A Research Of Friendship Prediction System On Location-based Social Networks

LBSNs (location-based social network) record both friendship information like traditional social networks and users’check-in information in time and spatial dimensions, connecting virtual network and real life. They not only enrich people’s network life, but also provide new research directions in data mining and mobile internet field for us. As one of the new directions, link prediction system in LBSNs usually considers the information in network structure and spatial dimension so that the prediction features are simple relatively. In addition, LBSNs have sparse network structure and large prediction space, posing challenges for the link prediction system performance. To cure the above problems, this paper presents a novel LBSN link prediction system framework, combined with the traditional link prediction methods based on node similarity. The simulation results on Brightkite and Gowalla data show that the new link prediction system has nice performance. More specifically, the main works and contributions of this thesis are summarized as follows.1. Analyze the network structure properties and users’behavior features in Brightkite and Gowalla dataset. The analysis showes that the users’check-in times and check-in locations, the locations’visitors and views all have long-tailed distributions. There are some relatively isolated users and locations to be processed at the same time.2. Due to the zombie fans, delete the isolated nodes in the Co-location Networks of Brightkite and Gowalla to reserve more users instead of the Co-located Friends Networks method.3. Divide Brightkite and Gowalla networks into several communities with Louvain algorithm to reduce link prediction space and improve average precision at the same time.4. Mine users’similarity from network structure and various check-in behavior aspects, and put forward two kinds of link prediction features based on users’check-in time and frequency proved to be effective with the statistical analysis.5. Establish a novel LBSN link prediction system framework based on node similarity and analyze different prediction features’performance of unsupervised and supervised methods. The results show that these two kinds of features make the LBSN link prediction system perform better, improving prediction precision by 15.5%and F1value by7.4%in comparison with traditional link prediction features based on network structure and check-in locations.