Periodic Topic Mining from Massive Amounts of Data

Social media keeps growing and providing us with rich sources of information to understand our everyday lives, customs, and culture in the form of periodic topics. This paper proposes a method of detecting periodic topics based on autocorrelation using the time series of the document frequencies of keywords. To deal with the massive amount of data collected from social media, this method is implemented using Hadoop, which is an open-source framework for distributed processing and data storage. The implementation is evaluated in comparison with a relational database management system. Using this method, this paper analyzes blogs, news sites, and spam as information sources which serve as social and cultural indicators. Data is collected from Japanese blogs and news sites, and spam blogs are then separated from legitimate blogs using a spam filtering system. Distribution periods of keywords within each information source and weekly keywords are then extracted, and the characteristics of each information source are illustrated in terms of distribution and keywords. The results obtained using this extraction method indicate that periodic blog topics tend to be TV programs, hobbies, and social events; periodic news topics tend to be political and economic events; and periodic topics in spam tend to be automatically copied-and-pasted e-mail newsletters and affiliate offers.