Bilkent University
Department of Computer Engineering


Machine Learning for Decision Making in Energy and Crowdsourcing Systems


Şeyda Ertekin, Ph.D
Computer Scientist
Sloan School of Management
Massachusetts Institute of Technology (MIT)

In today’s digital world, where we are surrounded by vast amount of data, Machine Learning is at the heart of almost every scientific discipline. The challenge of making accurate predictions from this data and inferring useful information to improve decision making is of utmost importance for many scientists, engineers and businesses. In this talk, I will discuss two different application domains where novel machine learning and statistical prediction algorithms take central stage in making accurate predictions from the data. I will show that making accurate predictions enable to build cost-efficient systems by allocating valuable resources wisely.

In the first part of my talk, I will discuss Crowdsourcing where collective intelligence of the large crowds is leveraged to accomplish a task. I will present a novel online algorithm, called “CrowdSense” to approximate the crowd. The problem of approximating the crowd" is that of estimating the crowd's majority opinion on a budget, -that is by querying only a subset of the crowd. Our goal is to determine the majority opinion of a crowd, on a series of questions with a limited budget. CrowdSense dynamically samples subsets of labelers based on an exploration/exploitation criterion. The algorithm produces a weighted combination of a subset of the labelers' votes that approximates the crowd's opinion. I will also introduce two variations of CrowdSense that make various distributional assumptions to handle distinct crowd characteristics.

In the second part of my talk, I will take you to the world of underground Energy Grids. I will present a new statistical model for predicting discrete events over time, called Reactive Point Processes (RPP's). RPP's are a natural fit for many different domains, but their development was motivated by the problem of predicting serious events such as fires, explosions and power failures in the underground electrical grid of New York City. Power grid reliability is today’s major source of concern as the electrical grid is aging but the demand for electricity is increasing. The order in which the electrical grid’s parts are inspected is critical to preventing serious events and power failures. The most vulnerable parts of the grid should be inspected and repaired with the highest priority before a serious event occurs in order to ensure electrical service reliability and public safety. I will show that RPP’s are very effective in modeling the vulnerability of the electric structures in the grid by using the historical data of events and inspections. With RPP’s serious events at the energy grid can be predicted in the short term, which permits the energy companies i) to take preventive action to keep vulnerability levels to the serious events low for the reliability of the power grid and ii) to help make broader policy decisions for power grid maintenance.

Bio: Seyda Ertekin currently works as a computer scientist at Massachusetts Institute of Technology (MIT) and is the co-PI of the Big Data analysis projects at MIT Sloan. She received her BSc degree in Electrical & Electronics Engineering from Middle East Technical University (METU) and PhD degree in Computer Science & Engineering from the Pennsylvania State University - University Park. During her PhD studies, she worked as a researcher at the Machine Learning group of the NEC Research Laboratories in Princeton, NJ. She also worked at wireless communication projects at ASELSAN Inc, in Turkey. Her current research at MIT bridges theoretical and experimental computer science to design machine learning algorithms and statistical models to solve real-world problems in the fields of data mining and information retrieval. At MIT, she is the member of the first prize winner team at the INFORMS 2013 Innovative Applications in Analytics Award.


DATE: 22 August, 2013, Thursday @ 13:40