Quantum Circuits Cut Data Imbalance Recall Gap by 10–25%

A new hybrid quantum-classical framework is demonstrating a measurable improvement in machine learning performance on imbalanced datasets, with researchers reporting a 5, 15% increase in F1-score and minority-class recall improving by 10-25%. The team, comprised of Tanapol Nuatho, Narisorn Sangnakara, Prapong Prechaprapranwong, and Rajchawit Sarochawikasit, developed a Quantum Circuit Born Machine (QCBM) to generate synthetic data that addresses limitations caused by data scarcity and class imbalance. Experiments utilizing the Iris dataset and the Telco Customer Churn dataset show that augmenting training data with QCBM-generated samples at 40-50% of the minority class improves F1-score by approximately 5-15%. Cross-domain evaluations revealed a performance gap of only 3-10%, indicating a high degree of distributional fidelity for the quantum-generated samples.

QCBM for Imbalanced Tabular Data Generation

A five to fifteen percent increase in F1-score and minority-class recall improving by ten to twenty-five percent is now achievable through a novel application of quantum computing, specifically by augmenting training data with synthetic samples generated using a Quantum Circuit Born Machine (QCBM). Researchers are addressing the longstanding challenge of class imbalance in machine learning, where algorithms struggle to accurately identify rare but critical instances, a common issue in fields like fraud detection and medical diagnosis. This performance boost demonstrates a measurable gain directly linked to harnessing quantum mechanical principles. The team’s approach centers on the QCBM, a parameterized quantum circuit designed to model complex probability distributions. Unlike classical generative methods, the QCBM leverages quantum superposition and entanglement to potentially capture nuanced relationships within data.

Before quantum processing, data undergoes classical pre-processing; normalization and Principal Component Analysis (PCA)-based dimensionality reduction are employed to enable efficient basis encoding for quantum circuits. This hybrid quantum-classical pipeline is crucial for translating real-world tabular data into a format suitable for quantum computation. The QCBM is then trained by minimizing the Kullback-Leibler (KL) divergence, a measure of difference between the real and generated data distributions, using a gradient-based parameter-shift optimization rule. The fidelity of the generated data is surprisingly high, which is a critical finding because the utility of synthetic data depends on its ability to accurately represent the underlying distribution.

PCA-Based Dimensionality Reduction for Quantum Encoding

Efficient data encoding is crucial for preparing classical data for processing on quantum hardware, and this task is essential for the promise of quantum machine learning. While the Quantum Circuit Born Machine (QCBM) offers a novel approach to synthetic data generation, its effectiveness hinges on this efficient encoding. This dimensionality reduction isn’t about arbitrarily simplifying the data; it’s about aligning the classical information with the constraints of quantum representation. By reducing the number of features, PCA minimises the number of qubits required to encode the data, a critical consideration given the limited availability and inherent fragility of qubits. This careful preparation allows for a more streamlined mapping of classical data into the quantum realm, maximizing the potential of the QCBM. This preprocessing step does not come at the cost of performance, and the boost in recall is particularly significant for datasets plagued by class imbalance, where identifying minority class instances is paramount.

The researchers also found that QCBM achieves competitive classification performance and produces lower Maximum Mean Discrepancy (MMD) on the Telco dataset, indicating a superior structural similarity compared to classical oversampling methods. This combination of improved recall, distributional fidelity, and structural similarity establishes the QCBM, when paired with careful classical preprocessing, as a promising tool for addressing data scarcity and imbalance in machine learning applications.

