From IceCube to IT-Sphere: A Hybrid Quantum-Classical GNN for Banking IT Root Cause Analysis

AI summary

Overview

Research area: Hybrid quantum–classical machine learning — specifically a Quantum Graph Neural Network (QGNN) applied to IT operations analytics (root cause analysis) in banking.

Technical level: Intermediate. Readers should be comfortable with basic graph neural network concepts, variational quantum circuits, and standard classification metrics (F1, average precision, ROC-AUC).

Scope: The paper presents HQ-RCA, an industrially grounded workflow in which the classification head of the IceCube neutrino-reconstruction GNN (DynEdge) is replaced by a 4-qubit Variational Quantum Circuit, and benchmarks it against classical baselines on 13 months of anonymised alarm data from a major European bank.

What This Paper Is About

Banks continuously monitor their IT systems and generate a constant stream of alarms. Industrial correlation tools group these alarms into clusters, but they are not primarily designed to reconstruct the causal chain across interconnected services — so identifying the actual root cause is still largely done by hand. The authors reframe root cause analysis as a graph-reconstruction problem borrowed from IceCube neutrino physics: each alarm cluster is a graph, a background "log wave" is analogous to an atmospheric muon, and a rare "IT Event" hiding a single root-cause alarm is analogous to a neutrino interaction vertex. The goal is to test whether swapping in a quantum circuit as the classifier head can match or improve on classical models — and whether such a model can actually run on today's noisy quantum hardware.

Key Contributions

1. A hybrid QGNN architecture (HQ-RCA) that reuses DynEdge's four-EdgeConv-block backbone and replaces its native classification head with a 4-qubit VQC using a HEE2 encoding plus a RandomLayers ansatz (12 VQC parameters), followed by a linear softmax head reading out the single-qubit observable ⟨Z₀⟩.
2. A systematic encoding–ansatz sweep — an "L-shaped" search over 13 distinct configurations, plus two starting baselines (RY + StronglyEntanglingLayer ×1 with 12 PQC parameters, and a one-layer hardware-efficient embedding HEE1 with an SO(4) two-qubit gate with 24 parameters) — to disentangle the effects of encoding choice from ansatz choice.
3. A Dimensional Expressivity Analysis (DEA) of the readout, computing the thin SVD of the empirical observable Jacobian J = ∂⟨Z₀⟩/∂θ (a 200 × 12 matrix from 200 uniformly sampled inputs, rank counted above ε = 10⁻³), together with a comparison of the baseline readout against two richer readouts and a parameter-matched classical head.
4. A NISQ hardware demonstration on ibm_fez (IBM Heron r2, no error mitigation), including a simulated depolarising-noise study, a shot-budget sweep, and a threshold-recalibration procedure validated by 5-fold cross-validation.

Main Findings

• The hybrid QGNN matches the strongest classical baseline on F1. On the 2,000-graph subset, the hybrid QGNN matches standalone DynEdge on F1, while standalone DynEdge leads the ranking metrics (AP and ROC-AUC). The exact per-model bar heights are shown in Fig. 4 but are not stated numerically in the text for RF, XGB, and standalone DynEdge.

• The architecture sweep selected HEE2 + RandomLayers. The best configuration (12 VQC parameters) reached F1 = 0.931, AP = 0.953 and ROC-AUC = 0.974. Both the encoding axis and the ansatz axis strongly affected performance; baseline (0a) with RY encoding and one StronglyEntanglingLayer outperformed baseline (0b) with the HEE1 + SO(4) embedding, though the two differed in both axes simultaneously.

• Performance is stable at scale. Going from the 2k subset to the full 13k alarm clusters gave F1 = 0.915 ± 0.017, AP = 0.947 ± 0.020 and ROC-AUC = 0.973 ± 0.013 — each within 1σ of its 2k value.

• Rank-1 collapse at the deployed layout. DEA showed the observable Jacobian is rank-1: σ₁ = 1.336 while σ₂ … σ₁₂ sit at machine precision (∼10⁻¹⁶), well below ε = 10⁻³. Because the output probability depends on θ only through the scalar ⟨Z₀⟩, the Fisher rank equals rank(J) exactly, collapsing optimisation to a one-dimensional problem.

• The 1-D landscape is a broad plateau. A 40-point grid scan over one period showed F1(θ₀) is a 2π-periodic step function with a single broad plateau (∼37% of the period) peaking at θ₀* = 6.241 rad with F1 = 0.925. Freezing the other eleven parameters and tuning θ₀ alone reached that value; re-initialising θ₀ at random and retraining still reached F1 = 0.911 (start-dependent).

