DSP seminar, Wed Mar. 08 Time and location: 4:00-5:00 PM 2269 Beckman Institute

Title: An Ensemble-Learning Approach for Missing Data Recovery in Sensor Networks

Speaker: Farinaz Koushanfar

Postdoctoral Research Associate, CSL - UIUC


Abstract:

Recovery of missing data is a canonical task in sensor networks, wireless communications, 
design space exploration, active learning, and many other research and engineering areas. 
Many important tasks in sensor networks such as fault detection, data cleaning, sampling, 
compression, and topology management are addressed by utilizing the missing data recovery 
methods. Numerous techniques have been proposed for missing data recovery, including 
maximmum likelihood, expectation maximization, nonlinear function minimization, clustering, 
and multiple imputations. Most of the proposed techniques are based on estimating the multi- 
dimensional density of the variables that is subject to the problems of curse of dimensionality, 
combinatorial explosion of the possible selection of the variables, and complexity of model 
derivation. Furthermore, many of the  techniques place strong prior assumptions on the data 
and its distributions. Often, such assumptions about the data are not justified for the recovery 
of missing sensor data and are rarely effective. 

I will talk about the complexity of the missing data recovery problem and prove that the 
problem is NP-complete. I will mention my few attempts to address the recovery of missing 
data. Specifically, I will discuss a new approach to address this problem using an ensemble 
learning method while we impose a very mild set of assumptions on the sensor data and its 
distributions. The underlying idea of ensemble learning is to find simple rules to form an 
ensemble of multiple hypothesis such that the performance of each single hypothesis is improved. 
I will present a way to map the missing data recovery to enseble learning of hypotheses and 
will give insights as to why the technique works in multi- dimensional environments. The 
effectiveness of the technique is demonstrated on retreiving the missing data on the traces 
of actually deployed sensor networks.


Biography: N/A