Analog Quantum Feature Selection with Neutral-Atom Quantum Processors

TL;DR

This work addresses the NP-hard problem of feature selection by presenting a quantum-native approach that uses analog simulation on neutral-atom Rydberg arrays to balance feature relevance and redundancy. Relevance to the target is encoded as local detunings and redundancy as distance-dependent Rydberg interactions, with the objective implemented in a programmable Hamiltonian and solved via adiabatic evolution. Data embedding maps mutual information structure into a 2D atomic layout through MDS, enabling physically faithful encoding of correlations with a blockade constraint. Empirical evaluation on three real-world binary classification datasets shows competitive or superior performance to classical baselines, especially for compact feature subsets (2–4 features), highlighting potential for scalable, interpretable quantum-assisted preprocessing in ML pipelines.

Abstract

We present a quantum-native approach to quantum feature selection (QFS) based on analog quantum simulation with neutral atom arrays, adaptable to a variety of academic and industrial applications. In our method, feature relevance-measured via mutual information with the target-is encoded as local detuning amplitudes, while feature redundancy is embedded through distance-dependent van der Waals interactions, constrained by the Rydberg blockade radius. The system is evolved adiabatically toward low-energy configurations, and the resulting measurement bitstrings are used to extract physically consistent subsets of features. The protocol is evaluated through simulations on three benchmark binary classification datasets: Adult Income, Bank Marketing, and Telco Churn. Compared to classical methods such as mutual information ranking and Boruta, combined with XGBoost and Random Forest classifiers, our quantum-computing approach achieves competitive or superior performance. In particular, for compact subsets of 2-5 features, analog QFS improves mean AUC scores by 1.5-2.3% while reducing the number of features by 75-84%, offering interpretable, low-redundancy solutions. These results demonstrate that programmable Rydberg arrays offer a viable platform for intelligent feature selection with practical relevance in machine learning pipelines, capable of transforming computational quantum advantage into industrial quantum usefulness.

Figures (5)

• Figure 1: Driving profiles for selected schedules: the global detuning $\Delta_g(t)$ initializes the system, $\Omega(t)$ controls exploration, and $\Delta_l(t)$ encodes feature relevance.
• Figure 2: Mapping redundancy onto atom positions for one dataset. The algorithm places highly redundant features within blockade range while keeping independent features farther apart.
• Figure 3: Relative error matrices $\epsilon_{ij}$ for the MDS embeddings of (left) Adult, (center) Bank Marketing, and (right) Telco Churn. Darker colors indicate smaller deviation between target and embedded distances.
• Figure 4: Percentage of coincident features between QFS-selected subsets and classical baselines as a function of subset size $n$.
• Figure 5: Median values of AUC, precision, and recall across the three datasets as a function of the number of selected features ($n_{\text{var}}$). The plots highlight that quantum feature selection (QFS) performs particularly well in the early stages of variable selection, achieving strong classification performance with small subsets of features.