The rapid development of the Internet brings us an overwhelming amount of information. To find a relevant item, online users nowadays need to search from a vast number of items. The recommender system (RS) is one of the most effective ways to solve this information overload problem. In this presentation, we will discuss the recommender system from a physics perspective. This presentation starts with a brief overview of recommender systems. Next, we will introduce some recent research related to network-based information filtering and introduce an evalua-tion method that uses a triple-folder data division framework to avoid the over-fitting issue, based on which we will reexamine eighteen net-work-based recommendation methods using three representative da-tasets. Our results not only gave a more objective overview of the per-formance of the existing recommendation algorithms but also opened a new door to justify the effectiveness of recommendation algorithms with different numbers of parameters. Later, we will introduce a motif-based recommendation method, which can precisely find the potential lead-ers for users in online news propagation systems such as Twitter, which has performed well in both recommendation accuracy and diversity. At last, we will discuss some methods that can further enhance recom-mender systems. We will introduce a more personalized recommenda-tion algorithm in which each user is assigned their own parameters to match their preference. Besides network-based recommendations, we will introduce two other widely-used recommendation models which can be combined with network-based recommendation models. Additional-ly, we will introduce procedures to optimize recommendation results, aiming to improve user satisfaction in real applications. We conclude with an outlook and some final remarks in the end.