Determination device, determination method, and determination program
A determination unit determines necessity of relearning of an initial learning model based on at least one of information regarding the initial learning model generated by learning initial learning data known to be normal, information regarding over-detection data over-detected by an abnormality detection system that uses the initial learning model, and information regarding an over-detection model generated based on the over-detection data. A notification unit notifies of determination result by determination unit.
1 . A determination device comprising:
processing circuitry configured to:
determine necessity of relearning of a first model based at least on:
information regarding the first model,
the first model has been generated by learning initial learning data that represent a ground-truth normal state,
information regarding over-detection data that describe the ground-truth normal state incorrectly as an abnormal state as over-detected by an abnormality detection system using the first model,
the information regarding the over-detection data further describes a ratio of a first number of types to a second number of types exceeding a predetermined value,
the over-detection data have been classified into the first number of types describing one or more types of a plurality of types according to a predetermined standard,
the initial learning data that have been classified into the second number of types describing one or more types of the plurality of types according to the predetermined standard,
information regarding a second model, and
the second model has been generated by learning the over-detection data; and
create and present a result of the determined necessity, thereby causing relearning of the first model using the information regarding the over-detection data as relearning data to classify the ground-truth normal state with accuracy by tracking change of the ground-truth normal state.
2 . The determination device according to claim 1 , wherein the processing circuitry is further configured to determine that the relearning of the first model is necessary when a ratio of the number of pieces of the over-detection data to the number of pieces of the initial learning data exceeds a predetermined value.
3 . The determination device according to claim 1 , wherein the processing circuitry is further configured to determine that the relearning of the first model is necessary when a loss function of the second model exceeds a predetermined value.
4 . The determination device according to claim 1 , wherein the processing circuitry is further configured to determine that the relearning of the first model is necessary when a ratio of data in which an abnormality is not detected by the abnormality detection system using the second model among detection target data exceeds a predetermined value.
5 . The determination device according to claim 1 , wherein the processing circuitry is further configured to determine that the relearning of the first model is necessary when a score indicating a degree of abnormality calculated by the first model exceeds a predetermined value.
6 . A determination method executed by a determination device, comprising:
determining necessity of relearning of a first model based at least on:
information regarding the first model,
the first model has been generated by learning initial learning data that represent a ground-truth normal state,
information regarding over-detection data that describe the ground-truth normal state incorrectly as an abnormal state as over-detected by an abnormality detection system using the first model,
the information regarding the over-detection data further describes a ratio of a first number of types to a second number of types exceeding a predetermined value,
the over-detection data have been classified into the first number of types describing one or more types of a plurality of types according to a predetermined standard,
the initial learning data that have been classified into the second number of types describing one or more types of the plurality of types according to the predetermined standard,
information regarding a second model, and
the second model has been generated by learning the over-detection data; and
creating and presenting a result of the determined necessity, thereby causing relearning of the first model using the information regarding the over-detection data as relearning data to classify the ground-truth normal state with accuracy by tracking change of the ground-truth normal state.
7 . A non-transitory computer-readable recording medium storing therein a determination program that causes a computer to execute a process comprising:
determining necessity of relearning of a first model based at least on;
information regarding the first model,
the first model has been generated by learning initial learning data that represent a ground-truth normal state,
information regarding over-detection data that describe the ground-truth normal state incorrectly as an abnormal state as over-detected by an abnormality detection system using the first model,
the information regarding the over-detection data further describes a ratio of a first number of types to a second number of types exceeding a predetermined value,
the over-detection data have been classified into the first number of types describing one or more types of a plurality of types according to a predetermined standard,
the initial learning data that have been classified into the second number of types describing one or more types of the plurality of types according to the predetermined standard,
information regarding a second model, and
the second model has been generated by learning the over-detection data; and
creating and presenting a result of the determined necessity, thereby causing relearning of the first model using the information regarding the over-detection data as relearning data to classify the ground-truth normal state with accuracy by tracking change of the ground-truth normal state.