Mode determination for multivariate time series data

摘要

Embodiments relate to mode determination for multivariate time series data. An aspect includes determining first within-mode and first cross-mode parameters for a first number of modes, each mode comprising one or more time periods in the multivariate time series. Another aspect includes determining a first likelihood of the at least one multivariate time series based on the first sets of within-mode parameters and first set of cross-mode parameters. Another aspect includes determining second within-mode and second cross-mode parameters for a second number of modes. Another aspect includes determining a second likelihood of the at least one multivariate time series based on the second sets of within-mode parameters and second set of cross-mode parameters. Another aspect includes based on the first likelihood being higher than the second likelihood, selecting the first number of modes to model the at least one multivariate time series.

主权项

1. A method for mode determination for multivariate time series data, comprising:
receiving at least one multivariate time series comprising historical data; determining a first number of modes for the at least one multivariate time series, each mode comprising one or more time periods in the multivariate time series; for each mode of the first number of modes, determining a first respective set of within-mode parameters that describe behavior of the at least one multivariate time series during the one or more time periods corresponding to the mode; determining a first set of cross-mode parameters for the first number of modes that describe behavior of the at least one multivariate time series across the first number of modes; determining a first likelihood of the at least one multivariate time series based on the first sets of within-mode parameters and first set of cross-mode parameters; determining a second number of modes for the at least one multivariate time series; for each mode of the second number of modes, determining a second respective set of within-mode parameters that describe behavior of the at least one multivariate time series during the one or more time periods corresponding to the mode; determining a second set of cross-mode parameters for the second number of modes that describe behavior of the at least one multivariate time series across the second number of modes; determining a second likelihood of the at least one multivariate time series based on the second sets of within-mode parameters and second set of cross-mode parameters; and based on the first likelihood being higher than the second likelihood, selecting the first number of modes to model the at least one multivariate time series.