Explanation method, information output device, and program

Abstract

This document provides methods for explaining the output results of predictive models that utilize machine learning. [Solution] The method for explaining the process includes the steps of: calculating the contribution of multiple features used as input to the output result; grouping the multiple features into two or more categories such that similar features belong to the same group; creating a matrix data of three or more dimensions, each containing each of the grouped categories and the contribution in each dimension, wherein in each dimension corresponding to each category, multiple features are arranged in relative order of their calculated contributions; performing learning for a predetermined task using the matrix data; calculating the contribution to the predetermined task; and outputting the calculated contribution to the predetermined task.

Claims

1. A method for explaining the output results of a predictive model using machine learning, A step of calculating the contribution of each of the multiple features used as input to the prediction model to the output result, The steps include: grouping the aforementioned multiple features into two or more categories such that similar features belong to the same group; A step of creating a matrix data of three or more dimensions, each containing each of the two or more categories that have been grouped and the contributions calculated for the multiple features, wherein in each dimension corresponding to each of the two or more categories, the multiple features included in the group are arranged in a way that they are ordered from closest to closest based on the calculated contributions. The steps include: using two or more of the created matrix data to train a neural network for a predetermined task; A step of calculating the contribution of each element included in the matrix data to the predetermined task, The explanation of the output result includes the step of outputting the contribution of each element included in the matrix data calculated for the predetermined task, Explanation method.

2. The predetermined task belongs to a specific group within one of the two or more categories. The explanatory method described in claim 1.

3. The aforementioned forecasting model is a demand forecasting model for predicting the amount of future demand for a specific product based on past data related to that specific product. The explanatory method according to claim 1 or 2.

4. A program for causing a computer to execute the method described in claim 1 or 2.

5. An information output device for explaining the output results of a predictive model using machine learning, A calculation unit that calculates the contribution of each of the multiple features used as input to the prediction model to the output result, A group division unit that groups the aforementioned multiple features into groups such that similar features belong to the same group in each of two or more categories, A creation unit creates a matrix data of three or more dimensions, each containing each of the two or more categories that have been grouped and the contributions calculated for the multiple features, wherein in each dimension corresponding to each of the two or more categories, the multiple features included in the group are arranged in a matrix data such that they are arranged in relative order of relative proximity as their calculated contributions are closer. A learning unit that uses two or more of the created matrix data to perform learning for a predetermined task using a neural network, A contribution calculation unit that calculates the contribution of each element included in the matrix data to the predetermined task, The system includes an output unit that outputs the contribution of each element included in the matrix data calculated for the predetermined task, as an explanation of the output result. Information output device.

Images