Thesis Title: "Point-of-Interest Recommendation Systems in Location-Based Social Networks"
Recommendation systems are becoming more and more popular with the development of web and mobile technologies. The next Point-of-Interest (POI) Recommendation problem under the Recommendation Systems is also one of the Recommendation Systems research areas. The main purpose of the Next POI Recommendation Systems is to analyze the past data of the users and make a new POI Recommendation to them. User data is derived from Location-based Social Networks (LBSNs) applications such as Foursquare, Gowalla, Yelp, etc. Next POI Recommendation problems also have many challenging factors because there are many parameters that affect the Next POI visit decision. Proposed Point-of-Interest Recommendation models in the literature have been reviewed in detail separately under traditional machine learning and deep learning subtitles. Using traditional machine learning algorithms for a Point-of-Interest Recommendation model from end to end cannot give users a new POI Recommendation with high accuracy compared to today's results. However, traditional machine learning approaches are often used in Deep Learning to solve some part of the problem. Therefore, knowing about traditional machine learning approaches is the key to developing a state-of-the-art approach. Deep Learning models used in the literature have also been reviewed in detail and compared with each other.
One of the biggest challenges in Next POI Recommendation problems is that there are many contexts that influence users' decisions such as static contexts, transition contexts, dynamic contexts, etc. We have developed a new Next POI Recommendation model using transition (spatio-temporal intervals, geographical probability), and dynamic contexts (time, weather, friendship relationship) with the Attention-Based Deep Learning Model with Transition and Dynamic Contexts for Next POI Recommendation (ATDL-POI) model we proposed. Rich context input allows us to make customized recommendations for each user. LSTM which is a Deep Learning model was utilized to understand the rich context and the complex behaviors relationships of users. LSTM has been customized by using power-law based geographical coefficient, therefore, ATDL-POI has geographical distance probability awareness. When analyzing the past check-in data of the users, it was noticed that some check-ins had less effect on the user's future preferences. The short-term and long-term preferences of the users may be different. The attention mechanism was employed in ATDL-POI to solve this problem. ATDL-POI outperformed traditional machine learning algorithms, base LSTM, GRU and many models in the literature.
The thesis study has been completed with both the detailed survey study and the proposed new model. Other context parameters or other deep learning models can be added/used to the proposed ATDL-POI model and is a flexible approach open to improvements.
Monday, March 29 at 11:15am to 12:05pmVirtual Event