Exploring user interest behind massive user behaviors is essential for online recommendations. Although recommendation models have been proposed recently with great success, existing studies ignore not only the timeliness of online users’ behaviors in terms of their interest, but also the sequential characteristics of their behaviors. To overcome this limitation, we construct a User Movie Interest Space (UMIS) model based on the sequential ratings of users. We define three indexes to elucidate the features of the interest of users for UMIS, which describe different patterns of behaviors of users related to their interests. Based on UMIS we propose a deep learning model named Dynamic Interest Flow (DIF) to provide dynamic movie recommendations. The DIF model achieves intelligently multi-dimensional observations on a user’s interest space and to predict simultaneously a variety of their future interests. Experimental results indicate that DIF outperforms traditional rating-based models and other state-of-the-art deep learning models. Results also demonstrate that modeling a dynamic recommendation as a sequential prediction is supposed to obtain outstanding advantages.

探索大量用户行为背后的用户兴趣对于在线推荐至关重要。尽管最近已经提出了推荐模型，但是推荐的模型非常成功，但是现有研究不仅忽略了在线用户行为的及时性，还忽略了其行为的顺序特征。为了克服此限制，我们基于用户的连续收视率构造了用户电影兴趣空间（UMIS）模型。我们定义了三个指标来阐明UMIS的用户兴趣特征，它们描述了与用户兴趣相关的用户行为的不同模式。基于UMIS，我们提出了一种称为动态兴趣流（DIF）的深度学习模型，以提供动态电影推荐。DIF模型可在用户的兴趣空间上实现智能的多维观察，并同时预测其未来的各种兴趣。实验结果表明，DIF优于传统的基于评分的模型和其他最新的深度学习模型。结果还表明，将动态推荐建模为顺序预测应该具有显着的优势。