• Richer readouts do not help. ⟨Z₀⟩ gave the best held-out F1 (0.931 ± 0.009, rank 1) and the smallest variance, versus 0.916 ± 0.004 for the parametric single-qubit readout M(φ) = Σᵢ φᵢ⟨Zᵢ⟩ (4 parameters, rank 10) and 0.892 ± 0.031 for the two-local Z readout including ⟨ZᵢZⱼ⟩ terms (10 parameters, rank 10). AP and ROC-AUC stayed comparable across all three.

• Rank-1 is the exception, not the rule. Sweeping the RandomLayers seed over its first 21 values (0–20), the ⟨Z₀⟩ rank ranged from 0 to 9 (mean 3.3). Across those 21 layouts, F1 stayed high and no correlation between F1 and rank was observed (Pearson −0.08, p = 0.73).

• The quantum head beats a parameter-matched classical head on F1, but loses on ranking metrics. A classical head with the same parameter count (Linear + tanh → scalar) reached F1 = 0.876 ± 0.063 at τ = 0.5 (AP/ROC-AUC 0.961/0.986); recalibrating only its threshold (τ* ≈ 0.76, 5-fold) lifted F1 to 0.907 ± 0.020, still below the VQC's 0.931 ± 0.009 — so its F1 deficit is not merely a calibration artefact. The classical head instead leads the threshold-free AP/ROC-AUC.

• Noise tolerance in simulation. Under simulated depolarising noise, F1 stayed ≥ 0.90 up to a 5% two-qubit gate-error rate (p₂q = 0.05). A shot-noise sweep at p₂q = 0.01 fixed 200 shots/node, giving F1 = 0.918, near the 0.922 plateau reached at ≥ 500 shots.

• Hardware results require threshold recalibration. With θ₀ at its optimum, the model ran on ibm_fez in ∼2 minutes. Device noise shifted the score distribution: at τ = 0.5 the model over-predicted and F1 fell to 0.587, while threshold-invariant ranking metrics were largely preserved (AP 0.927, ROC-AUC 0.957, versus 0.960/0.977 in simulation). Recalibrating only the threshold (τ* ≈ 0.68) with 5-fold cross-validation recovered F1 = 0.915 ± 0.037, matching the noisy-simulator value of ∼0.92. Noise shifts the optimal threshold more than it shifts class separability.

• The claim is feasibility, not advantage. The authors explicitly state that HQ-RCA is a hardware-executable, quantum-assisted root-cause pipeline rather than a demonstration of quantum advantage. The model choice is metric-driven: the hybrid head is justified when deployment optimises thresholded detection (F1), such as binary escalation of an alarm cluster, while ranked alarm triage favours standalone DynEdge (AP/AUC).

Methodology in Plain English

The team took an existing graph neural network built for particle physics — DynEdge, designed to reconstruct neutrino events in IceCube by updating each detector node from its neighbours — and repurposed it as the classical engine for a banking diagnosis task. The analogy is structural, not literal: a photon sensor in the detector becomes a monitoring-system sensor unit, an atmospheric muon background becomes a routine log wave, and a neutrino interaction becomes a genuine IT incident hiding a root cause, with the interaction vertex mapped to the root-cause node.

Concretely, alarm clusters are turned into graphs using Union–Find grouping, node features are compressed from 8 to 7 dimensions with PCA (retaining 95% of variance), and four EdgeConv blocks produce a 1031-dimensional node embedding (1031 = 7 + 4 × 256). A linear layer with tanh activation squeezes this into four numbers in (−1, +1), which are encoded into a 4-qubit variational circuit. The circuit's ansatz is trained, and the only thing read out is the expectation of the Pauli-Z observable on the first qubit, ⟨Z₀⟩. A small linear softmax layer converts that single scalar into a root-cause probability.

To decide which circuit design to use, the authors ran a structured two-stage sweep — first holding the encoding fixed and trying nine ansatz variants, then holding the best ansatz fixed and trying five encodings — rather than a full grid. To understand how many of the 12 circuit parameters actually matter, they borrowed Dimensional Expressivity Analysis: they measured how sensitively the readout responds to each parameter across 200 random inputs, assembled this into a 200 × 12 Jacobian matrix, and looked at its singular values. Because only one singular value rose above the cutoff, they could scan a single angle on a 40-point grid and solve the optimisation without gradients.

Finally, they tested the trained model under simulated gate noise and shot noise, chose a 200-shots-per-node budget, and executed it on IBM's Heron r2 device with no error mitigation. Because hardware noise shifted the output score distribution, they recalibrated only the decision threshold using cross-validation on the device, rather than retraining.

