Light-cone feature selection for quantum machine learning

Abstract: Feature selection plays an essential role in improving the predictive performance and interpretability of trained models in classical machine learning. On the other hand, the usability of conventional feature selection could be limited for quantum machine learning tasks; the technique might not provide a clear interpretation on embedding quantum circuits for classical data tasks and, more importantly, is not applicable to quantum data tasks. In this work, we propose a feature selection method with a specific focus on quantum machine learning. Our scheme treats the light-cones (i.e., subspace) of quantum models as features and then select relevant ones through training of the corresponding local quantum kernels. We numerically demonstrate its versatility for four different applications using toy tasks: (1) feature selection of classical inputs, (2) circuit architecture search for data embedding, (3) compression of quantum machine learning models and (4) subspace selection for quantum data. The proposed framework paves the way towards applications of quantum machine learning to practical tasks. Also, this technique could be used to practically test if the quantum machine learning tasks really need quantumness, while it is beyond the scope of this work.