Learning apparatus, trained model generation method, classification apparatus, classification method, and computer-readable recording medium

1 . A learning apparatus for machine learning a score function to be used for two-class classification, the learning apparatus comprising:

at least one processor; and

at least one memory,

wherein the at least one processor is configured to execute:

a score calculation unit configured to calculate a score for each piece of training data by inputting the training data to the score function; and

a score function optimization unit configured to:

identify a set of negative example training data that will be a false positive with respect to a specific fixed value of a false positive rate, based on the calculated scores; and

update a parameter of the score function through machine learning that is performed so as to maximize the index,

wherein the index is calculated by setting a weight for a pair of a positive example training data and a negative example training data to 1 when the negative example training data is within the identified set, and setting the weight for the pair to λ, which is a positive real number, when the negative example training data is not within the identified set,

wherein the update is performed while decreasing the value of the weight λ over a course of the machine learning,

wherein the index thereby represents a total value of a value obtained by multiplying an AUC by the weight λ and a value obtained by multiplying a pAUC by a weight (1−λ),

wherein the AUC indicates an area of a region on a horizontal axis side of a curve obtained by changing a threshold for determining a positive example and a negative example in the score function and setting a vertical axis as a true positive rate and a horizontal axis as a false positive rate, and

wherein the pAUC is a value of the AUC when the value of the false positive rate is set to a specific fixed value, and indicates an area of a region enclosed by the curve, the horizontal axis, and the vertical axis that passes through the fixed value.

2 . The learning apparatus according to claim 1 , further comprising:

a score calculation unit configured to calculate a score for each piece of the training data by inputting the training data to the score function,

wherein the score function optimization unit:

sets a weight for a pair with the training data of the positive example to 1, for the training data of the negative example which will be a false positive with respect to a false positive rate that has been set in advance, and sets a weight for a pair with the training data of the positive example to λ, for the training data of the negative example which will not be a false positive with respect to the false positive rate that has been set in advance; and

updates the parameter of the score function, while decreasing the value of the weight λ.

3 . A trained model generation method performed by a computer for machine learning a score function to be used for two-class classification, the trained model generation method comprising:

calculating a score for each piece of training data by inputting the training data to the score function;

identifying a set of negative example training data that will be a false positive with respect to a specific fixed value of a false positive rate, based on the calculated scores; and

updating a parameter of the score function through machine learning that is performed, so as to maximize the index,

wherein the index is calculated by setting a weight for a pair of a positive example training data and a negative example training data to 1 in a case where the negative example training data is within the identified set, and setting the weight for the pair to λ, which is a positive real number, in a case where the negative example training data is not within the identified set,

wherein the update is performed while decreasing the value of the weight λ over a course of the machine learning,

wherein the index thereby represents a total value of a value obtained by multiplying an AUC by the weight λ and a value obtained by multiplying a pAUC by a weight (1−λ),

wherein the AUC indicates an area of a region on a horizontal axis side of a curve obtained by changing a threshold for determining a positive example and a negative example in the score function and setting a vertical axis as a true positive rate and a horizontal axis as a false positive rate, and

wherein the pAUC is a value of the AUC when the value of the false positive rate is set to a specific fixed value, and indicates an area of a region enclosed by the curve, the horizontal axis, and the vertical axis that passes through the fixed value.

4 . The trained model generation method according to claim 3 , further comprising:

calculating a score for each piece of the training data by inputting the training data to the score function,

wherein the updating of the parameter of the score function comprises:

setting a weight for a pair with the training data of the positive example to 1, for the training data of the negative example which will be a false positive with respect to a false positive rate that has been set in advance, and setting a weight for a pair with the training data of the positive example to λ, for the training data of the negative example which will not be a false positive with respect to the false positive rate that has been set in advance; and

updating the parameter of the score function, while decreasing the value of the weight λ.

5 . A non-transitory computer-readable recording medium that includes a program, recorded thereon, for machine learning a score function to be used for two-class classification, the program including instructions that cause a computer to carry out:

calculating a score for each piece of training data by inputting the training data to the score function;

identifying a set of negative example training data that will be a false positive with respect to a specific fixed value of a false positive rate, based on the calculated scores; and

updating a parameter of the score function through machine learning that is performed so as to maximize the index,

wherein the index is calculated by setting a weight for a pair of a positive example training data and a negative example training data to 1 when the negative example training data is within the identified set, and setting the weight for the pair to λ, which is a positive real number, when the negative example training data is not within the identified set,

wherein the update is performed while decreasing the value of the weight λ over a course of the machine learning,

wherein the index thereby represents a total value of a value obtained by multiplying an AUC by the weight λ and a value obtained by multiplying a pAUC by a weight (1−λ),

wherein the AUC indicates an area of a region on a horizontal axis side of a curve obtained by changing a threshold for determining a positive example and a negative example in the score function and setting a vertical axis as a true positive rate and a horizontal axis as a false positive rate, and

wherein the pAUC is a value of the AUC in a case where the value of the false positive rate is set to a specific fixed value, and indicates an area of a region enclosed by the curve, the horizontal axis, and the vertical axis that passes through the fixed value.

6 . The non-transitory computer-readable recording medium according to claim 5 , wherein the program causes the computer to carry out:

calculating a score for each piece of the training data by inputting the training data to the score function; and

in the updating the parameter of the score function:

setting a weight for a pair with the training data of the positive example to 1, for the training data of the negative example which will be a false positive with respect to a false positive rate that has been set in advance, and setting a weight for a pair with the training data of the positive example to λ, for the training data of the negative example which will not be a false positive with respect to the false positive rate that has been set in advance; and

updating the parameter of the score function, while decreasing the value of the weight λ.