UCLA

The nature and size of data on the social web are different from those used to develop most algorithms for statistics, data mining, and machine learning. The social web is far more dynamic than actuarial tables, lab tests, or historical records, in that social algorithms will affect people's behavior. Google's PageRank, for instance, has undergone hundreds of announced changes. A multibillion-dollar search engine optimization (SEO) industry has been created to adapt organizations to social algorithms. To address the adaptive nature of the social web, I present an "Architecture of Participation": a methodology and toolkit that allow one to measure, visualize, and evaluate the effectiveness of a given social-web medium and to better understand two (overlapping) sets of factors: those that influence user participation in generating content, and those that influence the quality of that content. The goal of my Architecture of Participation, then, is to better measure the dynamics of social networks in the context of these factors, with the subsidiary objective of more effectively utilizing collective intelligence in biology.

Behavior in a social network can be dynamically affected by friends or contacts, community, time of day, and many other factors that reflect the state of the network as a whole. The data coming from a social network, in short, manifests a complex, dynamic system. Hence, measurement tools must take the whole system into account, not just its pieces, as well as the extent of the social system memory summarized by its current state. To this end, the Architecture of Participation enables one to measure how system state affects social algorithms and user behavior. This Architecture includes semantic lexica, algorithms, and software that allow one to incorporate the dynamics of data and the system state into social analytics. I show that information-theoretic measures such as Rényi entropy and mutual information can be used classify an infinite set of states. I create unsupervised efficient, scalable and parallelizable algorithms to classify text in "big data." I then develop a calculus that allows the combination these classifiers in to a search kernel in an intuitive and mathematically consistent way.