Real-time Forecasting of Non-linear Competing Online Activities

Abstract

Given a large, online stream of multiple co-evolving online activities, such as Google search queries, which consist of d keywords/activities for l locations of duration n, how can we analyze temporal patterns and relationships among all these activities? How do we go about capturing non-linear evolutions and forecasting long-term future patterns? For example, assume that we have the online search volume for multiple keywords, e.g., “HTML/Java/SQL/HTML5” or “Iphone/Samsung Galaxy/Nexus/HTC” for 236 countries/territories, from 2004 to 2015. Our goal is to capture important patterns and rules, to find the answer for the following issues: (a) Are there any periodical/seasonal activities? (b) How can we automatically and incrementally detect the sign of competition between two different keywords from the data streams? (c) Can we achieve a real-time snapshot of the stream and forecast long-range future dynamics in both global and local level? In this paper, we present RFC_AST, a unifying adaptive non-linear method for forecasting future patterns of co-evolving data streams. Extensive experiments on real datasets show that RFC_AST does indeed perform long-range forecasts and it surpasses other state-of-the-art forecasting tools in terms of accuracy and execution speed.