With the rapid development of Internet technology and the comprehensive popularity of Internet applications, online activities have gradually become an indispensable part of people's daily life. The original recommendation learning algorithm is mainly based on user-microvideo interaction for learning, modeling the user-micro-video connection relationship, which is difficult to capture the more complex relationships between nodes. To address the above problems, we propose a personalized recommendation model based on graph neural network, which utilizes the feature that graph neural network can tap deep information of graph data more effectively, and transforms the input user rating information and item side information into graph structure, for effective feature extraction, based on the importance sampling strategy. The importance-based sampling strategy measures the importance of neighbor nodes to the central node by calculating the relationship tightness between the neighbor nodes and the central node, and selects the neighbor nodes for recommendation tasks based on the importance level, which can be more targeted to select the sampling neighbors with more influence on the target micro-video nodes. The pooling aggregation strategy, on the other hand, trains the aggregation weights by inputting the neighborhood node features into the fully connected layer before aggregating the neighborhood features, and then introduces the pooling layer for feature aggregation, and finally aggregates the obtained neighborhood aggregation features with the target node itself, which directly introduces a symmetric trainable function to fuse the neighborhood weight training into the model to better capture the different neighborhood nodes' differential features in a learnable manner to allow for a more accurate representation of the current node features.