Analysis and Forecasting of Trending Topics in Online Media Streams Tim Althoff Damian Borth Jörn Hees Andreas Dengel German Research Center for Artificial Intelligence (DFKI) D-67663 Kaiserslautern, Germany [email protected], {damian.borth, joern.hees, andreas.dengel}@dfki.de ABSTRACT @total 1000 @google_news_usa Among the vast information available on the web, social me- @google_news_germany @google_search_usa dia streams capture what people currently pay attention to 800 @google_search_germany and how they feel about certain topics. Awareness of such @google_trends_usa @twitter_usa trending topics plays a crucial role in multimedia systems 600 @twitter_germany such as trend aware recommendation and automatic vocab- @twitter_daily @wiki_en ulary selection for video concept detection systems. 400 @wiki_de Correctly utilizing trending topics requires a better under- standing of their various characteristics in different social 200 media streams. To this end, we present the first compre- 0 hensive study across three major online and social media Jul 24 2012 Jul 31 2012 Aug 07 2012 Aug 14 2012 streams, Twitter, Google, and Wikipedia, covering thou- sands of trending topics during an observation period of an Figure 1: The trending topic “Olympics 2012” during sum- entire year. Our results indicate that depending on one’s re- mer 2012. The colored curves represent the contribution of quirements one does not necessarily have to turn to Twitter the different online and social media streams. for information about current events and that some media streams strongly emphasize content of specific categories. As our second key contribution, we further present a novel ap- 1. INTRODUCTION proach for the challenging task of forecasting the life cycle of Currently, as multimedia content consumption is rapidly trending topics in the very moment they emerge. Our fully increasing on the internet we can consider ourselves as living automated approach is based on a nearest neighbor forecast- in the zettabyte era1. To a large degree this vast amount of ing technique exploiting our assumption that semantically information is created, shared, and exchanged by the users similar topics exhibit similar behavior. themselves in social multimedia systems. These systems pro- We demonstrate on a large-scale dataset of Wikipedia vide us with streams of information that capture how we page view statistics that forecasts by the proposed approach are spending our time and what we are talking about. Over are about 9-48k views closer to the actual viewing statistics time, different topics arise in these media streams reflecting compared to baseline methods and achieve a mean average changing interests of groups of individuals. In this sense, percentage error of 45-19 % for time periods of up to 14 days. social media streams mirror our society. These streams con- tain rich information and immediate feedback about to what people currently pay attention to and how they feel about Categories and Subject Descriptors certain topics. H.3.3 [Information Storage and Retrieval]: Information Such trending topics [13] can be associated with a subject Search and Retrieval —Information filtering (i.e. a textual label) and experience sudden increases in user popularity (“trending”). They usually correspond to real world events such as sport events (Olympics 2012), product Keywords releases (iPhone 5), celebrity news (Steve Jobs’ death), inci- Trending Topics, Social Media Analysis. Lifecycle Forecast, dents (Sinking of the Costa Concordia), political movements Twitter, Google, Wikipedia (Occupy), and entertainment (Academy Awards). Aware- ness of trending topics plays a key role in building social Permission to make digital or hard copies of all or part of this work for personal or multimedia systems satisfying users’ information needs and classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation enhancing the multimedia consumption experience [17]. on the first page. Copyrights for components of this work owned by others than the Today, a large variety of social multimedia systems is author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or available to users on the web such as Twitter, YouTube, republish, to post on servers or to redistribute to lists, requires prior specific permission Facebook, Flickr, Pinterest, Tumblr, Google News/Trends, and/or a fee. Request permissions from [email protected]. and Wikipedia serving different kinds of needs such as in- MM’13, October 21–25, 2013, Barcelona, Spain. formation demand, social communication, as well as sharing Copyright is held by the owner/author(s). Publication rights licensed to ACM. ACM 978-1-4503-2404-5/13/10 ...$15.00. 