KL Divergence & MMD for Distribution Fidelity

Researchers are refining methods to assess the quality of synthetic tabular data generated by their Quantum Circuit Born Machine (QCBM), moving beyond simple classification accuracy to focus on distributional fidelity. Achieving high classification scores with augmented datasets is important, but the team emphasizes the need to verify how closely the synthetic data mirrors the statistical properties of the original, real-world data. This is particularly crucial when addressing the challenges of imbalanced datasets, where minority classes are underrepresented and require careful augmentation to avoid biased models. Alongside KL divergence, the team utilizes Maximum Mean Discrepancy (MMD) to evaluate structural similarity, providing a complementary perspective on the fidelity of the generated data. The researchers state they systematically compare QCBM against four SMOTE variants, offering the first comprehensive classical, quantum benchmark for tabular data augmentation.

Results indicate that the QCBM not only achieves competitive classification performance but also consistently produces lower MMD scores on the Telco Customer Churn dataset, suggesting a superior ability to capture the underlying data structure. The evaluation extends to cross-domain testing, assessing the QCBM’s ability to generalise. Through “Train on Synthetic, Test on Real” (TSTR) and “Train on Real, Test on Synthetic” (TRTS) protocols, the team observed a performance gap of only 3- 10%, demonstrating a surprisingly high level of distributional fidelity. This suggests that the synthetic data generated by the QCBM are not merely useful for improving classification but also closely resemble real-world data in their statistical characteristics. The team reports that augmenting training data with QCBM-generated synthetic samples, representing 40% to 50% of the minority class, improves minority-class recall by 10% to 25%, a significant gain for applications where accurate identification of rare events is critical.

F1-Score & Recall Gains with Synthetic Samples

Augmenting imbalanced datasets with synthetically generated samples using quantum computing techniques is yielding measurable improvements in machine learning performance, particularly in scenarios where identifying minority classes is critical. Researchers have demonstrated that a Quantum Circuit Born Machine (QCBM) can significantly boost recall by 10, 25% and F1-score by 5, 15% when applied to datasets struggling with uneven class representation. This gain is especially valuable in fields like fraud detection or medical diagnostics, where missing even a small number of positive cases can have substantial consequences. The QCBM, a parameterized quantum circuit, doesn’t simply replicate existing data points; it learns the underlying probability distribution and generates novel samples that closely mirror the real data. This fidelity is evidenced by cross-domain evaluations, where models trained on QCBM-generated data perform with a performance gap of only 3, 10% when tested on real-world datasets, and vice versa.

This suggests the synthetic data isn’t merely mimicking the training set but capturing the essential characteristics of the broader data distribution. The ability to generate high-fidelity synthetic data opens new avenues for improving model generalization and reducing bias in critical applications.

Cross-Domain Evaluation of QCBM Performance

A nuanced application of quantum computing is gaining traction: synthetic data generation. This cross-domain performance, where models are trained on QCBM-generated data and evaluated on real-world examples, challenges the assumption that quantum-created data is inherently limited to the characteristics of its training source. Researchers focused on assessing the QCBM’s ability to generalize beyond the datasets used for its initial training. Specifically, the team reports approximately a 5- 15% increase in F1-score and a noteworthy 10- 25% improvement in minority-class recall, a crucial metric when dealing with imbalanced datasets where identifying rare events is paramount. This suggests the QCBM isn’t simply memorizing the training data, but learning underlying patterns applicable to new, unseen instances. The team didn’t stop at evaluating performance within the same dataset.

Cross-domain evaluations reveal a performance gap of only 3, 10%, indicating strong distributional fidelity. This level of fidelity is particularly significant, as it suggests the QCBM is capable of capturing the essential characteristics of the data, rather than producing statistically plausible but structurally dissimilar samples. Comparative analysis further solidified the QCBM’s position. These findings establish QCBM as a viable complementary tool for data augmentation, particularly for low-dimensional structured tabular data with class imbalance.

Researchers meticulously evaluated the QCBM against four established classical oversampling techniques, SMOTE, Borderline-SMOTE, KMeans-SMOTE, and SVM-SMOTE, in a comprehensive benchmark designed to assess the viability of quantum approaches for tabular data augmentation. This improvement is particularly significant given the inherent difficulties in accurately predicting rare events or identifying minority groups.