The emergence of social media and advances in mobile technology and internet

has resulted in constant connectivity across users enabling them to post, share, and engage with content published on the web. Studying and learning from such data about

users, and their engagement with content can give insights into the current and emerging trends in society. However, studying social media data comes with its own set of

unique challenges. Social media data is highly unstructured because the content is not

curated to adhere to any formal structure. This makes the process of analyzing the data

challenging. Each message published on social media has Social media data is also

highly volatile since huge volumes of data is generated every second. In this thesis, we

propose machine learning based algorithms and methodologies to accommodate these

challenges; and apply the algorithms to solve problems in domains of public health and

journalism.

Chapter 1 proposes a new framework to combine the text and user engagement

data to detect trends from social networks.

Chapter 2 studies social media data to predict the impact of news events. The

chatter on social media surrounding news events is accurately quantified, and is found

to be the most distinguishing feature between high-impact and low-impact events.

Chapter 3 uses topic modeling to discover attitudes and trends about drug abuse.