1 http://dx.doi.org/10.1145/2502081.2502117. Cisco, The Zettabyte Era. May 30, 2012. and consumption of multimedia content. Because of this, 2. RELATED WORK many users turn to multiple of these online platforms, de- The overview over related work is divided into three parts. pending on their present individual needs, creating a hetero- We first report work that analyzes or combines information 2 geneous multi-channel environment within the social media from multiple media channels. Then, we focus on forecasting landscape. While researchers have a general intuition about behavioral dynamics before finally presenting applications of the nature of these channels (e.g. many current event de- trending topics to social multimedia systems. tection systems are based on Twitter feeds [14, 22]), stud- ies of the exact distribution of information across multiple Multi-Channel Analyses. channels have not been in the scope of the multimedia com- Detecting and tracking topics in news media has been munity until now. To the best of our knowledge, there does studied since 1997 with the goal of understanding news in not exist a study on the behavior and lifecycle of trending multiple languages and across multiple media channels (in- topics across multiple media channels and on the correla- cluding television and radio sources) [2]. tions between topic categories and media channels over a one Ratkiewicz et al. [16] argue that online popularity of spe- year observation period prior to this work. For example, the cific topics (e.g. “Barack Obama”) cannot in general be char- trending topic “olympics 2012” (the number one trend) man- acterized by the behavior of individual news-driven events ifests itself in multiple media channels which clearly exhibit (e.g. “Barack Obama inaugurated as U.S. President”) since different behavior (as illustrated in Figure 1). the former might subsume many different news stories. They A second and much more challenging task is forecasting further find bursts in Wikipedia traffic to be correlated with the life cycle of a trending topic, i.e. forecasting the amount bursts of Google search volume. of user engagement towards it in the very moment they Wikipedia article views have also been correlated with emerge. These forecasts enable us to anticipate changing behavior on Twitter to analyze the effectiveness of creating information needs of users ahead of time and to allocate new content for breaking news (e.g. Japan earthquake) or limited resources in trend aware multimedia systems such updating pages at the moment of news (e.g. Oscar winners) as recommender systems [9, 17] or video concept detection [22] and to filter spurious events on Twitter which indirectly systems [5]). For example, such limited resources include resulted in the finding that Wikipedia lags behind Twitter time of content consumers and producers, screen estate and by about two hours [14]. Adar et al. correlate several be- content recommendations on web pages [17], processing time havioral datasets such as general queries from an observation for training models in video concept detection [5], and server period of one month from MSN and AOL search logs with load and replication capacity [21]. the general idea of understanding past behavior to predict While obviously of great use, forecasting trending topics future behavior [1]. is a very challenging problem since the corresponding time The work of Yang and Leskovec [23] explores patterns of series usually exhibit highly irregular behavior also known temporal variation in online media by clustering them to as structural breaks that can lead to large forecasting errors analyze the rise and fall of human attention with regard to and unreliable models [7]. The complexity of this task will certain phrases or memes. They report six distinct temporal be illustrated in more detail in Sec. 4 (Figure 5). patterns and find that weblogs trailed mainstream media by We present a novel forecasting approach exploiting the as- one or two hours for most of the considered phrases. sumption that semantically similar topics exhibit similar be- In contrast to the above, we explicitly focus our study on havior in a nearest neighbor framework. Semantically sim- trending topics in online and social media channels over a ilar topics are discovered by mining topics of similar type long observation period. Similar to [8] using trending topics and category on DBPedia [3], a structured representation of such as “oil spill” and “iPhone” to evaluate tweet selection, the information on Wikipedia. We evaluate our fully auto- we utilize trending topics published by the media channels matic approach on a large-scale dataset of billions of views themselves and focus on multi-channel analysis and forecast. of several million articles on Wikipedia. We further analyze the correlation between topic categories Summarizing, our contribution in this paper is three-fold. and media channels and propose a fully automatic approach We present (1) the first comprehensive study of trending of forecasting trending topics in terms of time series (as op- topics in three major social and online media channels with posed to visualization tools [1] or predicting temporal clus- an observation period of one year, (2) a fully automatic ters [23]). forecasting technique for these trending topics based on a nearest neighbor approach exploiting semantic relationships Forecasting Behavioral Dynamics.