💻 computer science

Benchmarking Quantum Feature Encoding Strategies for Binary Classification with QSVM

This study demonstrates that incorporating statistical relationships into quantum feature encoding for Quantum Support Vector Machines can influence binary classification performance, but emphasizes that optimal strategies require balancing predictive accuracy with circuit complexity rather than simply increasing entanglement.

Original authors: Murat Kurt

Published 2026-09-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Murat Kurt

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the emerging field of quantum machine learning, researchers are trying to teach computers to recognize patterns using the strange rules of quantum physics. To do this, they must first translate ordinary data—like numbers describing a patient's health or a student's grades—into the language of quantum computers. This translation process is called encoding. Imagine trying to fit a complex, three-dimensional object into a flat, two-dimensional box; if you choose the wrong angle or the wrong way to squash the object, you lose the details that make it unique. In the quantum world, this translation happens by turning data points into specific configurations of quantum bits, or qubits. The way this translation is done is critical because it determines how well the computer can later find the differences between categories, such as distinguishing a healthy heart from a failing one. If the translation is too simple, the computer misses important clues. If it is too complicated, the computer gets confused by its own complexity or runs out of time before it can finish the calculation.

A researcher at Samsun University, Murat Kurt, recently set out to test exactly how different translation methods affect the ability of a quantum computer to sort data into two groups. The study focused on a specific type of algorithm known as a quantum support vector machine, which acts like a sophisticated sorter. The researcher tested five different real-world datasets, ranging from brainwave signals used to detect eye states to medical records predicting heart failure and credit risk assessments. For each dataset, the researcher tried several different ways of encoding the data. Some methods were simple, treating each piece of information independently. Others were more complex, attempting to link related pieces of information together within the quantum system, much like connecting dots on a map to reveal a hidden shape. The goal was to see if adding these connections, which represent statistical relationships between data points, actually helped the computer make better predictions, or if it simply made the process slower and more prone to errors.

The results of the study revealed a surprising truth: more complex is not always better. In some cases, the simplest method of encoding, which treated each data point on its own without trying to force connections between them, performed just as well as the most elaborate methods. In other instances, the simple method was actually superior. When the researcher tried to build a highly connected network where every piece of data was linked to every other piece, the computer often became too good at memorizing the training examples but failed to apply what it learned to new, unseen data. This is similar to a student who memorizes the answers to a practice test perfectly but fails the actual exam because they cannot recognize the questions when they are phrased differently. The study showed that these overly complex quantum circuits, while impressive in their design, often led to a sharp drop in performance when tested on fresh data.

The researcher also looked at a middle-ground approach where only the strongest statistical relationships between data points were used to create connections. This method did improve performance for some datasets, such as the heart failure prediction data, but it came with a significant cost. Building these connections required many more steps in the quantum calculation, which increased the time needed to run the simulation and the number of operations required. For other datasets, like the credit risk data, this extra effort provided no benefit at all; the simple method and the complex method produced identical results, meaning the extra work was wasted. The study found that the best approach depended entirely on the specific nature of the data being analyzed. There was no single "magic" encoding strategy that worked for every problem.

To make sense of these mixed results, the researcher developed a new way to score the different methods. Instead of just looking at how many correct answers the computer gave, this new score also weighed how much time the computer took to think and how much it struggled to generalize its learning. When this balanced score was applied, the most complex methods often fell to the bottom of the list. For example, on the student performance dataset, a simple encoding method achieved the highest score because it was fast, accurate, and reliable. In contrast, the most complex method, which tried to link every possible data point, scored the lowest because it was slow and made many mistakes on new data. Even on the dataset where the complex method achieved the highest raw accuracy, it still ranked lower than a slightly simpler method that was much faster and more stable.

The study concludes that the future of quantum machine learning lies not in building the most complicated circuits possible, but in choosing the right tool for the specific job. The research suggests that blindly adding more connections and entanglement to a quantum system does not guarantee better results. Instead, the most effective strategy is to understand the structure of the data first and then select an encoding method that matches that structure without unnecessary complexity. This approach ensures that the quantum computer remains efficient and capable of learning from new information, rather than just memorizing old examples. By carefully balancing the need for performance with the limits of current technology, researchers can build quantum models that are not only powerful but also practical and reliable.

Technical Summary: Benchmarking Quantum Feature Encoding Strategies for Binary Classification with QSVM

Problem Statement
In Quantum Machine Learning (QML), the method used to encode classical data into quantum states (quantum feature mapping) is a critical determinant of both classification performance and quantum circuit complexity. While various encoding strategies exist, such as angle encoding and standard ZZ feature maps, they often rely on predefined entanglement topologies (linear, circular, or fully connected) that do not necessarily reflect the statistical relationships inherent in the classical dataset. This study addresses the question of whether incorporating statistical dependency structures between features directly into the quantum feature map design can improve Quantum Support Vector Machine (QSVM) performance, and whether the resulting gains justify the increased circuit complexity and computational cost.

