Information processing method, information processing apparatus, and non-transitory computer-readable storage medium

An information processing apparatus, comprising:

• an obtaining unit that inputs each of plural sets of input data to be subjected to inference processing for violative content detection to a machine learning model, and obtains:
• plural output values output by the machine learning model and representing inference results respectively corresponding to the plural sets of input data;
• and reference information indicating a reference for evaluation of the machine learning model, wherein each output value has a value label that indicates evaluated result of corresponding input data;
• and
• a processing unit that selects an evaluated data group to be used in the evaluation of the machine learning model, from the plural output values, by using a threshold determined on the basis of the reference indicated by the reference information obtained by the obtaining unit, and calculates an index value representing an evaluation value of the machine learning model by using the evaluated data group selected and corresponding value labels,
• wherein plural output values are sorted in descending order of value, all output values equal to or larger than the threshold, among the plural output values that have been sorted, are set to a first value larger than the threshold, all output values less than the threshold, among the plural output values that have been sorted, are set to a second value smaller than the threshold, and the index value representing the evaluation value of the machine learning model is calculated by using the evaluated data group set to the first value,
• wherein the machine learning model is trained using the calculated index value as a feedback parameter to iteratively update model parameters, wherein the training improves violative content detectability of the machine learning model by dynamically optimizing detection thresholds based on the evaluation results.

A non-transitory computer-readable storage medium having stored therein an information processing program that causes a computer to execute:

• inputting each of plural sets of input data to be subjected to inference processing for violative content detection to a machine learning model, and obtaining:
• plural output values output by the machine learning model and representing inference results respectively corresponding to the plural sets of input data;
• and reference information indicating a reference for evaluation of the machine learning model, wherein each output value has a value label that indicates evaluated result of corresponding input data;
• and
• selecting an evaluated data group to be used in the evaluation of the machine learning model, from the plural output values, by using a threshold determined on the basis of the reference indicated by the obtained reference information, and calculating an index value representing an evaluation value of the machine learning model by using the evaluated data group selected and corresponding value labels,
• wherein the plural output values are sorted in descending order of value, all output values equal to or larger than the threshold, among the plural output values that have been sorted, are set to a first value larger than the threshold, all output values less than the threshold, among the plural output values that have been sorted, are set to a second value smaller than the threshold, and the index value representing the evaluation value of the machine learning model is calculated by using the evaluated data group set to the first value,
• wherein the machine learning model is trained using the calculated index value as a feedback parameter to iteratively update model parameters, wherein the training improves violative content detectability of the machine learning model by dynamically optimizing detection thresholds based on the evaluation results.

An information processing method executed by a computer, the information processing method comprising:

• performing, for each of a plurality of machine learning models, model inferencing for violative content detection using a plurality of input data sets to generate a corresponding plurality of inference results;
• receiving a plurality of value labels that corresponds to the plurality of input data sets, the plurality of value labels indicating actual evaluated results that correspond to the plurality of input data sets;
• performing, for each of a plurality of machine learning models:
• grouping ones of the plurality of inference results being equal to or greater than an inference threshold into a data group, and setting the ones of the plurality of inference results being equal to or greater than the inference threshold to a first value that is greater than the inference threshold;
• setting ones of the plurality of inference results being less than the inference threshold to a second value that is less than the inference threshold;
• and
• calculating an index value representing a machine learning evaluation value by using the data group and associated value labels from the plurality of value labels;
• and
• selecting a machine learning model with a highest index value from the plurality of machine learning models for deployment,
• wherein the selected machine learning model is trained using the associated index value as a feedback parameter to iteratively update model parameters, wherein the training improves violative content detectability of the selected machine learning model by dynamically optimizing detection thresholds based on the evaluation results.