Adaptive time series prediction and recommendation

The ubiquity of user-item interactions makes it essential and challenging to utilize the rich variety of hidden structural and temporal information for effective and efficient recommendation. In this work, our goal is to address the limitations of existing research: (i) inadequacy of popularity trend prediction and temporal recommendation (ii) failure to clarify the influence and mechanism of structural characteristics and the temporal evolution on recommendation. To this end, we first construct time sequences of grown popularity of items. Then we propose a family of time-series predictive models to predict the growing trend of popularity. Furthermore, we exploit the Broyden-Fletcher-Goldfarb-Shanno quasi-Newton optimization algorithm (BFGS) to adjust the predictive parameters adaptively. Moreover, to investigate the influence and interaction mechanism of structural and temporal information on recommendation, we propose a novel Hybrid Network Adaptive Time Series recommendation framework (HNATS), which improves synchronously the recommendation performance. Finally, we conduct comprehensive experiments on four real-world datasets of different sizes and time spans. The experimental results demonstrate that our proposed predictive models can capture the hidden temporal patterns and the HNATS method surpasses those compared state-of-the-art temporal ones, including the popularity-based, the time decay-based, and the Markov-based baselines.