| dc.description.abstracteng | User-generated big data mining is vital important for large online platforms in terms of security, profits improvement, products recommendation and system management. Personal attributes recognition, user behavior prediction, user identification, and community detection are the most critical and interesting issues that remain as challenges in many real applications in terms of accuracy, efficiency and data security. For an online platform with tens of thousands of users, it is always vulnerable to malicious users who pose a threat to other innocent users and consume unnecessary resources, where accurate user identification is urgently required to prevent corresponding malicious attempts. Meanwhile, accurate prediction of user behavior will help large platforms provide satisfactory recommendations to users and efficiently allocate different amounts of resources to different users. In addition to individual identification, community exploration of large social networks that formed by online databases could also help managers gain knowledge of how a community evolves. And such large scale and diverse social networks can be used to validate network theories, which are previously developed from synthetic networks or small real networks. In this thesis, we study several specific cases to address some key challenges that remain in different types of large online platforms, such as user behavior prediction for cold-start users, privacy protection for user-generated data, and large scale and diverse social community analysis.
In the first case, as an emerging business, online education has attracted tens of thousands users as it can provide diverse courses that can exactly satisfy whatever demands of the students. Due to the limitation of public school systems, many students pursue private supplementary tutoring for improving their academic performance. Similar to online shopping platform, online education system is also a user-product based service, where users usually have to select and purchase the courses that meet their demands. It is important to construct a course recommendation and user behavior prediction system based on user attributes or user-generated data. Item recommendation in current online shopping systems is usually based on the interactions between users and products, since most of the personal attributes are unnecessary for online shopping services, and users often provide false information during registration. Therefore, it is not possible to recommend items based on personal attributes by exploiting the similarity of attributes among users, such as education level, age, school, gender, etc. Different from most online shopping platforms, online education platforms have access to a large number of credible personal attributes since accurate personal information is important in education service, and user behaviors could be predicted with just user attribute. Moreover, previous works on learning individual attributes are based primarily on panel survey data, which ensures its credibility but lacks efficiency. Therefore, most works simply include hundreds or thousands of users in the study. With more than 200,000 anonymous K-12 students' 3-year learning data from one of the world's largest online extra-curricular education platforms, we uncover students' online learning behaviors and infer the impact of students' home location, family socioeconomic situation and attended school's reputation/rank on the students' private tutoring course participation and learning outcomes. Further analysis suggests that such impact may be largely attributed to the inequality of access to educational resources in different cities and the inequality in family socioeconomic status. Finally, we study the predictability of students' performance and behaviors using machine learning algorithms with different groups of features, showing students' online learning performance can be predicted based on personal attributes and user-generated data with MAE$<10\%$.
As mentioned above, user attributes are usually fake information in most online platforms, and online platforms are usually vulnerable of malicious users. It is very important to identify the users or verify their attributes. Many researches have used user-generated mobile phone data (which includes sensitive information) to identify diverse user attributes, such as social economic status, ages, education level, professions, etc. Most of these approaches leverage original sensitive user data to build feature-rich models that take private information as input, such as exact locations, App usages and call detailed records. However, accessing users' mobile phone raw data may violate the more and more strict private data protection policies and regulations (e.g. GDPR). We observe that appropriate statistical methods can offer an effective means to eliminate private information and preserve personal characteristics, thus enabling the identification of the user attributes without privacy concern. Typically, identifying an unfamiliar caller's profession is important to protect citizens' personal safety and property. Due to limited data protection of various popular online services in some countries such as taxi hailing or takeouts ordering, many users nowadays encounter an increasing number of phone calls from strangers. The situation may be aggravated when criminals pretend to be such service delivery staff, bringing threats to the user individuals as well as the society. Additionally, more and more people suffer from excessive digital marketing and fraud phone calls because of personal information leakage. Therefore, a real time identification of unfamiliar caller is urgently needed. We explore the feasibility of user identification with privacy-preserved user-generated mobile, and we develop CPFinder, a system which implements automatic user identification callers on end devices. The system could mainly identify four categories of users: taxi drivers, delivery and takeouts staffs, telemarketers and fraudsters, and normal users (other professions). Our evaluation over an anonymized dataset of 1,282 users with a period of 3 months in Shanghai City shows that the CPFinder can achieve an accuracy of 75+\% for multi-class classification and 92.35+\% for binary classification.
In addition to the mining of personal attributes and behaviors, the community mining of a large group of people based on online big data also attracts lots of attention due to the accessibility of large scale social network in online platforms. As one of the very important branch of social network, scientific collaboration network has been studied for decades as online big publication databases are easy to access and many user attribute are available. Academic collaborations become regular and the connections among researchers become closer due to the prosperity of globalized academic communications. It has been found that many computer science conferences are closed communities in terms of the acceptance of newcomers' papers, especially are the well-regarded conferences~\cite{cabot2018cs}. However, an in-depth study on the difference in the closeness and structural features of different conferences and what caused these differences is still missing.
%Also, reviewing the strong and weak tie theories, there are multifaceted influences exerted by the combination of this two types of ties in different context. More analysis is needed to determine whether the network is closed or has other properties. We envision that social connections play an increasing role in the academic society and influence the paper selection process. The influences are not only restricted within visible links, but also extended to weak ties that connect two distanced node.
Previous studies of coauthor networks did not adequately consider the central role of some authors in the publication venues, such as \ac{PC} chairs of the conferences. Such people could influence the evolutionary patterns of coauthor networks due to their authorities and trust for members to select accepted papers and their core positions in the community. Thus, in addition to the ratio of newcomers' papers it would be interesting if the PC chairs' relevant metrics could be quantified to measure the closure of a conference from the perspective of old authors' papers. Additionally, the analysis of the differences among different conferences in terms of the evolution of coauthor networks and degree of closeness may disclose the formation of closed communities. Therefore, we will introduce several different outcomes due to the various structural characteristics of several typical conferences. In this paper, using the DBLP dataset of computer science publications and a PC chair dataset, we show the evidence of the existence of strong and weak ties in coauthor networks and the PC chairs' influences are also confirmed to be related with the tie strength and network structural properties. Several PC chair relevant metrics based on coauthor networks are introduced to measure the closure and efficiency of a conference. | de |