Methodology
The study employed a state-vector-based computational framework to simulate QSVMs on five distinct binary classification datasets: EEG Eye State Classification, Heart Failure Prediction, Credit Risk, Student Performance, and Alzheimer's Disease. The experimental workflow involved:

Data Preprocessing: Datasets were normalized, and categorical variables were encoded numerically. For the high-dimensional Alzheimer's dataset, Principal Component Analysis (PCA) was applied to retain the first 14 principal components (explaining ~49.6% of variance) to manage computational constraints.
Encoding Strategies:
- Baseline: Standard Angle-RY encoding, H+RY (Angle-RY with Hadamard gates), Z Feature Map, and ZZ Feature Maps with linear, circular, and full entanglement topologies.
- Proposed Approach (RY+CRY): A novel encoding strategy where individual features are mapped via RY( $\theta$ ) rotations, while statistical relationships between feature pairs are incorporated via controlled-RY (CRY) gates. Five statistical dependency measures were used to determine the CRY rotation angles: Pearson correlation, Spearman correlation, Kendall's Tau, Mutual Information (MI), and Distance Correlation.
Evaluation Metrics: Performance was assessed using Accuracy, Precision, Recall, F1-score, and ROC-AUC. Generalization was evaluated via the training-test accuracy gap and Log-Loss. Circuit complexity was measured by logical gate counts, circuit depth, and decomposed CX gate counts. Computational cost was tracked via state vector generation time, kernel computation time, and training time.
Composite Metric (QFES): A Quantum Feature Encoding Score (QFES) was introduced to jointly evaluate classification performance, generalization (Accuracy Gap), and computational cost (normalized depth, gate count, and time).

Key Contributions

Data-Driven Feature Mapping: The study proposes and systematically evaluates a method for embedding statistical dependencies between classical features directly into quantum circuits using controlled gates, moving beyond static, topology-based entanglement.
Comprehensive Benchmarking: It provides a comparative analysis of five different statistical dependency measures against standard quantum feature maps across five diverse datasets.
Holistic Evaluation Framework: The paper moves beyond simple accuracy metrics by introducing the QFES, which penalizes high circuit complexity and poor generalization, offering a more realistic assessment of QSVM viability in the NISQ era.
Empirical Evidence on Complexity: It challenges the assumption that denser entanglement or more complex feature maps invariably lead to better performance, demonstrating that optimal strategies are dataset-dependent.

Results
The findings varied significantly across datasets, indicating no single "universal" optimal encoding strategy:

EEG Eye State Dataset: The fully connected ZZ Feature Map achieved the highest test accuracy (0.850) but exhibited a large training-test gap (0.150) and high computational cost. The Z Feature Map offered a superior balance, achieving 0.834 accuracy with significantly lower circuit depth and cost. RY+CRY methods improved upon basic Angle-RY but incurred substantial circuit overhead.
Heart Failure Prediction Dataset: RY+CRY methods based on Pearson and Distance Correlation achieved the highest test accuracy (0.883). However, ZZ-based maps showed severe overfitting (e.g., ZZ circular achieved 0.994 training accuracy but only 0.757 test accuracy). The Z Feature Map provided the best trade-off between performance and complexity.
Credit Risk Dataset: All encoding strategies yielded identical test accuracy (0.908), Precision, Recall, and F1 scores. However, ROC-AUC values differed, with Distance Correlation-based RY+CRY performing best (0.663) and Z Feature Map performing worst (0.464), suggesting that while threshold-dependent metrics were insensitive, decision score distributions varied.
Student Performance Dataset: The Z Feature Map significantly outperformed all other methods (Test Accuracy: 0.944, F1: 0.900) with minimal circuit depth (2). Conversely, the fully connected ZZ map suffered from extreme overfitting (Test Accuracy: 0.712, F1: 0.040).
Alzheimer's Disease Dataset: Simple Angle-RY and Pearson-based RY+CRY achieved the highest test accuracy (0.764). Complex ZZ maps resulted in near-zero Recall and F1 scores despite perfect training accuracy. The study notes that PCA dimensionality reduction likely influenced these results, as only ~50% of variance was retained.

Significance and Claims
The paper concludes that the effectiveness of a quantum feature encoding strategy is not unidirectional; it depends on the interaction between the data structure, the feature representation, and the entanglement topology.

Complexity vs. Performance: Increasing circuit complexity (e.g., via full ZZ connectivity or encoding all pairwise dependencies) does not guarantee improved classification performance and often leads to overfitting and reduced generalization.
Statistical Dependencies: Incorporating statistical relationships via CRY gates can improve performance over basic angle encoding for certain datasets (e.g., EEG, Heart Failure), but the marginal gain must be weighed against the quadratic increase in circuit depth and gate counts.
Resource Efficiency: Simpler maps like Z Feature Map and Angle-RY often provide the most favorable balance of performance, generalization, and computational cost, as reflected in the QFES rankings.
Evaluation Paradigm: The author argues that evaluating QSVMs requires a composite view of predictive performance, generalization behavior, and resource cost. Relying solely on accuracy or ROC-AUC can obscure significant differences in model robustness and efficiency.

The study suggests that future quantum machine learning designs should prioritize "data-aware" and "resource-conscious" encoding strategies, potentially utilizing sparse RY+CRY structures that encode only the most informative dependencies rather than all pairwise relationships.

More like this