Outlier determination rule generation device and outlier detection device, and outlier determination rule generation method and outlier detection method thereof

A computer implemented outlier detection device for detecting abnormal data in a data set, said outlier detection device comprising:

• an outlier rule preservation unit for holding a set of rules characterizing abnormal data, said abnormal data being indicative of fraudulent data,
• a filtering unit for determining whether each data of said data set is abnormal data or not based on said held rules,
• a degree of outlier calculation unit for calculating a degree of outlier indicative of a degree of being abnormal with respect to each data determined not to be abnormal data by said filtering unit, said each data determined not to be abnormal having assigned a score based on said calculated degree of outlier,
• a sampling unit for sampling a portion of higher scored data from said scored data and labeling said higher scored data as positive for abnormal data, and sampling a portion of lowest scored data and labeling said lowest scored data as negative for abnormal data, said positive and negative labeling being supervisor information, and
• a supervised learning unit for generating a new rule characterizing abnormal data by supervised learning using both of said positive labeled data and said negative labeled data as said supervisor information and adding the new rule to said held set of rules to update the rules.

A computer implemented outlier rule generation method for characterizing abnormal data in a data set, said method comprising:

• calculating a degree of outlier for each data of said data set, said degree of outlier being indicative of a degree of being abnormal, said each data having assigned a score based on said calculated degree of outlier, said abnormal data being indicative of fraudulent data,
• sampling a portion of higher scored data from said scored data,
• labeling said higher scored data as positive for abnormal data, and sampling a portion of lowest scored data and labeling said lowest scored data as negative for abnormal data, said positive and negative labeling being supervisor information, and
• generating a rule characterizing abnormal data by supervised learning using both of said positive labeled data and said negative labeled data as said supervisor information.

A computer implemented outlier rule generation method for characterizing abnormal data in a data set, said method comprising:

• an outlier rule preservation step of holding said set of rules characterizing abnormal data, said abnormal data being indicative of fraudulent data,
• a filtering step of determining whether each data of said data set is abnormal data or not based on said held rules,
• a degree of outlier calculation step of calculating a degree of outlier indicative of a degree of being abnormal with respect to each data determined not to be abnormal data by said filtering step, said each data determined not to be abnormal having assigned a score based on said calculated degree of outlier,
• a sampling step of sampling a portion of higher scored data of said scored data and labeling said higher scored data as positive for abnormal data, and sampling a portion of lowest scored data and labeling said lowest scored data as negative for abnormal data, said positive and negative labeling being supervisor information, and
• a supervised learning step of generating a new rule characterizing abnormal data by supervised learning using both of said positive labeled data and said negative labeled data as said supervisor information and adding the new rule to said held set of rules to update the rules.

An outlier rule generation program embodied in a computer-readable media for generating a rule characterizing abnormal data in a data set for controlling a computer, said program comprising:

• an outlier rule preservation function of holding said set of rules characterizing abnormal data, said abnormal data being indicative of fraudulent data,
• a filtering function for determining whether each data of said data set is abnormal data or not based on said held rules,
• a degree of outlier calculating function for calculating a degree of outlier indicative of a degree of being abnormal with respect to each data determined not to be abnormal data by said filtering function, said each data determined not to be abnormal having assigned a score based on said calculated degree of outlier,
• a sampling function for sampling a portion of higher scored data of said scored data and labeling said higher scored data as positive for abnormal data, and sampling a portion of lowest scored data and labeling said lowest scored data as negative for abnormal data, said positive and negative labeling being supervisor information, and
• a supervised learning function for generating a new rule characterizing abnormal data by supervised learning using both of said positive labeled data and said negative labeled data as said supervisor information and adding the new rule to said held set of rules to update the rules.