April 21, 2006: Tyler Neylon, CIMS
Linear Predictions and Occam's Razor
A time series is, mathematically, a sequence of real numbers which is
revealed to us one element at a time. Examples abound: stock prices,
temperature readings, medical data, an audio stream, etc. In many
cases, we are interested in learning a pattern among a set of many
time series, such as among prices in a stock market. In this talk, we
will explore the possibility of learning linear identities among time
series. In particular, we will focus on the discovery of _sparse_
linear identities  those which depend on only a small number of
time series (hence Occam's Razor). We will see theoretical evidence
that sparsity is indeed a good indicator of predictive power in our
model, as well as an efficient incremental algorithm for updating
predictions as the time series evolve.
