SP-3: Feature engineering
Summary: This pattern applies when feature engineering is delegated to a quantum computer, which serves as the first step in the inference pipeline by operating directly on the input data. The classical inference engine then carries out subsequent steps in the pipeline.

Figure 1. Graphical representation of the quantum feature engineering pattern.
Problem: Quantum computing offers benefits in the feature extraction procedure. However, limitations of NISQ-era quantum computers, such as the number of qubits and the depths of quantum circuits, prevent the efficient use of quantum components for further inference tasks.
Solution: This pattern leverages the power of quantum computing to initially extract features from raw data, taking advantage of its computational strengths and large dimensionality of the underlying Hilbert space. These extracted features are subsequently transferred to a classical system, which performs further analysis and inference using conventional machine learning algorithms. This pattern is often employed in support vector machines, where the quantum component is responsible for evaluating the kernel function. (Schuld and Killoran 2019) distinguishes between two types of quantum kernel models: the implicit approach, where the quantum device evaluates only the kernel function, and the explicit approach, where both kernel evaluation and classification are handled by the quantum computer. This pattern describes the former case.
Benefits:
- Efficiency. The quantum-feature engineering can help extract high-quality features, enhancing the performance of machine learning tasks. Quantum feature maps can, in principle, be computed with exponential speed-up for certain well-suited problems (Liu et al. 2021). For more practical applications, reports indicate that quantum support vector machines may operate with sub-quadratic run-time complexity (Park et al. 2023).
- Robustness. Several publications highlight the robustness of quantum feature engineering against noise in both data and hardware (Liu et al. 2021, Peters et al. 2021, Schetakis et al. 2022, Karimi et al. 2023).
Drawbacks:
- Restricted scalability for NISQ computers. Currently, quantum feature engineering can only be directly applied to data with relatively small dimensions due to the limited number of qubits available in NISQ-era quantum computers.
Known uses:
- (Vasques et al. 2023) proposed the utilisation of quantum feature engineering to classify brain neuron morphologies using multiclass classification with quantum kernel methods. They examined the impact of feature engineering on classification accuracy and found that quantum kernel methods performed similarly to classical methods.
- (Moradi et al. 2022) utilised a quantum distance classifier and a simplified quantum-kernel support vector machine on the 15-qubit IBM Melbourne quantum computer to address a classification problem using real clinical datasets.
- (Chalumuri et al. 2021) employed quantum computing techniques for feature extraction in image scene classification.