Brain Decoding via Graph Kernels

An emergent trend in data analysis of functional brain recordings is based on multivariate pattern recognition. Unlike univariate approaches, it is designed as a prediction task by decoding the brain state. fMRI brain decoding is a challenging classification problem due to the noisy, redundant and spatio-temporal correlated data, where there are generally much more features than samples. The use of a classifier requires that raw data is mapped into n-dimensional real vectors where the structural information of the data is not taken into account. Alternative methods propose a different data representation based on a graph encoding. While graphs provide a more powerful representation, machine learning algorithms for this type of encoding become computationally intensive. The contribution of this paper is the introduction of a graph kernel with a lower computational complexity that allows taking advantage from both the representative power of graphs and the discrimination power of kernel-based classifiers such as Support Vector Machines. We provide experimental results for a discrimination task between faces and houses on a fMRI dataset. We also investigate on synthetic data, how the brain decoding task differs according to the different encodings: vectorial and graph-based. A remarkable feature of the graph approach is its capability to handle data from different subjects, without the need of any intersubject alignment. An intersubject decoding experiment is also performed for the faces versus houses problem.