Why This Matters

Impact on research. The paper offers a rare end-to-end, industry-data demonstration of a hybrid quantum-classical graph model, and — more usefully — a critical one. Its central methodological point is that the effective rank of the quantum readout had no measurable correlation with F1 (Pearson −0.08, p = 0.73), so richer, higher-rank readouts bought nothing. This is a concrete counterexample to the assumption that more quantum expressivity yields better performance, and it provides a reusable diagnostic (observable-Jacobian DEA) for deciding when a circuit can be collapsed to a one-dimensional, gradient-free problem.

Real-world applications.

• Automated escalation of IT incidents: when the operational goal is a binary decision to page an engineer, F1 is the right metric and the quantum head outperformed a parameter-matched classical head (0.931 ± 0.009 versus 0.907 ± 0.020 after threshold recalibration).
• Ranked alarm triage: when operators want a prioritised queue rather than a hard decision, standalone DynEdge's stronger AP/AUC makes it the better fit.
• Threshold recalibration practice for quantum deployments: the finding that noise damages calibration far more than separability (F1 0.587 → 0.915 while AP moved only 0.960 → 0.927) is directly transferable to any NISQ classifier deployed with a decision threshold.
• Cross-domain transfer of scientific ML: the paper models how a physics-grade architecture can be re-targeted at enterprise telemetry with a structural analogy rather than a rewrite.

Industry relevance. The work comes from a collaboration between University of Siena, INFN Pisa, and IBM Consulting—Promontory (Banking & Financial Markets), and is accepted for the poster track at IEEE QCE26. It speaks directly to the practical question enterprise adopters ask: not "is there quantum advantage?" but "can this run on today's hardware, at what cost, and under which operating metric?" The answer here — ∼2 minutes on ibm_fez, no error mitigation, competitive F1 after threshold recalibration, on 13k real alarm clusters — is an unusually concrete data point.

Future Directions

• Other embedding families. The authors list alternative embedding strategies as the first open item, motivated by the fact that both the encoding axis and the ansatz axis strongly affected performance in their 13-configuration sweep.
• A new domain: mainframe transactions. Extending HQ-RCA beyond distributed IT monitoring to mainframe transaction data is named explicitly as future work, which would test whether the graph-reconstruction framing generalises.
• Rethinking what to optimise. The rank-versus-F1 null result raises an unresolved question: if higher observable rank does not improve discrimination, what architectural property should circuit designers target instead?
• Closing the AP/AUC gap. The hybrid head matched standalone DynEdge on F1 but trailed on ranking metrics, and the parameter-matched classical head led on both AP and ROC-AUC. Whether a quantum head can match classical ranking performance — rather than only thresholded detection — remains open.

Target Audience

This paper is best suited to researchers and practitioners working at the boundary of quantum machine learning and enterprise IT operations: QML researchers who want an honest, industry-scale evaluation rather than a toy benchmark; machine-learning engineers in banking and financial services evaluating whether hybrid quantum models are deployable; and physicists or scientific-computing specialists interested in how domain-specific architectures like DynEdge transfer to adjacent problem structures. It is also useful for technical decision-makers who need to understand that model selection in this space is metric-dependent — a quantum head can be the right operational choice under F1 while a classical model remains the right choice under AP/AUC.

Authors’ abstract

We present Hybrid Quantum Root Cause Analysis (HQ-RCA), an industrially grounded workflow for root cause analysis in banking IT operations, built on a hybrid Quantum Graph Neural Network (QGNN): the classical backbone of DynEdge (the IceCube neutrino-reconstruction GNN, which we call standalone DynEdge), with its classification head replaced by a Variational Quantum Circuit (VQC). On 13 months of anonymised IT data (13k alarm clusters) from a major European bank, the hybrid QGNN matches standalone DynEdge -- the strongest classical baseline -- on $F_1$, while standalone DynEdge leads the ranking metrics. A readout-sensitivity and layout-robustness study, analysed via Dimensional Expressivity Analysis (DEA), shows that the effective parameter dimensionality (rank) of the quantum observable has no measurable correlation with $F_1$; we therefore keep the simplest readout $\langle Z_0\rangle$ (the Pauli-$Z$ expectation on the first qubit), which in the deployed layout is rank-1, collapsing optimisation to a 1-D problem solvable by a gradient-free grid scan. Execution on IBM Heron r2 (no error mitigation) shows this gradient-free readout is executable on NISQ hardware after threshold recalibration.