Finding Relevant Tweets

Abstract

When a user of a microblogging site authors a microblog post or browses through a microblog post, it provides cues as to what topic she is interested in at that point in time. Example-based search that retrieves similar tweets given one exemplary tweet, such as the one just authored, can help provide the user with relevant content. We investigate various components of microblog posts, such as the associated timestamp, author’s social network, and the content of the post, and develop approaches that harness such factors in finding relevant tweets given a query tweet. An empirical analysis of such techniques on real world twitter-data is then presented to quantify the utility of the various factors in assessing tweet relevance. We observe that content-wise similar tweets that also contain extra information not already present in the query, are perceived as useful. We then develop a composite technique that combines the various approaches by scoring tweets using a dynamic query-specific linear combination of separate techniques. An empirical evaluation establishes the effectiveness of the composite technique, and that it outperforms each of its constituents.