MINING USER-GENERATED CONTENTS ON THE WEB AND SOCIAL NETWORKS

Abstract

In solving diverse data management problems, underlying social network between users and semantics hidden deep in User-generated Contents (UGC) can be useful from many perspectives. Finding and applying such hidden semantics of UGC and social correlations illustrates a new way in solving various problems. In this thesis, we study several challenging data management problems to investigate how to apply the framework of UGC mining and social network analysis to substantially improve existing solutions. In particular, we focus on the following four problems:

First, we propose a novel query expansion technique in Information Retrieval that exploits the “location-based” correlation between users and search engine user logs. We explore the vocabulary of users from different geographic locations and investigate the semantic relations among the documents they search for. Based on that, a hierarchical location and topic based query expansion model is proposed to improve the accuracy of web search. Our proposed model predicts the query location sensitivity with more than 80% precision. Using the model, the final search result is significantly better than several existing query expansion methods.

Second, we explore the aggregate social activity and evaluate the significance of various activity features in determining the social activity evolution. In particular, we look in to various formats of social activities and measure how member activity impacts the evolution of the active population. Several activity features are extracted and their impact on the community evolution is evaluated with a feature selection model. Based on the model, the most significant features are identified.

Third, we study UGC on Twitter, a large online platform of social media, to identify tweet topics and sentiments towards some preset brands/products. To help understand brand perception and customer opinions, we utilize the correlation of tweet sentiments and topics, and propose a multi-task multi-label (MTML) classification model that performs classification of both sentiments and topics simultaneously. It incorporates results of each task from prior steps to promote and reinforce the other iteratively. Meanwhile, by using multiple labels, the class ambiguity can be addressed. Compared with baselines, MTML produces a much higher accuracy of both sentiment and topic classification.

Furthermore, based on tweet sentiment analysis, social network among Twitter users is also taken into consideration to investigate the impact of events on tweet sentiment change. By mining tweets about 2012 USA presidential campaign, we analyze the sentiments towards the presidential candidates. Meanwhile, we incorporate social correlation between Twitter users and present a method to predict the impact of events based on social activities. Analysis on tweets collected over 8 months shows that our method can predict the sentiment change with high accuracy. Mining UGC and social network is not only efficient but also effective in predicting the impact of events.

Chapter  1

Introduction

The studies on mining user generated contents(UGC) and interpretation have been rapidly evolving in recent years. With the emergence of digital storage and a variety of online services, sources arise to enable in-depth research in relevant applications in various data management problems. In this dissertation, the semantic computing framework is adopted to learn from UGC and improve techniques in information retrieval, social activity mining, and sentiment analysis as well as prediction in social media.

UGC has a variety of formats, which are supported by different social media websites, such as Facebook, Tumblr, Twitter, and Youtube. Figure 1.1 shows an example of UGC from Twitter. On these websites, users can write and post UGC about any subject. At the same time, a user can also follow other users, which forms online social network. Therefore any update of the followed users will be delivered to the follower. On one hand, social network determines the diffusion of UGC, thus analyzing social network structure can help predict the generation of UGC. On the

Figure 1.1. An example of user generated contents: twitter stream

other hand, UGC also have impacts on individuals, which furthermore promotes the change of social network topology. Studying UGC and social connections can provide insights to this interactive relationship. Mining and understanding the correlations will help with solving a lot of problems.

As a study of interpretation analysis, UGC mining focuses on analyzing words, signs and symbols, as well as understanding their meaning. Based on that, semantic computing on UGC explores hidden semantics in UGC content, such as topics and sentiments. Semantic relations embodied in UGC reveal deep patterns and domain rules in a variety of fields. The objective of semantic computing on UGC is to understand the meaning of various sorts of computational content and furthermore find out mapping rules.

3

Generally, the topics addressed in mining UGC and semantic computing can be grouped into three categories: intention understanding, content interpretation, and semantic mapping. The first category involves understanding the intentions of human expression and convert them into a machine-processable language. In the second category, the focus is on understanding and converting various sorts of user generated content, including but not limited to text, video, audio, and image. Semantic mapping is based on the results of the first two categories. By understanding semantic objectives and extracting their relations, algorithms are developed to create mapping between semantic content for different purposes.

As the relations generated with semantic mapping are integrated, the meaning of semantic objectives can be understood within a common framework. Thus, relevant patterns for different purposes can be summarized from the computational content through the embedded semantic metadata. With the patterns extracted, many methods are proposed to apply UGC mining to different applications, including document parsing, semantic relation extraction, semantic interpretation, and entity disambiguation.

Existing researches have developed many techniques of semantics analysis from different perspectives. To study the semantic relations, Bollegala et al. proposed a relational similarity measure to compute the similarity between semantic relations by using a web search engine [1] . Furthermore, they presented a clustering algorithm to train the logistic regression model to identify the relation patterns expressed by each cluster [2] . By using a wide-coverage parser and semantic analyzer trained from newspaper text, a method is proposed to generate large scale semantic knowledge networks efficiently [3] . Another technique is proposed for scalable semantic retrieval by adopting summarization and refinement [4] . Based on visual and semantic consistency, a social image retagging scheme is developed to assign images with better content description [5] . These studies analyze the semantic metadata in text and summarize the mapping rules for further analysis.

In data management, several studies are conducted to explore applications of semantic mining on the web and social networks. In [6] , a technique is introduced to perform large scale semantic integration and reduce the redundancy while indexing on semantic web. Also, a decentralized infrastructure is proposed to efficiently tackle the graph-based entity disambiguation problem [7] . As an enhanced application, a semantic web search engine Falconer is developed to support friends auto-discovery, semantic annotation, as well as topic trend analysis [8] . Furthermore, by projecting each post into a topic space, a sparse coding-based model is constructed in [9] to simultaneously model semantics and structure of threaded discussions. In a variety of applications, relevant problems are addressed by mining semantics from UGC and utilizing the patterns extracted.

In this dissertation, we make use of UGC mining to help solve data management problems in a few domains, including query expansion in web search, social activity analysis, multi-label sentiment and topic co-classification, and prediction on impact of events. The UGC involved includes not only search engine user logs, but also online posts and microblogs in social networks.

MINING USER-GENERATED CONTENTS ON THE WEB AND SOCIAL NETWORKS

Sharing is caring!