Hossein Keshavarz


Sep
06
2013

Add to Cal
  • Host Department: Statistics
  • Date: 09/06/2013
  • Time: 2:00PM

  • Location: 438 WH

  • Description:

    Contextual anomaly and outlier detection is about to finding the peculiar phenomenons
    which don’t conform the normal pattern. Anomaly detection has vast applications
    in many areas such as fraud detection for credit cards, disease outbreak detection,
    sensor networks and environmental science. We address the anomaly detection
    problem in high dimensional spatio-temporal data. This study is motivated by finding
    non-conforming patterns in environmental data sets such as NDVI and CO2 flux
    linkage. In the first part, we propose two class of detection algorithms based on the
    low rank and sparse Fourier representation of data. The simulation studies show
    that the suggested algorithms can efficiently detect the abnormal patterns. We also
    state the theoretical conditions under which the sparse Fourier anomaly detection
    scheme attains the minimax consistency rate. The second part is devoted to study
    the fused lasso signal approximator (FLSA) as a change point detection algorithm.
    We present some upper bounds on `1 and `2 norm of estimation error of convex and
    non-convex FLSA under possible model mismatch.

  • On the anomaly detection in the spatio-temporal data