| CPC G06N 3/082 (2013.01) [G06F 17/18 (2013.01); G06F 18/217 (2023.01); G06F 18/285 (2023.01); G06N 3/08 (2013.01); G06N 20/20 (2019.01)] | 12 Claims |
|
1. An apparatus for generating an integrated model from a plurality of individual prediction models, the apparatus comprising:
memory storing a set of computer program instructions; and
at least one computer configured to perform, in accordance with the set of computer program instructions, an integrated model determination that includes:
a selection task of selecting, from the plurality of individual prediction models, a subset of plural individual prediction models as a combination of candidate models, the candidate models constituting a candidate integrated model;
a training task of training the candidate models using a training dataset;
an evaluation metric calculation task of applying a test dataset to the candidate integrated model to calculate an evaluation metric for evaluating an output of the candidate integrated model;
a regression expression generating task of generating a regression expression that represents a relationship between the evaluation metric and the candidate integrated model in accordance with an expression (1) as follows:
Z=Σi,jaijsisj+Σibisi (1)
where:
Z represents the evaluation metric;
each of si and sj is a selection indicator for each of the candidate models representing whether a corresponding one of the candidate models is included in the candidate integrated model;
aij represents a weight parameter for the selection indicator si and a weight parameter for the selection indicator sj; and
bi represents the weight parameter for the selection indicator si;
a repetition task of repeatedly performing a sequence of the selection task, the training task, the evaluation metric calculation task, and the regression expression generating task until a predetermined termination condition is satisfied, the selection task being configured to use an expression (2) as follows to accordingly select, from the plurality of individual prediction models, a new subset of new candidate models that constitute a new candidate integrated model as a new candidate for each sequence:
argminSc∈{0,1}D(Σi,jaijsisj+Σibisi+λΣi∥si∥1) (2)
where:
argmin(x) stands for argument of a minimum of a function x,
λ represents a hyper parameter,
Σi∥si∥1 represents a L1 norm of si,
Sc represents the combination, and
D represents the number of the plurality of individual prediction models, and
wherein:
the term λΣi∥si∥1 introduced in the expression (2) enables the selection task to search the plurality of individual prediction models for the new subset of new candidate models whose number is minimized, and
the regression expression generated with an additional value of the evaluation metric from the new candidate combination of candidate models for each sequence has a higher prediction accuracy on average as the number of the repeated sequences increases; and
a determination task of determining an integrated model based on the candidate integrated models and values of the evaluation metric after the termination condition is satisfied.
|