Node Behavior Prediction for Large-Scale Approximate Information Filtering

In this paper we investigate methods that allow us to identify the publishing behavior of individual nodes in large-scale distributed information filtering systems. The work presented here is based on our system MAPS (Minerva Approximate Publish/Subscribe), a novel approach to support approximate information filtering func-tionality in a peer-to-peer environment. In MAPS, a user subscribes to and monitors only carefully selected publisher nodes, and re-ceives notifications from these information sources only. In this way, document-granularity dissemination is known from exact in-formation filtering approaches is avoided, and the system is able to support very high publication rates. However this scalability bene-fits come at the cost of lower recall. To improve node selection and thus recall, in previous work we have proposed a ranking method that predicts nodes ’ publishing behavior based on time-series anal-ysis techniques. In this work, we focus on the prediction parameters used by these techniques and show how their choice can affect the overall recall achieved. Additionally, we demonstrate how these parameters can be computed in a per node fashion, using only lo-cal information available at each node. Our method incurs no extra message traffic in the network and experiments show improvements in both prediction error and recall. When compared to an oracle op-ponent and an opponent with global information, our method per-forms almost as good, by resorting only to local node computa-tions. 1