Discriminant Model Learning with Domain Knowledge Regularization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Supervised learning methods for discriminant model learning face challenges with small datasets, mismatched user knowledge, and inability to capture unseen phenomena, leading to reduced accuracy and reliability, especially when data nature changes.
Innovation Solution
A discriminant model learning device and method that incorporates domain knowledge to generate a regularization function, optimizing both the loss function and the regularization function to reflect user intentions and maintain data fitting, using a query candidate storage and model learning process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning methods are used to learn discriminant models from data with discriminant labels, then the model can be optimized to minimize discrimination error, but the model may not match user's knowledge when data amount is small
Solution Approach 1:
The patent introduces domain knowledge as an intermediary element that mediates between the data-driven learning process and user expectations. The domain knowledge module provides constraints or prior information that guides the learning process, ensuring the learned model aligns with user knowledge while still utilizing available data for optimization.
Solution Approach 2:
The patent applies preliminary action by incorporating domain knowledge before the actual learning process. The domain knowledge module prepares prior information, constraints, or guiding principles in advance, which are then integrated into the learning algorithm to ensure the model starts from a foundation that aligns with user expectations, rather than learning purely from data without such guidance.
2Measurement precision
If a discriminant model is optimized using loss functions to minimize discrimination error, then the model fits the training data well, but the model cannot capture phenomena not present in the training data
Solution Approach 1:
The patent changes the parameters of the learning process by incorporating domain knowledge constraints or prior information into the optimization objective. This modifies the effective parameter space and optimization criteria, enabling the model to capture patterns and phenomena that extend beyond the training data distribution while maintaining good fit on the training data itself.
Solution Approach 2:
The patent performs preliminary action by pre-incorporating domain knowledge about potential phenomena, relationships, or patterns into the learning framework before training. This allows the model to be prepared to recognize and capture unseen phenomena that align with domain expectations, rather than being limited solely to patterns present in the training data.
3Quantity of substance
If semi-supervised learning methods are used to utilize data without discriminant labels, then the amount of usable data increases, but the model still cannot guarantee matching user knowledge when data is limited
Solution Approach 1:
The patent uses domain knowledge as an intermediary that bridges the gap between the increased quantity of usable data from semi-supervised learning and the requirement to match user knowledge. The domain knowledge module processes and integrates this additional data while maintaining alignment with user expectations through the constraints or prior information provided by the domain knowledge.
Solution Approach 2:
The patent changes the learning parameters by combining semi-supervised learning objectives with domain knowledge constraints. This allows the model to utilize the increased quantity of data from semi-supervised approaches while the parameter modifications ensure the learned relationships remain consistent with user knowledge, even when data is limited.
Data Source
AI summary
To provide a discriminant model learning device capable of efficiently learning a discriminant model on which domain knowledge indicating user's knowledge or analysis intention for a model is reflected while keeping fitting to data. A query candidate storage means 81 stores candidates of a query as a model to be given with domain knowledge indicating a user's intention. A regularization function generation means 82 generates a regularization function indicating compatibility with domain knowledge based on the domain knowledge to be given to the query candidates. A model learning means 83 learns a discriminant model by optimizing a function defined by a loss function and the regularization function predefined per discriminant model.