Social Engagement at Scale

This project examines the dynamics of human attention, in the times where so much of it is aggregated and commoditized by the online social networks. The digital townhalls of Twitter, Facebook and Instagram track our collective attention via an increasingly diverse set of engagement metrics. The first study, proves it is possible to advance the state-of-the-art in virality prediction without compromising on explainability, robustness or privacy compliance. The approach combines content signal available at the time of posting, with high accuracy ground-truth and sentiment analysis in 18 languages to achieve stateof-the-art results on multiple benchmark datasets. The second study, questions virality as the best predictor of social influence. We examine a diverse set of content engagement metrics from Twitter. Correlations discovered lead us to propose a new, more holistic, one-dimensional engagement signal. We then show it is more predictable than any individual influence metric previously investigated. We propose the ability to engage the audience as a new, more holistic target for social influence maximization and share the compound engagement workflow to ensure reproducibility. In the third study, we examine the transferability of the proposed framework beyond Twitter. We address the problem of multi-modal popularity prediction on Instagram. We use deep neural networks to advance user-generated content representation. Through the ablation of transfer learning, we offer a detailed explanation of popularity dynamics. The models of virality, engagement and popularity are the first to achieve strong ranking performance in a robust and explainable way. The compound engagement model, in particular, is the first to explain half of the variance with features available early, and to offer strong ranking performance simultaneously. I deliver new models of understanding, via scientific avenues and Microsoft cloud services. The era of big data offers significant advancements in data collection, storage and analysis methods, creating new opportunities for researchers to achieve high relevance and impact. Extracting knowledge from social big data, however, remains extremely difficult. Much of the recent work is still plagued by anecdotal evidence from short timeframe samples or black-box approaches, while the relevant technology, data and knowledge appear siloed in separation. One ambition of this Industrial PhD project is to rise above the divide between engineering and science and prove the potential of a holistic approach. The proposed data collection and analysis framework positioned this project among the largest studies on social media to date. The proposed model operationalization framework enabled Microsoft customers to respond to pre-viral content, including support request, before anyone else.