Prediction of Influencers from Word Use Chan Shing Hei

Detecting influential users before they show observable signals of influence Method: psycholinguistic category scores from word usage Twitter Data  built predictive models of influence from such category based features Introduction

Measuring Influence measure influence score  the average number of retweets generated from a user's tweets

Dataset Twitter's streaming API Nov 1, 2013 to Nov 14, 2013 randomly sample 1000 users Tweets from last one month (Oct 2013) historical tweets (max 200)  average influence score: SD: 0.098

Psycholinguistic Analysis from text How to measure word use ?  users’ historical tweets with the Linguistic Inquiry and Word Count (LIWC) 2001 dictionary How to computed his/her LIWC based scores?  in each category as the ratio of the number of occurrences of words in that category in one’s tweets and the total number of words in his/her tweets

Influence and Word Use

Finding from analysis LIWC category that negatively correlated with influence score: Negative emotion physical states inhibition

LIWC category that positively correlated with influence score: more interactive positive feelings or emotion determination and desires for the future Finding from previous analysis

Prediction Models Using Weka regression analysis and a classification  influence score Evaluate the performance

Regression analysis linear regressions  predict influence score using LIWC measures

Classification study 1 supervised binary machine learning algorithms

divide influence scores into 10 equal sized bins, and trained supervised classifiers with 10 classes Classification study 2

Suggestion Adding more criteria for measure influence Other social media Real application – political campaigns

Conclusion Correlations of word usage with influence behavior discovers a set of psycholinguistic categories identify users early to be an influencer