Invited Talk, Hanyang University (ERICA), Ansan, South Korea
Introducing theory and techniques of data analysis and machine learning, from traditional to state-of-art, and the research on machine-learning-based applications in Social Science.
Invited Talk, Hanyang University (ERICA), Ansan, South Korea
Introducing basic graph theroy and how the theory and its techniques can be used in state-of-art research on Computational Social Science.
Tutorial, Montreal, Canada
The word-of-mouth diffusion has been regarded as an important mechanism to advertise a new idea, image, technology, or product in online social networks (OSNs). In this talk, I introduce the study on the prediction of popular and viral image diffusion in Pinterest. We first characterize an image cascade from two perspectives: (i) volume – how large the cascade is, i.e., total number of users reached, and (ii) structural virality – how many users in the cascade are responsible for attracting other users. Our model predicts whether an image will be (a) popular in terms of the volume of its cascade, or (b) viral in terms of the structural virality. Our analysis reveals that a popular image is not necessarily viral, and vice versa. This motivates us to investigate whether there are distinctive features for accurately predicting popular or viral image cascades. To predict the popular or viral image cascades, we consider the following feature sets: (i) deep image features, (ii) image meta and poster’s information, and (iii) initial propagation pattern. We find that using deep image features alone is not as effective in predicting popular or viral image cascades. We show that image meta and poster’s information are strong predictors for predicting popular image cascades while image meta and initial propagation patterns are useful to predict viral image cascades. We believe our exploration can give an important insight for content providers, OSN operators, and marketers in predicting popular or viral image diffusion.
Talk, Stanford University, California, USA
It becomes the norm for people to communicate with one another through various online social channels, where different conversation structures are formed depending on platforms. One of the common online communication patterns is a threaded conversation where a user brings up a conversation topic, and then other people respond to the initiator or other participants by commenting, which can be modeled as a tree structure. In this talk, I introduce the work which investigates (i) the characteristics of online threaded conversations in terms of volume, responsiveness, and virality and (ii) what and how content properties and user participation behaviors are associated with such characteristics. To this end, we collect 700 K threaded conversations from 1.5 M users in Reddit, one of the most popular online communities allowing people to communicate with others in the form of threaded conversations. Using the collected dataset, we find that ‘social’ words, difficulties of texts, and document relevancy are associated with the volume, responsiveness, and virality of conversations. We also discover that large, viral conversations are mostly formed by a small portion of users who are reciprocally communicate with others by analyzing user interactions. Our analysis on discovering user roles in conversations reveal that users who are interested in multiple topics play important roles in large and viral conversations, whereas heavy posting users play important roles in responsive conversations. We expand our analysis to topical communities (i.e., subreddits) and find that news-related, image-based, and discussion-related communities are more likely to have large, responsive, and viral conversations, respectively.