User Profiles Inferring Methods On Social Networks

Online social networking site is a symbolic product in the era of web2.0. It has changed people’s life. Millions of users chat with friends, share interesting things and take part in activities via online social network sites. Users of social network sites usually share some personal information, such as gender, b irthday, location, educational background, hobbies on these sites. However, the reality is: the data there is often full of cracks, such as incomplete and inaccurate.In this paper, we study the methods of user profiles inferring on social networks. The main contents and contributions are as follows:First, the traditional relational data classification methods are proposed based on different scenarios. In this paper, we examine the relevant classical algorithm in detail, and compare them on same social network datasets. Then, we make a discussion on the pros and cons of those algorithms.Second, we analyze the nature of the use of homogeneity in some inferring algorithms combined with the notion of “strong ties” and "weak ties" in sociology. We propose the concept of "homogeneous edge" and "heterogeneous edge", stating that the noises come from "heterogeneous edge". Then we introduce the extended Euclidean distance, propose a self-learning metric for the homogeneity of edges. We apply a simple noise reduction method of removing those heterogeneous edges based on the self-learning metric, it was proved effectively in the experiments. This part of the dissertation shows that effective noise reduction method can be helpful for inferring user profiles.Third, we propose a new method for user profiles inferring——Hops Limited Relational Neighbor, which focused on the instability issue of classical algorithm LI(Local Iterative).One important flaw of LI is that when the number of iterations increases, the accuracy of the algorithm will firstly increase and then decrease. We follows the framework of LI, and redesign the voting process, we add three detail improvements including meticulous voting control, propagation distance limit and inflation operation. Then we compare the two algorithms in our experiments, and find that they are comparable on the accuracy measure, but HLRN is far more stabler.