Social Media Analytics for User Behavior Modeling

A Task Heterogeneity Perspective

Arun Reddy Nelakurthi, Jingrui He

• Publisher: CRC Press
• Year: 2020
• ISBN:
  9781000025408
• Edition: 1
• Topic: Informatique
• Subtopic: Vision par ordinateur et reconnaissance de formes

Winner of the "Outstanding Academic Title" recognition by Choice for the 2020 OAT Awards.

The Choice OAT Award represents the highest caliber of scholarly titles that have been reviewed by Choice and conveys the extraordinary recognition of the academic community.

In recent years social media has gained significant popularity and has become an essential medium of communication. Such user-generated content provides an excellent scenario for applying the metaphor of mining any information. Transfer learning is a research problem in machine learning that focuses on leveraging the knowledge gained while solving one problem and applying it to a different, but related problem.

Features:

• Offers novel frameworks to study user behavior and for addressing and explaining task heterogeneity

• 
  
  

• Presents a detailed study of existing research

• 
  
  

• Provides convergence and complexity analysis of the frameworks

• 
  
  

• Includes algorithms to implement the proposed research work

• 
  
  

• Covers extensive empirical analysis

• 
  
  

Social Media Analytics for User Behavior Modeling: A Task Heterogeneity Perspective is a guide to user behavior modeling in heterogeneous settings and is of great use to the machine learning community.

1

Introduction

In recent years, social media has gained significant popularity and become an essential medium of communication. According to a survey, about 88% of the public in the United States use some form of social media, a 53% growth in the last decade. Also, the average number of accounts per user has increased from two in 2012 to seven in 2016 (Pew Research Center, d). The rise in social media usage both vertically in terms of the number of users by platform and horizontally in terms of the number of platforms per user has led to a data explosion. Popular social media platforms like Facebook, Instagram and Twitter manage tens of petabytes of information with daily data flows of hundreds of terabytes and a continually expanding userbase (Pew Research Center, c). Such huge volumes of user-generated content provide an excellent opportunity to mine data of interest. We can, thus, look for valuable nuggets of information by applying diverse search (information retrieval) and mining techniques (data mining, text mining, web mining, opinion mining).

User-generated content is diverse based on the need the social media platform caters to. Per one survey, amongst those who use social media roughly 67% stated staying in touch with current friends and family as a major reason, while 17% felt social media enabled them to connect with friends they have lost touch with (Pew Research Center, a). Other research indicated about 67% of the United States population use social media to stay updated on the latest news and seniors are driving that number up (Pew Research Center, b).

Social media usage has also seen a spike when it comes to personal healthcare. In the United States, nearly 90% of adults, in the age group 50-75, have used social media to seek and share health information (Tennant et al., 2015a). Research demonstrates that online social support programs like health care forums and social media websites (e.g. Facebook and Twitter) can help patients gain knowledge about their diseases and cope better with their daily management routine (Petrovski et al., 2015). Effectively mining information from these healthcare-related social media platforms can, thus, have a wide range of applications resulting in improved healthcare. For example, healthcare social networks can connect patients suffering from major chronic diseases such as diabetes mellitus, with physicians as well as other patients. Compared to generic social networks such as Twitter and Facebook, disease-specific social networks (e.g., TuDiabetes¹ and DiabetesSisters² ) have a greater concentration of patients with similar conditions and relevant resources. However, when it comes to such social networks, the patient is more likely to stick to a single social network, and would rarely look at other networks, thus limiting their access to online resources, especially patients with similar questions and concerns. Identifying patient groups with similar conditions can help connect patients across networks, thereby opening doors for knowledge sharing to help the community as a whole. Additionally, in a world of “fake news”, a lot of health information is misrepresented and therefore calls for authenticity. Motivated by the immense scope of leveraging social media information for healthcare and addressing underlying challenges with usage and reliability, in this book we explore answers to the following questions:

• Can social media serve as a platform for improved healthcare? Specifically, why would patients leverage social media and how would it impact their healthcare? Would it equip them to make better choices? And finally, does it help in communicating effectively with doctors and health care providers?
• How can we efficiently learn and build algorithms to mine knowledge from these healthcare-dedicated social networks? What are the challenges involved?
• Finally, how can we provide meaningful explanations to justify the behavior of algorithms and learning methods?

Unlike traditional mining settings where data is considered to be homogeneous for most mining tasks, user-generated social media data is intrinsically heterogeneous and thus poses a set of challenges. It can be both structured (ratings, tags, links) as well as unstructured (text, audio, video). Similar health-related social media websites that cater to users from different geographical locations can suffer from a distributional shift in user-generated data, either features or class labels. This shift could also be due to user bias or personal preferences. Transfer learning addresses the problem of distribution shift in data (Pan and Yang, 2010). In particular, task heterogeneity is reflected in inconsistent user behaviors across social media platforms, similar actors across social networks, etc. Therefore, in this book work, we aim to design efficient models and tools to help us leverage and learn from data heterogeneity in real-world scenarios that help in improving healthcare.

In scenarios where parts of data in one social network are hidden, missing or not available, leveraging it partially for mining is very challenging and has not been well studied. Motivated by the applications of task heterogeneity, in this book, we present our work on techniques for addressing task heterogeneity and the underlying challenges in social media analytics.

In lieu of the above questions and challenges for this research, three main research directions have been investigated:

D1.Social media in healthcare: To study the real-world impact of social media as a source to seek and offer support to patients with chronic health conditions.

