<p>The volume of data accumulating daily poses difficulties to consumers in the decision-making process, all the while troubling distributors in an attempt to comprehend customer behavior. This signals the importance of recommendation systems in order to resolve the data overloading issue automatically. Various approaches have been proposed with excellent results, however, few are applicable to the Vietnamese dataset. This paper provides an advanced approach to Vietnamese e-commerce data. Based on the graph-based deep learning method, combined with multimodal fusion, the model can leverage multimodal features. A combination of heterogeneous graphs and homogeneous graphs aids the model in capturing both the internal and external relationships of entities, thus making the model perform better in learning user/item representations and improving the accuracy of the provided suggestions.</p>