Predicting Responses to Microblog Posts

Predicting Responses to Microblog Posts Yoav Artzi ∗ Computer Science & Engineering University of Washington Seattle, WA, USA yoav@cs.washington.edu Patrick Pantel, Michael Gamon Microsoft Research One Microsoft Way Redmond, WA, USA {ppantel,mgamon}@microsoft.com Abstract Microblogging networks serve as vehicles for reaching and influencing users. Predicting whether a message will elicit a user response opens the possibility of maximizing the viral- ity, reach and effectiveness of messages and ad campaigns on these networks. We propose a discriminative model for predicting the like- lihood of a response or a retweet on the Twit- ter network. The approach uses features de- rived from various sources, such as the lan- guage used in the tweet, the user’s social net- work and history. The feature design process leverages aggregate statistics over the entire social network to balance sparsity and infor- mativeness. We use real-world tweets to train models and empirically show that they are ca- pable of generating accurate predictions for a large number of tweets. 1 Introduction Microblogging networks are increasingly evolving into broadcasting networks with strong social as- pects. The most popular network today, Twitter, re- ported routing 200 million tweets (status posts) per day in mid-2011. As the network is increasingly used as a channel for reaching out and marketing to its users, content generators aim to maximize the impact of their messages, an inherently challeng- ing task. However, unlike for conventionally pro- duced news, Twitter’s public network allows one to observe how messages are reaching and influencing users. One such direct measure of impact are mes- sage responses. ∗ This work was conducted at Microsoft Research. In this work, we describe methods to predict if a given tweet will elicit a response. Twitter provides two methods to respond to messages: replies and retweets (re-posting of a message to one’s follow- ers). Responses thus serve both as a measure of dis- tribution and as a way to increase it. Being able to predict responses is valuable for any content gener- ator, including advertisers and celebrities, who use Twitter to increase their exposure and maintain their brand. Furthermore, this prediction ability can be used for ranking, allowing the creation of better op- timized news feeds. To predict if a tweet will receive a response prior to its posting we use features of the individual tweet together with features aggregated over the entire so- cial network. These features, in combination with historical activity, are used to train a prediction model. 2 Related Work The public nature of Twitter and the unique char- acteristics of its content have made it an attractive research topic over recent years. Related work can be divided into several types: Twitter Demographics One of the most fertile av- enues of research is modeling users and their inter- actions on Twitter. An extensive line of work char- acterizes users (Pear Analytics, 2009) and quantifies user influence (Cha et al., 2010; Romero et al., 2011; Wu et al., 2011; Bakshy et al., 2011). Popescu and Jain (2011) explored how businesses use Twitter to …

Content truncated to 500 words