Researchers at Beijing University of Posts and Telecommunications have developed a new approach to quantum machine learning that effectively combines information from multiple data sources. A team led by Fei Gao proposed quantum multi-view kernel learning using local information, called L-QMVKL, to address performance limitations when processing complex data. Existing quantum kernel methods often struggle with local structural patterns because they rely on a single perspective and global data structures. By fusing cross-view information through quantum multi-kernels and exploiting local data characteristics, L-QMVKL demonstrates a significant increase in accuracy on the Mfeat dataset, achieving competitive results when compared to traditional machine learning models. “Our research has the potential to advance the theoretical and practical understanding of quantum kernel methods,” said Fei Gao, a researcher at the National Key Laboratory of Networking and Switching Technology.
Quantum multiple kernel learning for multi-view data
Quantum computing offers the potential to solve previously unsolvable problems by exploiting an exponentially larger computational space. A new study from Beijing University of Posts and Telecommunications details how to apply this power to complex real-world datasets. This problem occurs due to overreliance on global data structures and poor representation of individual data views, which hinders performance on complex datasets. L-QMVKL is built on multi-kernel learning, an already established technique for processing data from multiple sources by building “quantum multi-kernels” designed to effectively integrate information across these views. The researchers further enhanced the system by incorporating local information, aiming to capture structural details inherent in the data itself. This is achieved through a sequential training strategy of quantum circuit parameters and connection weights using hybrid global/local kernel alignment. The research team says they evaluated the effectiveness of L-QMVKL through comprehensive numerical simulations on the Mfeat dataset, demonstrating significant improvements in accuracy.
While quantum kernel methods promise solutions to classically difficult problems by mapping data into vast Hilbert spaces, they have historically struggled with complex datasets that exhibit local patterns. This limitation is due to over-reliance on single-view feature representations and global data structures. The key innovation lies in the construction of quantum multi-kernels that combine view-specific quantum kernels to enhance cross-view data fusion and capture subtle relationships.
Quantum kernels offer the possibility of solving problems that cannot be solved with classical models by mapping data into an exponentially large Hilbert space.
