Inspiration
The Atomiq challenge connects three practical questions: who is likely to subscribe, which six customers should a constrained campaign contact, and how likely is the campaign to achieve at least three subscriptions? Our project connects quantum learning, optimization, and simulation in one reproducible workflow with classical benchmarks.
What it does
The pipeline learns from 300 labeled training customers and predicts subscription probabilities for 100 test customers. It ranks customers using the official tie-breaking rule, selects exactly six from the top ten under budget and capacity constraints, and estimates the probability of at least three subscriptions.
Selected customers: C0005, C0007, C0009, C0020, C0022, C0036.
The selected campaign has expected net value 6.1375, uses budget 9/25 and capacity 19/22, and achieves a quantum-simulated conversion probability of 0.022796630859375, approximately 2.28%.
How we built it
Stage 1 uses training-only preprocessing and three-fold out-of-fold validation. We compare classical models with a four-qubit Qiskit fidelity-kernel QSVC. The quantum model uses a disclosed 64-row stratified training subset and was selected by the lowest out-of-fold Brier score, approximately 0.1000.
Stage 2 uses ten-qubit, two-layer XY-QAOA. Its mixer preserves the required six-customer cardinality. We compare measured candidates with exhaustive enumeration of all 210 possible six-person subsets. The measured QAOA result reached the exact optimum on this instance; the final selection explicitly uses classical exact certification.
Stage 3 encodes the six unchanged Stage 1 probabilities using RY rotations on six qubits. We run 32,768 ideal shots and compare the estimated probability of at least three subscriptions against exact probability enumeration and Monte Carlo simulation.
All quantum experiments use simulators. We do not claim hardware execution or quantum advantage.
Noise and validation
We also run 32,768 shots with a synthetic independent readout-error model and apply readout mitigation. The noisy probability is approximately 0.03445, the mitigated estimate is approximately 0.02242, and the exact analytical probability is approximately 0.02405.
The repository includes circuit depths, gate counts, shot counts, documented seeds, pinned dependencies, classical comparisons, repeated QAOA experiments, sensitivity analysis, and 44 passing tests.
Challenges and limitations
The training set is small and imbalanced. The selected feature representation produces a highest-score tie among 86 test customers, making the official alphanumeric tie-breaking rule consequential. We disclose this limitation rather than interpreting the ranking as strong evidence of individual customer differences.
Exact enumeration is inexpensive for this small optimization instance. QAOA matching its optimum demonstrates correctness on this instance, not a computational speedup.
The conversion simulation assumes independent customer outcomes and depends on the quality of Stage 1 probabilities. Its sampling uncertainty does not capture all predictive uncertainty.
What we learned
A useful quantum-classical workflow needs consistent probabilities across stages, explicit feasibility checks, classical reference solutions, and transparent reporting of noise and model limitations. Preserving cardinality in the mixer helps make sampled optimization candidates directly interpretable.
What's next
Future work would improve probability calibration and feature representation, evaluate larger optimization instances, test correlated customer outcomes, and assess hardware execution with realistic device noise.
Reproducibility
The public repository contains runnable code, dependency versions, documented seeds, the three root-level result files, technical reports, and presentation slides.
Repository: https://github.com/shylankamesh/atomiq-quantum-campaign-2026
Built With
- aer
- learning
- machine
- numpy
- pandas
- python
- qiskit
- scikit-learn
- scipy
