In this thesis, We study the application of two types of recommender systems.

In the rst part, we propose a novel tensor-based method to provide POI recommendations to users. Location-based social networks(LBSNs) have attracted millions of users to share their social friendship and daily activity locations. Plenty of the available check-in records make it possible to mine users preferences on locations and provided POI recommendations to users. However, It is a challenge to learn the preference of users since check-in data are sparse, long-tail, and inuenced by time and social relationships. Our proposed method can simultaneously deal with the above issues and experiments on a real-world dataset show that our model could provide more accurate location recommendations to users.

We illust...[ Read more ]

In this thesis, We study the application of two types of recommender systems.

In the rst part, we propose a novel tensor-based method to provide POI recommendations to users. Location-based social networks(LBSNs) have attracted millions of users to share their social friendship and daily activity locations. Plenty of the available check-in records make it possible to mine users preferences on locations and provided POI recommendations to users. However, It is a challenge to learn the preference of users since check-in data are sparse, long-tail, and inuenced by time and social relationships. Our proposed method can simultaneously deal with the above issues and experiments on a real-world dataset show that our model could provide more accurate location recommendations to users.

We illustrate a novel SAVER model for handling a dynamic treatment recommendation setting involving multiple diseases and medications in the second part. Most of the existing works based on reinforcement learning just construct the reward function and state simply since it is a hard task to nd a proper metric to evaluate the condition of a patient with multi-morbidity and to model the complicated daily condition of patients. In this thesis, a novel reward function is dened to provide a preliminary signal for which action is responsible for a good or bad outcome. A state with a more complicated structure is used to capture more features of a patients condition. Besides, a supervised constraint is fused into a reinforcement learning framework to avoid unacceptable risks.