Overseas short video recommendations: A multimodal graph convolutional network approach incorporating cultural preferences

Abstract

In an age of cultural globalization, short video platforms are springing up around the globe, making it challenging to cater to a diverse mix of users with varied preferences and cultural backgrounds. In our research, we propose a novel suggestion model of short video material for international video apps through user preference modelling via hybrid multi-modal GCN (graph convolutional network). Unlike traditional methods that rely on the overall metadata of the short movies only, our approach jointly considers visual, linguistic and audio features of short movies, as well as user interactions, to propose personalized recommendations. Due to the effectiveness of the proposed method on TikTok and MovieLens dataset with a recall of 0.590 and video label classification accuracy more than 94.9%, The approach demonstrates effective use of resources with a maximum CPU utilization of only 44% whilst maintaining high user satisfaction across different age groups. Overall, the results have an implication that the proposed approach can lead to better user interaction and satisfaction in a culturally diverse environment.