FriendBurst: Ranking people who get friends fast in a short time

Abstract

Number of friends (or followers) is an important factor in social network. Attracting friends (or followers) in a short time is a strong indicator of one person for becoming an influential user quickly. Existing studies mainly focus on analyzing the formation of relationship between users, however, the factors that contribute to users' friend (or follower) numbers increment are still unidentified and unquantified. Along this line, based on users' different friends (or followers) increasing speeds, firstly, we get a number of interesting observations on a microblog system (Weibo) and an academic network (Arnetminer) through analyzing their characteristics of structure and content from the diversity and density angles. Then we define attribute factors and correlation factors based on our observations. Finally we propose a partially labeled ranking factor graph model (PLR-FGM) which combines these two kinds of factors to infer a ranking list of the users' friends (or followers) increasing speed. Experimental results show that the proposed PLR-FGM model outperforms several alternative models in terms of normalized discounted cumulative gain (NDCG).

Introduction

With the success of social Web, the online social networks, such as Facebook, Twitter and DBLP, significantly enlarge our social circles. The friends (or followers) in social networks are important resources not only for transferring messages but also for being popular, which can be considered as an crucial indicator of social status. For example, in microblogs, the increment of followers of a user means his/her published contents have more audiences and his/her actions could affect more people. And if a person's follower number has a great “burst” suddenly, he/she would probably become a “new star”. On the other hand, in academic social networks, an author who get more co-authors in a short time means he/she is more active and can build reputation in his/her research area quickly. Understanding the characteristics of users who attract friends fast is an important issue for social influence analysis, which can provide suggestions for users' behaviors and benefit many applications such as “virtual market” and recommendation systems.

Generally speaking, quick increments of friends (or followers) means that the users get new relationships in a short time. In the literature, there exists some studies on relationship analysis, for example, link prediction and unfollow behavior analysis. The goal of link prediction [1], [2] is to predict whether one user will follow another in the future, while unfollow behavior analysis [3], [4] mainly focuses on analyzing the reason of unfollow behaviors. In sum, most of the existing literatures study the formation of friendship between users. But another perspective, the factors that impact on users' friends increasing speed are still unidentified and unquantified. Although Hutto et al. [5] did a longitudinal study on followers increment, where they built a regression model for follower count prediction, the correlation between magnitude of content-based and structural factors and the friends increasing speed is ignored; moreover, their dataset is small.

Different with these works, we want to propose a method to infer the ranking list of friends increasing speed for candidate users. There are several challenges for the friends increasing speed ranking. First, how to capture the rich structural and content-based information for friend increment analysis? Second, how to construct an algorithm to model both the users' attributes and the relationships between users? Third, how to validate the proposed model in real large social networks.

To address the above challenges, we firstly perform some statistical analysis on the correlations between users' friend (or followers) increasing speed and their structural and content-based properties. The analysis is conducted based on two networks, namely, a microblog system (Weibo) and an academic network (Arnetminer). The structural and content-based information are studied with in-depth exploration. For the structure based analysis, we utilize calculations such as diversity and density of circles and structures; for the content based analysis, we also define diversity and density based on topic distribution and hashtags. We then propose a partially labeled ranking factor graph model (PLR-FGM) to infer the ranking list of friends increasing speed. The model cannot only use the structural and contents-based properties of individuals as attribute factors, but also model the relationships between uses as the correlation factors. The ranking list can be obtained by sorting the marginal probabilities which are calculated by the model. Experimental results show that our PLR-FGM model significantly outperforms several alternative methods in terms of normalized discounted cumulative gain (NDCG) with augments ranging from 6% to 51% .

The contributions of this paper are summarized as follows:

• Based on large datasets from two real social networks – a microblog system (Weibo) and an academic network (Arnetminer), we derive observations and analyze the correlation between users' friends increasing speed and the users' structures from diversity and density angles. Besides, we also analyze the effects of their contents properties (such as circle diversity and density) on the friends increasing speed.

• We propose a ranking factor graph model which not only incorporates the structural and contents-based properties of individual users but also model the relationships between them. Then we use the model to detect the users who have a higher friends increasing speed in the social network.

• We conduct experiments on the two real social networks. Experimental results verified the effectiveness of our observation factors, and the proposed model can achieve a better performance than several alternative models.

The paper is organized as follows: In Section 2, we give a brief description of the datasets we used and perform some preprocessing on these two datasets; Section 3 presents our observations on users' friends burst states with their attributes such as structure and content. Section 4 explains the proposed ranking factor graph model. Section 5 illustrates experimental results and validates the effectiveness of our methodology. Finally, Section 6 discusses related work and Section 7 concludes.