Existing micro-video recommendation methods are mainly based on graph neural network modeling the complex interaction between users and videos. Although there have been research attempts to introduce multimodal information to enhance the recommendation effect, most of the methods still use simple multimodal feature summation, which lacks the in-depth mining of the association between different modalities; at the same time, the existing micro-video recommendation methods tend to ignore the potential interest features in the user’s multiple interaction behaviors, which leads to the limitation of the recommendation results. In this regard, a micro-video recommendation method (Multi-Modal User Behavior Graph Neural Network, MMUB-GNN) that fuses multimodal data and user’s multi-behavior is proposed to explore a more effective multimodal fusion method to further explore the potential features of the items; then, design a more comprehensive user interest modeling mechanism by combining users’ multiple interaction behaviors; and then better express the connection between users and items to improve the accuracy and diversity of suggestions. Through comprehensive experiments on three public datasets, it is experimentally demonstrated that the proposed model performs considerably better than the leading multimodal micro-video recommendation techniques.