Design of short video recommendation preference algorithm considering multimodal information features

With the continuous evolution of artificial intelligence technology, personalized recommendation systems in short video platforms urgently need more refined multimodal information understanding and semantic modeling. In order to improve the recommendation accuracy and robustness, the study proposes a short video recommendation algorithm that integrates multimodal features and graph neural networks. The algorithm models user-video multi-granularity associations by extracting video, audio, and text features, splicing them together, and then employing graph neural networks with hypergraph structure, combined with attention-gating mechanism. The application test results on real short video log datasets show that the proposed algorithm achieves an accuracy of 0.95 and 0.92 on the training and test sets, respectively, which significantly outperforms the two compared methods. Under different video types, the algorithm performs optimally in terms of click rate, average recommendation ranking and hit rate, in which the click rate of entertainment category reaches 0.098, the average recommendation ranking is 0.492, and the hit rate is 0.641. The experiments validate that the proposed algorithm has good practical application prospects in terms of recommendation accuracy and generalization ability in diverse content scenarios.