With the rapid development and digital transformation of online education, how to provide users with accurate personalized recommendations in the vast amount of educational resources has become an urgent problem to be solved. Traditional recommendation algorithms have shortcomings in capturing the complex and multi-level interactive relationships between students, teachers, and courses, which can easily lead to information overload and unsatisfactory recommendation results. This article proposes an online education recommendation algorithm based on relation aware heterogeneous graph neural network(RAHG-FKAN). This method first splits the global heterogeneous graph into a teacher centered graph and a course centered graph, capturing implicit relationships between students, teachers, and courses from different perspectives; Subsequently, the Fourier KAN module is introduced to map node features to the frequency domain and perform nonlinear feature transformation, constructing a dual tower structure to fuse fine-grained and coarse-grained information; Finally, feature fusion and personalized prediction are achieved using a shared multi-layer perceptron. The experiment verified on the MOOCCube dataset that this method significantly outperforms traditional models in terms of recommendation accuracy and efficiency, providing effective technical support for optimizing resource allocation and improving user experience on online education platforms.