1.0 - 8.00 - ShapeletSupervised Arguments - Teradata Vantage

Teradata® Vantage Machine Learning Engine Analytic Function Reference

Product
Teradata Vantage
Release Number
1.0
8.00
Release Date
May 2019
Content Type
Programming Reference
Publication ID
B700-4003-098K
Language
English (United States)
IDColumn
Specify the name of the column in input_data_table and input_categories_table that contains the unique identity of a time series.
TimeColumn
Specify the name of the input_data_table column that contains the time axis of the data.
TargetColumn
Specify the name of the input_data_table column that contains the data points.
CategoryColumn
Specify the name of the input_categories_table column that contains the category (class) of the time series.
SaxSymbolsPerWindow
[Optional] Specify the SAX argument SymbolsPerWindow, which specifies the number of SAX code symbols to create from a window. The symbols_per_window must an INTEGER in the range [1, 1000000]. If the symbols_per_window is greater than the length of the shortest time series in input data set (d), its value becomes d.
Default: 10
SaxMinWindowSize
[Optional] Specify the SAX argument WindowSize , which specifies the size of the sliding window. The min_window_size defines the length (number of data points) of the shortest shapelet; the minimum span (time series length) used to distinguish two time series from each other.
The min_window_size must be an integer in the range [1, 1000000]. If the min_window_size is greater than the length of the shortest time series in input data set (d), its value becomes d. If min_window_size is smaller than symbols_per_window, its value becomes symbols_per_window.
Default: 10
SaxMaxWindowSize
[Optional] Specify the SAX argument WindowSize , which specifies the size of the sliding window. The max_window_size defines the length of the longest shapelet; the maximum span used to distinguish two time series from each other. The max_window_size must be an integer in the range [1, 1000000] that is greater than or equal to min_window_size.
If the max_window_size is greater than the length of the shortest time series in input data set (d), its value becomes d.
A greater difference between min_window_size and max_window_size increases the probability of identifying better shapelets at the cost of higher execution time. The function uses this formula to compute the number of sliding windows, n:

n = ((max_window_size min_window_size) / symbols_per_window)+1

The maximum value of n is 20.

Default: 70
SaxOutputFrequency
[Optional] Specify the SAX argument OutputFrequency, which specifies the number of data points to skip between successive sliding windows. The gap_between_windows must be an integer in the range [1, 1000]. A smaller value increases accuracy (the chance of distinguishing time series from each other) at the cost of higher execution time.
Default: 10
RandomProjections
[Optional] Specify the number of iterations required for random masking of SAX words during shapelet training. The projections must be an INTEGER in the range [1, 40].
Default: 10

Specifying a greater projections for a longer input time series increases the probability of identifying better shapelets at the cost of higher execution time.

ShapeletCount
[Optional] Specify the maximum number of shapelets in the output model table. The num_shapelets must be an INTEGER in the range [1, 100000].
Default: 20
TimeInterval
[Optional] Specify the number of data points in a time series to skip between consecutive time series windows when calculating the distance of a shapelet from a time series.

The function builds a shapelet classification tree based on the distance of a shapelet from the time series data. Because a shapelet is typically much smaller than a complete time series, the function calculates the distance of a shapelet from a time series by sliding the shapelet across time series windows of shapelet length, calculating the distance between the shapelet and each window, and then selecting the smallest distance.

The num_data_points is the number of data points to skip when sliding from one time series window to the next. The num_data_points must be an INTEGER in the range [1, 1000000]. The value 1 gives optimal results at the cost of higher execution time.

Default: 10

Seed
[Optional] Specify the random seed the algorithm uses for repeatable results (for more information, see Nondeterministic Results). The seed must be an INTEGER in the range [1, 100000].
Default: 23