A personalized movie recommendation method considering user consensus and interest changes

Personalized movie recommendations depend on an accurate characterization of user interests. User interests not only evolve dynamically over time but also differ across users in their sentiment toward the semantic content of movies. Most existing studies rely on users’ historical ratings or interaction behaviors, making it challenging to jointly capture from reviews both the temporal evolution of their interests and their sentiment divergence toward semantic content. This results in coarse interest representations, which limit recommendation accuracy. To address these issues, a personalized movie recommendation method considering user consensus and interest changes is proposed. First, the BERTopic model is employed to mine fine-grained interest tags from movie reviews, and these tags characterize users’ content-level focus. Second, an improved forgetting function is introduced to compute users’ interest levels in these tags, thereby modeling dynamic interests. Additionally, considering that users may hold divergent sentiments toward the same tag, the ERNIE model is adopted to identify sentiment preferences toward movie content. Third, users’ interest levels, sentiment preferences, ratings, and genre preferences are integrated into a consensus measurement framework to more comprehensively capture the consistency of user preferences across multiple dimensions and generate personalized recommendations accordingly. Experimental results show that the proposed method outperforms baseline models in terms of overall performance on the Douban and Amazon datasets. Compared with the best-performing baseline, the method achieves reductions of 1.0% and 7.7% in MAE, and 3.3% and 3.1% in RMSE, respectively. These results confirm that integrating dynamic interest modeling with user consensus can effectively enhance recommendation accuracy.