Multi-Source Trendwatching

Currently available trend watchers, such as the trends-list on Twitter, are capable of supplying a user of a list of frequently used terms. Such trend watchers are limited because the relationship between trending terms and related terms are not presented, resulting in a minimal understanding why a term is trending. Elvers et al. [2011] proposed a trend watcher that is capable of creating networks between trending terms and related terms on textual YouTube content. In this paper we have applied their method to Twitter data. A disadvantage is that the resulting trends are based one resource solely. We propose an extension that uses the TF-IDF ratings of words in data from news websites to improve the relevance of the trends. We found that while the trend watcher as described in Elvers et al. [2011] was capable of detecting relevant trends, over 62% of the found trends could not be defined. The addition of the news filter resulted in an overall decrease of 3,6% of unidentifiable trends and increased the amount of news related trends from 8,3% to 13,7%. However, it also indiced a delay in detecting some news events. These results show that the method of Reed et al. can be used for detecting trends in Twitter. Also, if a bias towards news related content is preferred over the prematurity of the trends, the proposed news filter can improve results. This paper was written as part of the honours project 2013-2014 of the Bachelor Artificial Intelligence at the University of Amsterdam