Material Prediction Data Sectioning for Accuracy and Unexpectedness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing prediction models for material development suffer from low prediction accuracy when data deviates significantly from training data, leading to a lack of unexpectedness and reduced efficiency in developing useful new materials.
Innovation Solution
A prediction device that determines sections based on attribute value frequency distributions, evaluates prediction target data appropriateness, and displays predicted values with an evaluation result, using a trained model to exclude data unlikely to yield useful new materials.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If prediction target data significantly deviates from training data, then unexpectedness increases, but prediction accuracy deteriorates
Solution Approach 1:
The patent segments the continuous attribute value range into multiple discrete sections based on frequency distribution. By dividing the data space into distinct sections, the system can evaluate prediction target data against multiple sections simultaneously, identifying which sections the target data falls into and comparing against corresponding training data sections to maintain accuracy while allowing deviation for unexpectedness.
Solution Approach 2:
The patent changes the parameter representation by transforming continuous attribute values into discrete section classifications. This parameter transformation allows the system to evaluate both the degree of deviation (for unexpectedness) and the correspondence with training data sections (for prediction accuracy) in a structured manner, resolving the contradiction between exploring new territory and maintaining reliability.
2Measurement precision
If prediction target data is close to training data, then prediction accuracy is maintained, but unexpectedness decreases
Solution Approach 1:
By segmenting the attribute value range into multiple sections, the system can identify when prediction target data falls into sections that are less frequently represented in training data. This allows the system to maintain prediction accuracy by still using the trained model while flagging data points that represent rarer scenarios, thereby preserving unexpectedness without sacrificing accuracy.
Solution Approach 2:
The patent implements feedback by evaluating which sections prediction target data falls into and comparing this against the frequency distribution of training data sections. This feedback mechanism allows the system to identify predictions that are both accurate (model can predict) and unexpected (rare in training data), thus resolving the contradiction between accuracy and unexpectedness.
3Productivity
If machine learning processes are reduced, then development efficiency improves, but prediction accuracy deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-calculating the frequency distribution of training data across multiple sections before actual prediction. This pre-computed section-based frequency information is stored and reused during prediction, eliminating the need for repeated complex machine learning processes while maintaining the ability to evaluate both accuracy and unexpectedness, thus improving efficiency without sacrificing precision.
Data Source
AI summary
To improve the development efficiency of a new material, a prediction device includes a section determination unit configured to acquire a training data set used for generating a trained prediction model, and determine a plurality of sections for classifying attribute values from a frequency distribution of the attribute values calculated between a plurality of data included in the training data set, an evaluation unit configured to determine sections to which attribute values calculated between prediction target data and the plurality of data are classified into, among the plurality of sections, and evaluate an appropriateness of the prediction target data with respect to conflicting indexes, and a display unit configured to display a predicted value predicted by the trained model in association with an evaluation result of the evaluation unit, by inputting the prediction target data to the trained model.