Micro-video recommendation based on multi-dimensional sequence features in content e-commerce platform

Compared with traditional e-commerce platforms, content-based e-commerce has obvious advantages in commodity display, as it can present commodities more comprehensively, three-dimensionally and flexibly, providing better user service experience. However, the micro-video format of content-based e-commerce, with its complex information and high real-time requirements, limits recommendation accuracy. Most existing micro-video recommendation methods focus on static user interest modeling, while few construct dynamic interest models via short-term interaction sequences to adapt to rapid preference changes. Moreover, most dynamic modeling studies ignore multi-dimensional transfer relationships in interactions, leading to interest perception biases in recommendation tasks. To address these issues, this study proposes a multi-dimensional sequence and graph neural network based micro-video recommendation algorithm (MSAGNN). Specifically, the algorithm first constructs user preference sequences into three-dimensional graphs (item, category, author) chronologically. These graphs are input into the graph neural network to extract multi-dimensional micro-video feature representations, which are then fused via a dedicated strategy. Each sequence item is assigned positional information and processed by the attention mechanism to obtain the final session representation, which is dot-multiplied with candidate items to generate recommendations. Extensive experiments on two KuaiRec datasets with different densities verify the effectiveness and rationality of MSAGNN.