Real-life transactional data-sets often involve millions of customers
and thousands of products, recorded over a period of time.
Typical market-basket tools try to use the full data or a random
sampling thereof. However, such datasets have strong temporal behaviour
especially in companies that are ramping up rapidly in terms of the customer base
and work in fast evolving markets. Therefore techniques such as
clustering have to adapt to the dynamics and trends in such data.
This paper deals with seasonality detection and data partitioning
using a novel approach based on "learned" windowing and filtering.
It quantifies how much back in time one can look at for a
given data and still treat it as describing one model. The technique
is able to capture both periodic and divergent trends in the data.
It serves as a key preprocessing stage for subsequent transformation
of seasonally partitioned and sampled high dimensional 
($>10,000$) categorical transactional data into a 
lower dimensional ($30-50$) continuous space using a graph based 
clustering approach called VBACC (GGupta01Spie,GGupta00thesis). 
A window-based approach for motion estimation
in this low dimensional cluster space is proposed along with
a simple visualization scheme. Results on an industrial data-set
are presented.