Dataset Division for Domain Generalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information processing methods face challenges in generating datasets for domain generalization, particularly when there is limited data across multiple domains, leading to reduced model performance due to domain shift issues.
Innovation Solution
An information processing method that selects domain candidate variables to divide datasets into pseudo domains, determining whether each candidate dataset represents different domains, and generating datasets for each domain using these variables, allowing for the creation of multiple pseudo domains from a single domain dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If data is collected from a single domain for training, then data collection cost and time are reduced, but model prediction accuracy deteriorates due to domain shift when applied to different facilities
Solution Approach 1:
The patent creates pseudo-datasets by copying and transforming data from a single source domain. A generator network synthesizes artificial data in target domains that mimics the statistical properties and patterns of real data, enabling training without collecting actual data from multiple facilities. This copying approach maintains model reliability while avoiding the time and cost of multi-domain data collection.
Solution Approach 2:
The patent transforms data parameters by applying domain adaptation techniques that modify statistical properties of the source data. By changing parameters such as data distribution characteristics and feature representations through the generator, the system adapts single-domain data to represent multiple domains, resolving the contradiction between using limited data and achieving domain-generalization.
2Adaptability or versatility
If multiple domain datasets are collected for domain generalization learning, then model robustness against domain shift is improved, but data collection cost and complexity increase
Solution Approach 1:
Instead of physically collecting data from multiple domains, the system copies data from a single domain and uses a generator to create synthetic representations of other domains. This copying mechanism provides the diversity needed for domain generalization without the complexity of coordinating data collection across multiple facilities.
Solution Approach 2:
The generator network acts as an intermediary that bridges the gap between single-domain source data and multi-domain training requirements. This intermediary transforms the source data into multiple pseudo-domains, eliminating the need for direct data collection from multiple domains while still achieving the desired adaptability.
3Reliability
If the number of domains in training data is increased, then model performance across different domains is improved, but the difficulty of data preparation increases
Solution Approach 1:
The system copies data from a single verified domain and uses automated generation to create additional domain representations. This approach maintains data quality and consistency while achieving the necessary domain diversity, significantly reducing the manual preparation and verification effort required compared to collecting data from multiple actual domains.
Data Source
AI summary
Provided are an information processing method, an information processing apparatus, and a program that realize generation of a dataset of a user behavior history of different domains. An information processing method includes acquiring a dataset in one domain, the dataset being a dataset in which the response variable, the explanatory variable, and a plurality of variables excluding the response variable and the explanatory variable are applied, selecting a plurality of domain candidate variables that are domain candidates from the plurality of variables excluding the response variable and the explanatory variable, generating a dataset candidate that divides the dataset by using the domain candidate variables, and generating, in a case where each of the dataset candidates is a dataset in a different domain, a divided dataset by setting the domain candidate variables as domains.