The Interplay of Internal and External Traffic on Online Networks in the Prediction of Collective Attention

Understanding the interplay between internal and external traffic across networks is an important issue in the analysis of dynamic processes on graphs, with applications across a range of disciplines, including the sociological contexts of models of online popularity, attention prediction for news events, and collective memory. Previous related research has used time series data to learn graphs [1], or typically applies multivariate forecasting techniques without explicit consideration of underlying network structure [2]. In this work, we reconstruct the time series for a process on a target node using a known graph and the time series of its neighbours. We apply our models (Figure 1) to Wikipedia data; the underlying network of articles connected by hyperlinks, with edge weights the aggregate monthly traffic across each link, and the processes on each node as the time series for daily page views. In effect, using traffic on the network to assist in predicting traffic to nodes from external sources.
Current results with the models reproduce a given article’s time series with a median of 21% symmetric mean absolute percentage error and 46% explained variance, despite just 17% of traffic being internal to the Wikipedia article network. Given the success of the current research, we will further develop our models towards the problem of time series forecasting. Network interactions have been underutilised in studying online popularity and attention prediction, yet these results indicate that ‘sympathetic’ responses of page neighbours are likely a rich resource for accuracy improvements.