Quantum Feature Engineering for Credit Default Prediction

Appending 16 features derived from an 8-qubit Instantaneous Quantum Polynomial-time (IQP) circuit improves the F1 score of a logistic regression model for credit default prediction from 0.462 to 0.517. Researchers W.J. Zeng and others, in their paper 'Quantum Feature Engineering for Credit Default Prediction: When and Why IQP Circuits Help Linear Classifiers', identified this gain by comparing the quantum-augmented linear model against standard classical baselines, including Kernel Principal Component Analysis (KPCA), using the UCI Default of Credit Card Clients dataset. This result demonstrates that quantum feature engineering provides a measurable performance boost specifically for this credit dataset when using linear classifiers.## The Mechanism of Quantum Feature Encoding

IQP circuits use a sequence of Hadamard gates, controlled-phase rotations, and single-qubit rotations to map input variables into a high-dimensional feature space. To convert a classical number like a credit balance into a quantum state, the value is first normalized to a range between 0 and 2π. This value is then assigned as the rotation angle θ for an Rx gate, which rotates the qubit state by that specific amount. The controlled-phase gates are central to this process because they introduce entanglement between qubits, which corresponds to the interaction terms between input features.

By applying these multi-qubit interactions, the circuit maps non-linear correlations into explicit, separate feature columns based on the expectation values of the quantum operators. The logistic regression model does not need to learn the complex non-linear physics of the circuit; it simply treats these new columns as additional, independent linear inputs. The quantum circuit performs the projection into a Hilbert space—a vector space with many dimensions—where patterns that were previously entangled and inseparable by a flat plane become linearly separable. The logistic regression model then defines a hyperplane that partitions this high-dimensional space, effectively turning a difficult non-linear classification task into a series of weighted linear sums.

Limits and Scalability

This approach functions as a specialized pre-processing step. The observed gains are confined to models with linear expressivity. When applied to algorithms like Random Forest, XGBoost, Support Vector Machines (SVM), and k-Nearest Neighbors (k-NN), these extra features provided no additional benefit. These models do not exhibit performance gains because they are already equipped to handle the non-linearities in the dataset internally. The quantum transformation provides no information that these sophisticated models are not already extracting on their own.

Computational overhead for generating these features remains manageable because the circuit depth is kept constant, avoiding the exponential time costs associated with deep circuits. However, as the number of qubits increases, the Hilbert space grows as 2 to the power of N. Classical computers simulate these circuits by calculating the state vector; eventually, the memory requirement for this state vector exceeds available hardware, which creates a simulation ceiling for the performance of these feature engineering techniques. What remains unknown is whether this specific approach scales to significantly larger datasets or more complex tabular inputs beyond the credit domain, or if the computational cost of simulation will always outpace the gain in predictive accuracy for larger qubit counts.