one sensor broke" over the idea that "three different parts of the machine changed at once."

4. The Results: A Sharp Detective

The authors tested CALI on two types of data:

Fake Data: They built computer simulations where they knew exactly what was broken. CALI was incredibly accurate at finding the broken part and correctly identifying if it was a sensor error or a machine error.
Real Data: They tested it on:
- Biological Data: Protein interactions in cells (Sachs dataset).
- Physics Data: A controlled light tunnel experiment (Causal Chambers).
- Taxi Data: New York City taxi trips.

The Taxi Case Study:
The authors found some truly weird taxi trips.

Measurement Error: A trip that was 9 miles long but took 1 minute and cost $2.50. The math didn't add up. CALI correctly flagged this as a measurement error (the distance or time was recorded wrong).
Mechanistic Anomaly: A trip that cost $52 flat, regardless of distance. CALI realized this wasn't a glitch; it was a mechanistic change (a special flat-rate fare for JFK airport).

5. The Bottom Line

The paper claims that CALI is the first tool that can:

Locate the exact root cause (which variable is the problem).
Classify the type of problem (is it a broken sensor or a broken process?).
Do this without needing a separate "clean" dataset to learn from (it figures it out from the messy data itself).

In short, CALI helps you decide whether to calibrate your sensor or fix your machine, saving you from making the wrong repair.

Technical Summary: Root Cause Analysis of Measurement and Mechanistic Anomalies

Problem Statement

Real-world datasets frequently contain anomalies—samples generated by processes that deviate from the norm. While existing Root Cause Analysis (RCA) methods aim to identify which features are responsible for an anomaly, they typically fail to distinguish between two fundamentally different generative processes:

Measurement Anomalies: The underlying causal process is normal, but one or more values are recorded incorrectly (e.g., sensor malfunctions, data entry errors).
Mechanistic Anomalies: The causal mechanism generating the sample has genuinely changed (e.g., a shift in the physical process or system dynamics).

Distinguishing between these is critical for practical applications. A measurement error may require data correction or sensor recalibration, whereas a mechanistic shift often necessitates intervention in the system itself. Existing methods blur this distinction, often treating all outliers as deviations in the observed data without modeling the latent causal state. Furthermore, many approaches assume a clean training set, whereas in practice, anomalies must often be identified directly within a contaminated dataset (unsupervised in-sample setting).

Methodology

Causal Formalization

The authors propose a causal framework that explicitly models both anomaly types as interventions on a Structural Causal Model (SCM).

Variables: The model distinguishes between latent causal state variables ( $X^*$ ) and their observed readouts ( $X$ ).
Anomaly-Free Process: In the absence of anomalies, $X^*_j = f_j(Pa^*_j, U_j)$ and $X_j = X^*_j$ , where $f_j$ is the structural assignment and $U_j$ is independent noise.
Mechanistic Anomalies: Modeled as a hard intervention on the latent variable $X^*_j$ . The structural assignment is replaced by an exogenous outlier variable $O^{mech}_j$ . This perturbation propagates to downstream variables.
Measurement Anomalies: Modeled as a hard intervention on the observed variable $X_j$ only. The latent state $X^*_j$ remains unchanged, and the perturbation does not propagate downstream.
Sparse Mechanism Shift (SMS): The framework adopts the SMS hypothesis, positing that explanations requiring fewer simultaneous mechanism failures are more likely. This serves as an inductive bias to prefer simple explanations (e.g., a single measurement error) over complex concerted mechanism changes.

Inference and Identifiability

The problem is cast as a latent-variable maximum likelihood estimation task.

Assignment-Conditional Densities: The likelihood of a sample given a specific assignment of anomalies (which variables are intervened upon and how) is derived. For measurement anomalies, this requires marginalizing over the unobserved latent pre-replacement values.
Identifiability: The paper establishes conditions under which the reduced intervention probabilities are identifiable.
- For non-sink nodes, mechanistic and measurement probabilities are separately identifiable.
- For sink nodes (variables with no descendants in the latent graph), only the probability of any intervention (mechanistic or measurement) is identifiable, as both types produce identical observable distributions at the sink.
- Generic identifiability is proven for linear Gaussian SCMs and fixed-degree polynomial Gaussian SCMs.
Optimization: Since exact marginalization over all possible anomaly assignments is computationally infeasible (exponential in the number of variables), the authors introduce a hard-assignment approximation. They maximize the likelihood of the single most probable assignment rather than summing over all assignments. This is justified when outliers are strong enough to induce well-separated assignment-conditional densities.

The CALI Algorithm

The authors develop CALI (Causal Anomaly Localization and Inspection), a practical algorithm that:

Estimates the Clean Model: Uses robust regression (quantile regression with generalized additive models) and trimmed kernel density estimation to learn the anomaly-free SCM and noise distributions from contaminated data.
Greedy Assignment: Iteratively updates the anomaly assignment for each sample. It prioritizes candidates with low residual likelihood under the clean density, accepting an intervention only if the gain in likelihood outweighs the sparsity cost (penalty for adding an outlier).
Output: Returns a discrete assignment identifying root causes and their types (measurement vs. mechanistic), along with confidence scores based on likelihood gains.

Key Contributions

Formal Causal Model: The first formal definition of measurement and mechanistic anomalies within a causal framework, treating them as distinct interventions on latent and observed variables.
Identifiability Characterization: A theoretical characterization of when the reduced intervention probabilities can be recovered from data, specifically addressing the indistinguishability of anomaly types at sink nodes.
Practical Algorithm (CALI): An efficient inference procedure that achieves state-of-the-art performance in both localizing root causes and classifying anomaly types in an unsupervised, in-sample setting.

Experimental Results

The authors evaluate CALI on synthetic data and real-world benchmarks (Sachs protein signaling, Causal Chambers, and NYC Taxi data).

Root Cause Localization: CALI achieves state-of-the-art Top-k Recall on synthetic data, closely matching the performance of an oracle exhaustive search (OPT*). On real-world datasets, it outperforms existing RCA methods (including SCM-based attribution, SMTR, SCOR, and AE+SHAP), particularly under medium to strong outlier strengths.
Anomaly Type Classification: CALI significantly outperforms baselines (like MARG, which uses marginal outlierness) in distinguishing between measurement and mechanistic anomalies. It maintains high accuracy even with weak outliers where other methods struggle.
Robustness: The method demonstrates robustness to violations of assumptions, including correlated root causes, soft interventions, DAG misspecification, confounding, and heteroscedastic noise.
Case Study: On the NYC Taxi dataset, CALI successfully identified plausible measurement errors (e.g., impossible trip distances vs. fares) and mechanistic shifts (e.g., specific flat-fare pricing rules), demonstrating actionable insights.

Significance and Claims

The paper claims that distinguishing between measurement errors and mechanistic shifts is a critical, previously overlooked gap in anomaly analysis. By modeling these as distinct causal interventions, the proposed method enables more appropriate remediation strategies (e.g., fixing a sensor vs. fixing a process).

The authors emphasize that their approach is modest regarding its limitations:

It requires a known or learned DAG; performance degrades if the causal structure is severely misspecified.
It assumes anomalies are rare and that the "clean" process dominates the data.
For extremely large shifts, simpler marginal methods may become competitive, as the model relies on density estimates that may extrapolate poorly.
The method is currently restricted to tabular data.

The authors position CALI as a decision-support tool rather than an automated decision-maker, noting that its conclusions depend on the validity of the assumed causal graph and modeling assumptions. They suggest future work could leverage the structure of anomalies to improve causal discovery in contaminated data.