Modeling Dynamic Aspects of Context-Aware Recommender Systems
Abstract:
In contrast to conventional Recommender Systems (RSs), which only utilize information about users and items, Context-Aware Recommender Systems (CARSs) incorporate contextual information and make the recommendation problem multi-dimensional. Recent evidence shows that the use of different contextual information (e.g., user's mood, current activity, companion, current location, and time) enhances the accuracy of recommendations. The advancement of smart devices such as smartphones plays an important role in capturing contextual information via their physical sensors and applications. The use of contexts poses challenges in non-stationary recommender environments such as inferring ever-changing latent contexts and balancing the trade-off between efficiency and information loss. However, the existing CARSs have found it difficult to address these challenges. In this study, we discuss how CARSs should be devised to handle a stream of user interactions that occur in a real-world recommender environment while improving efficiency and the quality of recommendations.
Firstly, we present two strategies to select discriminative instances (i.e., user interactions) from a stream of incoming instances. These strategies are based on the notion of Self-Paced Learning and the rating profiles of users, items, and contextual feature values. We combine these strategies with Factorization Machines to provide online recommendations. We aim to make the online updating process efficient via selecting "best" instances while reducing the impact of information loss.
Secondly, we present how to efficiently infer ever-changing latent contexts using a dynamic network embedding approach and to integrate these contexts with an incremental RS. To make the network embedding process efficient, our neighborhood sampling is designed to focus on the most discriminative (influential) nodes. Moreover, we design a biased random walk to explore the nodes that are changed over time, with higher transition probabilities. We show the feasibility of our approach by inferring the dynamics of spatiotemporal contexts in the Point-of-Interest (POI) recommender task.
Finally, we present how to use the notion of user curiosity as contexts. We argue that it is important to model the changes in user curiosity, and a CARS should be able to fine-tune recommendations according to subtle changes in user curiosity. We present a data analysis of how users reveal their curiosity in different application domains. We show how to use the classical Multi-armed bandits with Matrix Factorization to enhance prediction accuracy and temporal diversity in the POI recommender task.