D2.Learning from task heterogeneity: To propose various models and algorithms to learn and model user behaviors on social media platforms, to identify similar actors across social networks, to adapt and leverage information from existing black-box models to improve classification accuracy under domain adaptation settings.

D3.Model explainability: To provide interpretable explanations for heterogeneous predictive models in the presence of task heterogeneity.

The book is organized as follows. The related work, Chapter 2 discusses existing research and how the proposed methods differ from it. Chapter 3 discusses the impact of social media on patients with diabetes mellitus. Chapter 4 presents algorithms and models to learn from task heterogeneity in social media. Chapter 5 discusses methods to explain task heterogeneity. Finally, Chapter 6 concludes our research.

¹ http://www.tudiabetes.org/

² https://diabetessisters.org

2

Literature Survey

Since 2004, the growth of social media has been near exponential [We Are Social]. According to a survey, about 88% of the public in the United States use some form of social media, a 53% growth in the last decade [Pew Research Center, d]. This growth in social media usage led to an information explosion. Mining valuable nuggets of data from such information generated through social media has immense applications [Zafarani et al., 2014]. Machine learning techniques have been widely adopted to mine and analyze the large social media data to address many real-world problems. Mining from social media platforms has many applications, (1) Event detection - Social networks enable users to freely communicate with each other and share their recent news, ongoing activities or views about different topics. As a result, they can be seen as a potentially viable source of information to understand the current emerging topics/events [Nurwidyantoro and Winarko, 2013]; (2) Community detection - identifying communities on social networks, how they evolve, and evaluating identified communities, often without ground truth [Zafarani et al., 2014]; (3) Recommendation in social media - recommending friends or items on social media sites [Ricci et al., 2011]; (4) Sentiment and opinion mining - identifying collectively subjective information, e.g. positive and negative, from social media data [Liu, 2012]; (5) Network embedding - assigning nodes in a network to low-dimensional representations and effectively preserving the network structure [Cui et al., 2017].

As mentioned earlier in the introduction chapter, the intrinsic property of data heterogeneity in social media data poses a set of challenges. In this chapter, we present the existing research on handling data heterogeneity and study the impact of social media. In this chapter, we present existing work on impact of social media and its implications in Section 2.1, Section 2.2 presents existing research addressing data heterogeneity with a focus on task heterogeneity. Finally, we discuss the existing research on explaining models under task heterogeneity in Section 2.3.

2.1 Impact of Social Media

The growing popularity in the usage of social media platforms and applications has an impact on the individuals and society as a whole [Bishop, 2017]. These platforms have revolutionized the way we view ourselves, the way we see others and the way we interact with the world around us. Social media has many positive implications. Khurana [2015] studied the impact of social networking sites on the youth; it was shown that social media enables connecting with people all across the globe by not hampering their work hours and schedules and it also helps in education. Hudson and Thal [2013] studied the impact of social media on the consumer decision process and its implications for tourism marketing. Pew Research Center (b), showed that about 67% of the United States population uses social media to stay updated on the latest news. Also, the use of social media in politics including Twitter, Facebook, and YouTube has dramatically changed the way campaigns are run and how Americans interact with their elected officials [Bonilla and Rosa, 2015]. Social media usage has also seen a spike when it comes to personal healthcare. Tennant et al. [2015b] showed that nearly 90% of adults who use the internet and social media platforms like Facebook and Twitter used these platforms to find and share healthcare information. With a lot of growing interest and immense benefits from healthcare applications to society, we are motivated to work on addressing challenges in healthcare-related social media platforms.

Research demonstrates that online social support programs like health care forums and social media websites (e.g. Facebook and Twitter) can help patients gain knowledge about their diseases and cope better with their daily management routine [Petrovski et al., 2015]. Patel et al. [2015] studied the impact of social networks on perceived social support (e.g., of patients with chronic diseases). Researchers also studied how social media users gather and exchange health-related information and share personal experiences [Naslund et al., 2016; Shepherd et al., 2015]. Fung et al. [2016] researched the spread of misinformation about disease outbreaks to inform public health communication strategies.

2.2 Heterogeneous Learning and Social Media

Mining from healthcare-related social media platforms is challenging. The key to building applications from social media data is user-behavior modeling. Social media data is intrinsically heterogeneous - generated by users from different demographical locations, who speak different languages and have different cultural backgrounds. This makes user-behavior modeling under heterogeneity very challenging. Further, to mine across multiple social media platforms, the likelihood of the same user having multiple accounts is very low. Often they stick to one or two platforms that are popular based on the geographical location or demographics. To efficiently design applications that serve across multiple platforms, it is essential to identify similar users across the networks. Finally, it is very costly to collect labels for data from multiple platforms. A more practical approach would be to leverage knowledge from one platform to another. Motivated by this we identified three major problems: (1) modeling user-behavior; (2) identifying similar actors and (3) adapting to new domains. In this section, we discuss existing research on each of the problems.

In traditional machine learning models, it is considered that the training data on which the model is trained has similar data distributions to the data at the test time. Due to data heterogeneity and the dynamic nature of social media platforms, it is not possible to use traditional machine learning models. In the past, researchers have addressed these issues through a new branch of machine learning called Transfer Learning. In transfer learning, given data from the source domain and target domain, models are trained on a source domain and the underlying knowledge is transferred to target domain [Pan and Yang, 2010]. Different supervised, unsupervised and semi-supervised methods have been proposed for a wide variety of applications such as image classification [Tan et al., 2015], WiFi-localization on time variant data [Pan et al., 2008], and web document classification [He et al., 2009; Pan et al., 2010]. ...