Temporal Data Mining
Any information having a time component can be represented in
a general way in a temporal database. Our task is to develop a query
language that is flexible enough to access this general kind of
representation, and generate (as output) information to be processed
by a time series analysis package.
A time series is a sequence of observations on a particular variable
over time. This statistical method provides a natural way of storing
time related information in a temporal database.
Uses for temporal databases:
- Forecasting events in the future
- Analyzing patterns
We are interested in performing time series analysis on the information
held in temporal databases. Some of the examples that make time
series analysis interesting are:
- Identifying the number of workers in different job categories,
in order to plan recruiting and training
- Identifying the demand for each product line needed, for accurate
production
- Calculating patterns of minimum, maximum, etc. growth in employees'
salaries over different periods of service
- Calculating patterns of expenses in projects over different
periods of time
The Classical Multiplicative Model views a time series as being
built up of four different components. In order to identify patterns
in a time series, it is convenient to think of a time series as
consisting of several components:
- Trend
- upward or downward growth (may be linear or exponential), to
characterize the time series over a period of time
- Cycle
- refers to recurring up and down movements around trend levels.
For example, the peaks and troughs of a business cycle: expansion
followed by contraction (not necessarily affected by changes in
economic factors).
- Seasonal
- patterns that complete themselves in a year. (e.g. monthly housing
starts related to weather; increase in sales during Christmas).
- Irregular
- erratic movement in a time series, that follows no regular pattern.
(e.g. leftover or unaccountable parts after considering trend,
cycle, or seasonal variations).
We can use these components to:
- Calculate moving trend averages
- to smooth out the series
- Obtain the ratio-to-moving average
- for derivation of seasonal components (i.e., extracting periodic
fluctuations)
- Plot the final graph having trend, seasonal, and cyclical components
- Extrapolate the final graph to forecast new values for variables
The approaches described above can be used for analyzing patterns
and for forecasting.
|