Speaker: Mike Conover (LinkedIn)
Title: Building Machine Learning Systems at LinkedIn
Room: Informatics East 122
Abstract: This talk details patterns and machine learning systems to provide our members with actionable, relevant opportunities to nurture their professional networks. Featuring the Connected mobile app as an in-depth case study of how to combine compute-intensive features describing billions of relationships with information that isn’t known until the moment a user opens the app, in this talk we’ll discuss the architectural, modeling, and experimentation patterns leveraged by the Connected relevance team to rank and serve mobile content. Additionally, this session will touch on the human element of machine learning product development, outlining collaboration and communication patterns for working effectively across the organization – from reporting and documentation to evangelism, skills transfer and user experience research. Taken together, these insights will provide a detailed picture of some of LinkedIn’s best practices for building data products at a global scale.
Biography: A graduate of the Indiana University School of Informatics and Computing, Mike Conover builds machine learning technologies that leverage the behavior and relationships of hundreds of millions of people. A senior data scientist at LinkedIn, Mike has a Ph.D. in complex systems analysis with a focus on information propagation in large-scale social networks. His work has appeared in the New York Times, the Wall Street Journal, Forbes, Science, MIT Technology Review and on National Public Radio.
Prof. Luis Rocha from CNETS at IU Bloomington, Prof. Lang Li from IUPUI Medical School, and Prof. Hagit Shatkay from the University of Delaware have been awarded a four-year, $1.7M grant from NIH/NLM to study the large-scale extraction of drug-Interaction from medical text. Drug-drug interaction (DDI) leads to adverse drug reactions, emergency room visits and hospitalization, thus posing a major challenge to public health. To circumvent risk to patients, and to expedite biomedical research, both clinicians and biologists must have access to all available knowledge about potential DDI, and understand both causes and consequences of such interactions. However, mere identification of interactions does not directly support such understanding, as evidence for DDI varies broadly, from reports of molecular interactions in basic-science journals, to clinical descriptions of adverse-effects in a myriad of medical publications. This project will develop tools that focus directly on large-scale identification and gathering of various types of reliable experimental evidence of DDI from diverse sources. The successful completion of the project will provide clinicians and biologists with substantiated knowledge about drug interactions and with informatics tools to obtain such information on a large-scale, laying the basis for preventing adverse drug reactions and for exploring alternative treatments.
The DESPIC team at the Center for Complex Systems and Networks Research (CNetS) presented a demo of a new tool named BotOrNot at a DoD meeting held in Arlington, Virginia on April 23-25, 2014. BotOrNot (truthy.indiana.edu/botornot) is a tool to automatically detect whether a given Twitter user is a social bot or a human. Trained on Twitter bots collected by our lab and the infolab at Texas A&M University, BotOrNot analyzes over a thousand features from the user’s friendship network, content, and temporal information in real time and estimates the degree to which the account may be a bot. In addition to the demo, the DESPIC team (including colleagues at the University of Michigan) presented several posters on Scalable Architecture for Social Media Observatory, Meme Clustering in Streaming Data, Persuasion Detection in Social Streams, High-Resolution Anomaly Detection in Social Streams, and Early Detection and Analysis of Rumors. See more coverage of BotOrNot on PCWorld, IDS, BBC, Politico, and MIT Technology Review.