Multi-scale Basis Learning

Under construction

This work, described here, is an outgrowth of an earlier method, proposed in our VLDB03 paper. Upon further experimentation, we found that most of the benefit was due to the sparse wavelet representation, whereas the additional benefit of fitting a forecasting model within each frequency band was typically small in comparison. However, the wavelet basis is sometimes sub-optimal in representing the key features of the time series.

This naturally lead to seeking ways to learn the basis from the data itself, rather than using

Open questions:

• The learned basis is overcomplete; can a
• TODO