Tutorial: Social Media Predictive Analytics

Time
Afternoon.
Responsible
Svitlana Volkova, Benjamin van Durme, David Yarowsky, and Yoram Bachrach.
Prerequisites
This tutorial assumes basic knowledge of probability, machine learning (supervised classification, regression, and feature engineering) and basic coding skills in Python.

Abstract

The recent explosion of social media services like Twitter, Google+ and Facebook has led to an interest in social media predictive analytics – automatically inferring hidden information from the large amounts of freely available content. It has a number of applications, including: online targeted advertising, personalized marketing, large-scale passive polling and real-time live polling, personalized recommendation systems and search, and real-time healthcare analytics etc.

In this tutorial, we will describe how to build a variety of social media predictive analytics for inferring latent user properties from a Twitter network including demographic traits, personality, interests, emotions and opinions etc. Our methods will address several important aspects of social media such as: dynamic, streaming nature of the data, multi-relationality in social networks, data collection and annotation biases, data and model sharing, generalization of the existing models, data drift, and scalability to other languages.

We will start with an overview of the existing approaches for social media predictive analytics. We will describe the state-of-the-art static (batch) models and features. We will then present models for streaming (online) inference from single and multiple data streams; and formulate a latent attribute prediction task as a sequence-labeling problem. Finally, we present several techniques for dynamic (iterative) learning and prediction using active learning setup with rationale annotation and filtering. The tutorial will conclude with a practice session focusing on walk-through examples for predicting latent user properties e.g., political preferences, income, education level, life satisfaction and emotions emanating from user communications on Twitter.

Outline

Part I
Overview of the existing approaches for social media analytics (15 mins)
Part II
Static (Batch) Prediction (30 mins)
  • Classifiers: user-based vs. tweet-based predictions
  • Features: lexical, network structure, affect, behavior etc.
  • Homophily and attribute assortativity
  • Generalization of the classifiers
Part III
Streaming (Online) Inference (30 mins)
  • Bayes’ rule: iterative Bayesian updates
  • Predictions from multiple data streams
  • Batch vs. online performance
  • Data collection/annotation biases and classification performance
  • User attribute prediction task as a sequence labeling problem
Part IV
Dynamic (Iterative) Learning and Prediction (30 mins)
  • Iterative batch retraining with and without rationale filtering
  • Active learning with and without oracle annotations
  • Active learning with interactive iterative rationale annotations
Break
(15 mins)
Part V
Practice session: building analytics for inferring user demographics, personality, emotions and opinions (1 hour)
  • Data and annotation schema description (data and models will be released)
  • Walk-through python examples for static inference – extracting features, training, testing for tweet-based e.g., emotions, opinions and user-based e.g., demographic predictions.
  • Walk-through python examples for streaming inference from multiple data streams e.g., friends, followers, user-mentions, retweets etc.

Instructors

Svitlana Volkova
Ph.D. Candidate, Center for Language and Speech Processing, Johns Hopkins University, Baltimore, MD, svitlana@jhu.edu.
Svitlana is a Ph.D. Candidate in Computer Science at the Center for Language and Speech Processing, Johns Hopkins University. She works on machine learning and natural language processing techniques for social media predictive analytics. She develops batch and streaming (dynamic) models for automatically inferring psycho-demographic profiles from social media data streams, fine-grained emotion detection and sentiment analysis for under-explored languages and dialects in microblogs, effective interactive and iterative rationale annotation via crowdsourcing.
Benjamin Van Durme
Human Language Technology Center of Excellence, Center for Language and Speech Processing, Johns Hopkins University, Baltimore, MD, vandurme@cs.jhu.edu.
Benjamin Van Durme is the Chief Lead of Text Research at the Human Language Technology Center of Excellence, and an Assistant Research Professor at the Center for Language and Speech Processing. He works on natural language processing (specifically computational semantics), predictive analytics in social media and streaming/randomized algorithms.
David Yarowsky
Professor, Center for Language and Speech Processing, Johns Hopkins University, Baltimore, MD, yarowsky@jhu.edu.
David Yarowsky is a Professor at the Center for Language and Speech Processing, Johns Hopkins University. His research interests include natural language processing and spoken language systems, machine translation, information retrieval, very large text databases and machine learning. His research focuses on word sense disambiguation, minimally supervised induction algorithms in NLP, and multilingual natural language processing.
Yoram Bachrach
Researcher, Microsoft Research, Cambridge, UK, yobach@microsoft.com.
Yoram Bachrach is a researcher in the Online Services and Advertising group at Microsoft Research Cambridge UK. His research area is artificial intelligence (AI), focusing on multi-agent systems and computational game theory. Computational game theory combines the theoretical foundations of economics and game theory with creative solutions from AI and computer science.

Three of the four organizers co-led a workshop on Social Dynamics and Personal Attributes in Social Media at ACL 2014. Two of four organizers developed a demo system for predicting later user properties from texts published in social media presented at AAAI 2015.