CONVEX: Similarity-Based Algorithms for Forecasting Group Behavior

Abstract

A proposed framework for predicting a group's behavior associates two vectors with that group. The context vector tracks aspects of the environment in which the group functions; the action vector tracks the group's previous actions. Given a set of past behaviors consisting of a pair of these vectors and given a query context vector, the goal is to predict the associated action vector. To achieve this goal, two families of algorithms employ vector similarity. CONVEXk _NN algorithms use k-nearest neighbors in high-dimensional metric spaces; CONVEXMerge algorithms look at linear combinations of distances of the query vector from context vectors. Compared to past prediction algorithms, these algorithms are extremely fast. Moreover, experiments on real-world data sets show that the algorithms are highly accurate, predicting actions with well over 95-percent accuracy.