A Deep Temporal Collaborative Filtering Recommendation Framework via Joint Learning from Long and Short-Term Effects
Qianqian Ji, Xiaoyu Shi, Mingsheng Shang · 2019
Recommendations based on deep learning technologies are attracting increasingly attentions from academia and industry recently, due to it has the powerful ability of data representation learning, and can capture the complex and non-linear user-item interaction patterns. Previous studies on deep learning based recommenders heavily emphasize on learning the short-term or long-term temporal effects in separate ways. Furthermore, most of the existing recommenders studied on temporal dynamics hidden in user-item interactions by using ratings or review texts solely, without utilizing these heterogeneous side information in a comprehensive manner. As a result, they own a limited ability to exploit all of available user-item interaction data. To address the above issues, we propose a temporal collaborative filtering recommendation framework with utilizing deep learning technologies, which can capture the drift of users' preferences and items' attributes over time from available heterogeneous feedback. For studying the rich sentiment information about particular item features hidden in reviews and capturing the short-term temporal dynamics in user preference, a parallel recurrent neural network (RNN) coupled in the last layer is presented. With consideration of the life cycle of each item, we further propose an item-based time evolution model as a supplement to enrich the long-term temporal effects when recommendation. In addition, a modified cosine similarity function is incorporated to produce the personalized candidate list for each user. Finally, extensive experimental are conducted on Amazon's three datasets. They indicate that the proposed method outperforms the other state-of-the-art recommenders in terms of precision and recall when doing the Top-n recommendation task.