Model building server and model building method thereof

A model building method for a model building server, the model building server comprising a network interface, a storage and a processor, the storage storing a plurality of raw data, a plurality of output labels and a model building program, the output labels corresponding to the raw data, the model building program having a configuration combination, the configuration combination comprising a feature retrieving configuration, a model algorithm configuration and a model algorithm hyperparameter configuration, a feature configuration range of the feature retrieving configuration comprising a plurality of feature categories, each of the feature categories comprising a plurality of features, a model algorithm configuration range of the model algorithm configuration comprising a plurality of model algorithm categories, each of the model algorithm categories comprising at least one model algorithm, a hyperparameter configuration range of the model algorithm hyperparameter configuration comprising a plurality of hyperparameter sets, each of the hyperparameter sets corresponding to one of the model algorithms and comprising at least one hyperparameter, the feature retrieving configuration being configured to enable the model building program to generate a plurality of training data from the raw data and the model algorithm configuration and the model algorithm hyperparameter configuration being configured to enable the model building program to perform a training of supervised learning based on the training data and the output labels to build a model when the model building program is running, the processor being electrically connected to the network interface and the storage, the model building method being executed by the processor and comprising the following steps:

• performing a configuration combination coding based on the feature categories, the model algorithm categories and the hyperparameters to randomly generate a plurality of first configuration combination codes, wherein each of the first configuration combination codes has a plurality of first integer fields and a plurality of first float-point fields, the first integer fields correspond to the feature categories and the model algorithm categories respectively and the first float-point fields correspond to the hyperparameters respectively, an integer value of each of the first integer fields is randomly generated to be 1 or 0, the integer value of the first integer field being 1 represents that the feature category or the model algorithm category corresponding to the first integer field is applied, and the integer value of the first integer field being 0 represents that the feature category or the model algorithm category corresponding to the first integer field is not applied, and a float-point value of each of the first float-point fields is randomly generated and falls within a preset range of the corresponding hyperparameter;
• generating a plurality of first-stage optimization configuration combination codes from the first configuration combination codes based on a first optimization algorithm;
• setting the configuration combination of the model building program according to the first-stage optimization configuration combination codes respectively, and running the model building program to build a plurality of first-stage models;
• deciding a first model from the first-stage models, wherein the first model has a relative maximum accuracy value among the first-stage models, and the first-stage optimization configuration combination code corresponding to the first model indicates at least one selected feature category among the feature categories and at least one selected model algorithm category among the model algorithm categories;
• performing the configuration combination coding based on a plurality of selected features of the at least one selected feature category, at least one selected model algorithm of the at least one selected model algorithm category, and at least one selected hyperparameter corresponding to the at least one selected model algorithm to randomly generate a plurality of second configuration combination codes, wherein each of the second configuration combination codes has a plurality of second integer fields and at least one second float-point field, the second integer fields correspond to the selected features and the at least one selected model algorithm respectively, the at least one second float-point field corresponds to the at least one selected hyperparameter respectively, the integer value of each of the second integer fields is randomly generated to be 1 or 0, the integer value of the second integer field being 1 represents that the selected feature or the selected model algorithm corresponding to the second integer field is applied, and the integer value of the second integer field being 0 represents that the selected feature or the selected model algorithm corresponding to the second integer field is not applied, and the float-point value of each of the at least one second float-point field is randomly generated and falls within the preset range of the corresponding selected hyperparameter;
• generating a plurality of second-stage optimization configuration combination codes from the second configuration combination codes based on a second optimization algorithm;
• setting the configuration combination of the model building program according to the second-stage optimization configuration combination codes respectively, and running the model building program to build a plurality of second-stage models;
• deciding a second model from the second-stage models, wherein the second model has a relative maximum accuracy value among the second-stage models;
• and
• storing the second model into the storage as an optimization model.