Computationally Efficient Scoring of Activity using Demographics and Connectivity of Entities

Abstract

Consider a collection of entities, where each may have some demographic properties, and where the entities may be linked in some kind of, perhaps social, network structure. Some of these entities are “of interest”—we call them active. What is the relative likelihood of each of the other entities being active? AFDL, Activity from Demographics and Links, is an algorithm designed to answer this question in a computationally-efficient manner. AFDL is able to work with demographic data, link data (including noisy links), or both; and it is able to process very large datasets quickly. This paper describes AFDL’s feature extraction and classification algorithms, gives timing and accuracy results obtained for several datasets, and offers suggestions for its use in real-world situations.