Learning from Networks of Examples

Abstract

Most machine learning algorithms assume that examples are independent of each other, but many (or most) domains violate this assumption. For example, in real markets customers’ buying decisions are influenced by their friends and acquaintances, but data mining for marketing ignores this (as does traditional economics). In this talk I will describe how we can learn models that account for example dependences, and use them to make better decisions. For example, in the marketing domain we are able to pinpoint the most influential customers, “seed” the network by marketing to them, and unleash a wave of word of mouth. We mine these models from collaborative filtering systems and knowledge-sharing Web sites, and show that they are surprisingly robust to imperfect knowledge of the network. I will also survey other applications of learning from networks of examples we are working on, including: combining link and content information in Google-style Web search; automatically translating between ontologies on the Semantic Web; and predicting the evolution of scientific communities.