Analysis of Evolutionary Patterns in Short-Video Content Preferences Based on Audience Behavior Data

User preferences on short-video platforms exhibit dynamic evolutionary characteristics that traditional static recommendation models struggle to capture in terms of drift and abrupt shifts. Based on 32 million user behavior records from a leading platform, this study constructs multi-dimensional feature vectors and proposes an LSTM-based preference drift-and-mutation detection model. Through multi-granularity temporal window dynamic modeling of evolutionary trajectories, the detection accuracy reaches 87.3%, representing a 23.6 percentage point improvement over baseline methods. The study further quantifies the driving mechanisms of content features, social events, and algorithmic interventions, and designs a dynamic recommendation mechanism accordingly. A/B testing results show that user dwell time increased by 18.7% and content diversity improved by 31.2%, providing a theoretical basis and technical pathway for precise recommendation and content ecosystem optimization on short-video platforms.