Identification of Online Users' Social Status via Mining User-Generated Data
by Tao Zhao
Date of Examination:2019-09-05
Date of issue:2019-09-11
Advisor:Prof. Dr. Xiaoming Fu
Referee:Prof. Dr. Xiaoming Fu
Referee:Prof. Dr. Marcus Baum
Files in this item
Name:PhD_thesis_TaoZhao.pdf
Size:3.19Mb
Format:PDF
Description:PhD thesis
Abstract
English
With the burst of available online user-generated data, identifying online users’ social status via mining user-generated data can play a significant role in many commercial applications, research and policy-making in many domains. Social status refers to the position of a person in relation to others within a society, which is an abstract concept. The actual definition of social status is specific in terms of specific measure indicator. For example, opinion leadership measures individual social status in terms of influence and expertise in an online society, while socioeconomic status characterizes personal real-life social status based on social and economic factors. Compared with traditional survey method which is time-consuming, expensive and sometimes difficult, some efforts have been made to identify specific social status of users based on specific user-generated data using classic machine learning methods. However, in fact, regarding specific social status identification based on specific user-generated data, the specific case has several specific challenges. However, classic machine learning methods in existing works fail to address these challenges, which lead to low identification accuracy. Given the importance of improving identification accuracy, this thesis studies three specific cases on identification of online and offline social status. For each work, this thesis proposes novel effective identification method to address the specific challenges for improving accuracy. The first work aims at identifying users’ online social status in terms of topic-sensitive influence and knowledge authority in social community question answering sites, namely identifying topical opinion leaders who are both influential and expert. Social community question answering (SCQA) site, an innovative community question answering platform, not only offers traditional question answering (QA) services but also integrates an online social network where users can follow each other. Identifying topical opinion leaders in SCQA has become an important research area due to the significant role of topical opinion leaders. However, most previous related work either focus on using knowledge expertise to find experts for improving the quality of answers, or aim at measuring user influence to identify influential ones. In order to identify the true topical opinion leaders, we propose a topical opinion leader identification framework called QALeaderRank which takes account of both topic-sensitive influence and topical knowledge expertise. In the proposed framework, to measure the topic-sensitive influence of each user, we design a novel influence measure algorithm that exploits both the social and QA features of SCQA, taking into account social network structure, topical similarity and knowledge authority. In addition, we propose three topic-relevant metrics to infer the topical expertise of each user. The extensive experiments along with an online user study show that the proposed QALeaderRank achieves significant improvement compared with the state-of-the-art methods. Furthermore, we analyze the topic interest change behaviors of users over time and examine the predictability of user topic interest through experiments. The second work focuses on predicting individual socioeconomic status from mobile phone data. Socioeconomic Status (SES) is an important social and economic aspect widely concerned. Assessing individual SES can assist related organizations in making a variety of policy decisions. Traditional approach suffers from the extremely high cost in collecting large-scale SES-related survey data. With the ubiquity of smart phones, mobile phone data has become a novel data source for predicting individual SES with low cost. However, the task of predicting individual SES on mobile phone data also proposes some new challenges, including sparse individual records, scarce explicit relationships and limited labeled samples, unconcerned in prior work restricted to regional or household-oriented SES prediction. To address these issues, we propose a semi-supervised Hypergraph based Factor Graph Model (HyperFGM) for individual SES prediction. HyperFGM is able to efficiently capture the associations between SES and individual mobile phone records to handle the individual record sparsity. For the scarce explicit relationships, HyperFGM models implicit high-order relationships among users on the hypergraph structure. Besides, HyperFGM explores the limited labeled data and unlabeled data in a semi-supervised way. Experimental results show that HyperFGM greatly outperforms the baseline methods on individual SES prediction with using a set of anonymized real mobile phone data. The third work is to predict social media users’ socioeconomic status based on their social media content, which is useful for related organizations and companies in a range of applications, such as economic and social policy-making. Previous work leverage manually defined textual features and platform-based user level attributes from social media content and feed them into a machine learning based classifier for SES prediction. However, they ignore some important information of social media content, containing the order and the hierarchical structure of social media text as well as the relationships among user level attributes. To this end, we propose a novel coupled social media content representation model for individual SES prediction, which not only utilizes a hierarchical neural network to incorporate the order and the hierarchical structure of social media text but also employs a coupled attribute representation method to take into account intra-coupled and inter-coupled interaction relationships among user level attributes. The experimental results show that the proposed model significantly outperforms other stat-of-the-art models on a real dataset, which validate the efficiency and robustness of the proposed model.
Keywords: Socioeconomic Status; Social Community Question Answering; Mobile Phone Data; Social Media Content; Factor Graph; Hypergraph; Coupled Attribute Representation; Bidirectional Long Short-Term Memory; User-Generated Data; Social Status; Topical Opinion Leader