From Sparsity and Efficiency to Persuasiveness: Addressing Key Challenge of RecSys

As the amount of online information increases, recommender systems have become more effective at reducing information overload. The usage of the recommender system has overwhelmed everyone's normal life, from the widespread application adoption in e-commerce applications to our daily life decisions, such as choosing a restaurant for dinner (food recommendation), recommending a movie for a family reunion (movie recommendation), or recommending music for your personal playlist (music recommendation). In most cases, recommender systems provide suggestions to users to help them make good decisions. In recent years, deep learning has garnered considerable interest in many research fields such as computer vision and natural language processing, owing not only to stellar performance but also to the attractive property of learning feature representations from scratch. The influence of deep learning is also pervasive, recently demonstrating its effectiveness when applied to recommender systems research. However, towards to the way to modern recommender systems, there are still a number of unresolved issues 1. Data sparsity: Due to the distinctiveness of the datasets in which most users interact with a limited number of items. As a result, all datasets used for recommendation are sparse. 2. Model efficiency: Recommender systems always have to be accurate and efficient in order to function effectively. While incorporating deep learning techniques into recommender systems can greatly increase the accuracy of the recommendation, the training workloads increase. To improve the user experience, the system must always respond quickly and accurately. 3. Explainability: The majority of recommendation models pursue accuracy in the recommendation, which directly impacts user satisfaction, but the user's trust in the system and its transparency are also significant factors. This thesis explores the opportunities that exist in three phases of recommender systems: data sparsity in the data preparation stage, efficiency in the training stage, and persuasiveness in the recommendation stage. Hence, it explores various deep learning methods, including transfer learning, attention, graph neural networks, and multi-task learning. As part of the new version of modern recommender systems, efforts are made to improve user trust and satisfaction.