Improving CCA-based methods for SSVEP classification using a new graph reference signal

Brain-computer interface (BCI) systems enable individuals to control external devices through brain activity. Among the various paradigms in the BCI domain, steady-state visual evoked potentials (SSVEPs) are particularly dominant. One of the most effective methods for frequency detection in SSVEP-based BCIs is canonical correlation analysis (CCA). In standard CCA, a sine-cosine signal is used as a reference signal, which may not be optimal for SSVEP recognition. In this paper, we propose a novel graph reference signal that preserves a sinusoidal form, yet exhibits temporal smoothness on a graph learned from the reference signal. We employ an alternating optimization approach to obtain the graph reference signal and its Laplacian matrix representation. The proposed method was validated using 10 subjects from the benchmark SSVEP dataset with six target frequencies. Our proposed method achieved an average accuracy of $81.6 \%$ within a 1.5 s time window, representing a $3.3 \%$ improvement over the classical CCA method.