Machine Learning Model Training Rarity Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional prediction models struggle to accurately predict rare events that occur rapidly, as they are overshadowed by continuous events in learning data, leading to poor representation and prediction accuracy for rare events.
Innovation Solution
A model generation apparatus that adjusts the learning process by setting a rarity degree for each dataset, prioritizing more extensive training on datasets with higher rarity, ensuring that rare event data is adequately reflected in the prediction model, using techniques like increased sampling frequency or weighting to emphasize rare event data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning is executed using obtained learning data, then learning data relating to continuous events is processed, but learning data relating to rare events gets buried and prediction accuracy for rare events deteriorates
Solution Approach 1:
The patent applies local quality by assigning different weights to different data points based on their rarity. Rare events are assigned higher weights while continuous events receive lower weights, allowing the model to focus computational attention on critical rare events without completely ignoring the abundant continuous event data. This differential weighting approach ensures that rare events are adequately represented in the learning process despite their small proportion in the overall dataset.
Solution Approach 2:
The patent changes the parameter of data weighting in the machine learning process. By introducing a rarity-based weighting parameter that dynamically adjusts the importance of each learning dataset based on its rarity degree, the system transforms the uniform treatment of all data into a differentiated approach where rare events receive amplified influence on model training, thereby improving prediction accuracy for rare events.
2Reliability
If conventional machine learning processes all learning data uniformly, then processing is simple, but rare events cannot be appropriately reflected in the prediction model
Solution Approach 1:
The patent introduces a rarity degree parameter that quantifies how rare each event type is in the learning data. This parameter is then used to compute weights for each learning dataset, transforming the uniform processing approach into a differentiated weighting scheme. The complexity increase is minimal, involving only the calculation and application of these weights during the machine learning process, while yielding significant improvements in rare event prediction accuracy.
Solution Approach 2:
The patent implements a feedback mechanism where the rarity degree of events is calculated based on the learned distribution from continuous events, and this rarity information is fed back into the weighting scheme for training the prediction model. This feedback loop ensures that the model adapts to the actual rarity patterns in the data, continuously improving its ability to predict rare events while maintaining a relatively simple overall process structure.
Data Source
AI summary
A model generation apparatus according to one aspect of the present invention acquires a plurality of learning datasets each constituted by a first sample of a first time of predetermined data obtained in time series and feature information included in a second sample of the predetermined data of a future second time relative to the first time, and trains a prediction model, by machine learning, to predict feature information of the second time from the first sample of the first time, for each learning dataset. In the model generation apparatus, a rarity degree for is set each learning dataset, and, in the machine learning, the model generation apparatus trains more preponderantly on learning datasets having a higher rarity degree.