Determining a window size for outlier detection
AssigneeInternational Business Machines Corporation
InventorsJayanta Basak, Martin Keller, Laurent Sebastien Mignet, Sourashis Roy
Life of this patent
Estimating the term…
Abstract
A window size for outlier detection in a time series of a database system is determined. Strength values are calculated for data points using a set of window sizes, resulting at least in one set of strength values for each window size. The strength values increase as a distance between a value of a respective data point and a local mean value increases. For each set of strength values, a weighted sum is calculated based on the respective set of strength values. A weighting function is used to suppress the effect of largest strength values and a window size is selected based on the weighted sums.
Claims19
A method for determining a window size for outlier detection in a time series of a database system, said method comprising:
- calculating strength values for data points using a set of window sizes, resulting at least in one set of strength values for each window size, said strength values increasing as a distance between a value of a respective data point and a local mean value increases;
- calculating, for each set of strength values, a weighted sum based on the respective set of strength values using a weighting function suppressing effect of largest strength values, and
- selecting a window size based on said weighted sum.
A data processing system for determining a window size for outlier detection in a time series of a database system, said data processing system comprising:
- means for calculating strength values for data points using a set of window sizes, resulting at least in one set of strength values for at least a subset of window sizes, said strength values increasing as a distance between a value of a respective data point and a local mean value increases,
- means for calculating, for each set of strength values, a weighted sum based on the respective set of strength values using a weighting function suppressing effect of largest strength values, and
- means for selecting a window size based on said weighted sum.
A computer program product for determining a window size for outlier detection in a time series of a database system, said computer program product comprising a computer usable medium having computer usable program code embodied therewith, said computer usable program code comprising:
- computer usable program code configured to calculate strength values for data points using a set of window sizes, resulting at least in one set of strength values for at least one subset of window sizes, said strength values increasing as a distance between a value of a respective data point and a local mean value increases;
- computer usable program code configured to calculate, for each set of strength values, a weighted sum based on the respective set of strength values using a weighting function suppressing effect of largest strength values, and
- computer usable program code configured to select a window size based on said weighted sum.
A method for outlier detection for time series in database systems comprising:
- performing a local search for outliers on a sliding window with a window size;
- maintaining a data structure representing the degree to which a value of a measuring point can be an outlier;
- measuring an uncertainty in said data structure;
- optimizing window size by maximizing the uncertainty;
- and
- detecting outliers with a given threshold.
A computer program product for outlier detection for time series in database systems comprising a computer usable medium having computer usable program code embodied therewith, said computer usable program code comprising;
- computer usable program code configured to perform a local search for outliers on a sliding window with a window size;
- computer usable program code configured to maintain a data structure representing the degree to which a value of a measuring point can be an outlier;
- computer usable program code configured to measure an uncertainty in said data structure;
- computer usable program code configured to optimize window size by maximizing the uncertainty;
- and
- computer usable program code configured to detect outliers with a given threshold.
Dependencies are inferred from the claim text.