Rare-Event Training Data Augmentation for Balanced ML Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning algorithms for predicting rare events are less efficient due to unbalanced training data, where data associated with the absence of a rare event vastly outnumber data associated with its presence, leading to poor performance in detecting rare occurrences in industrial systems and medical diagnostics.
Innovation Solution
A method and device that generate balanced training data by identifying subsets of variables associated with a risk threshold, applying pseudo-random variations, and grouping data into three categories using unsupervised learning, followed by reclassification with multiple machine learning models to rebalance the dataset and improve prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If training data are used in bulk for common events, then the dataset size is large, but the data distribution becomes unbalanced and rare events are underrepresented
Solution Approach 1:
The patent generates synthetic copies of rare event data through pseudo-random variations. Step B creates a third set of data by applying pseudo-random variations to the subset of data from rare events, effectively copying and transforming limited rare event samples to increase their representation in the training dataset without simply duplicating original data.
Solution Approach 2:
The patent transforms the distribution of training data by applying parameter changes through pseudo-random variations. Step B modifies the parameters of the subset data (from rare events) by introducing controlled random variations, thereby changing the data distribution to better represent rare events while maintaining their essential characteristics.
2Quantity of substance
If the cardinal of the first set (absence of rare event) is much larger than the second set (presence of rare event), then common events are well-represented, but the classification algorithm becomes less efficient for predicting rare events
Solution Approach 1:
The patent applies local quality by treating different data sets differently. Step A identifies a specific subset of variables from the rare event data that are most relevant. Step B then applies pseudo-random variations specifically to this subset, creating localized improvements in the representation of rare events without affecting the overall dataset structure or common event data quality.
Solution Approach 2:
The patent performs preliminary action by pre-processing the rare event data before main training. Steps A and B identify and enhance rare event characteristics in advance by selecting relevant variables and generating varied samples, so that when the augmented dataset is used for training, the algorithm is already prepared to detect rare events more effectively.
3Productivity
If unbalanced training data are used, then the first set (common events) dominates the dataset, but the algorithm performance for rare event prediction deteriorates
Solution Approach 1:
The patent merges multiple data sets to create a balanced augmented training dataset. Step C groups the first set (common events), second set (original rare events), and third set (synthetically generated rare events) into a unified augmented dataset, combining the benefits of large common event samples with enhanced rare event representation.
Solution Approach 2:
The patent introduces an intermediary process to bridge the gap between unbalanced original data and balanced training data. The pseudo-random variation process acts as an intermediary transformation that converts limited rare event samples into a larger set of varied rare event samples, mediating the imbalance without requiring additional real-world rare event data collection.
Data Source
AI summary
The invention relates to a method and a device for generating training data for machine learning of a model for predicting a risk of occurrence of a rare event, said training data being formed by vectors of variables representative of a state of at least one monitored system before the occurrence of said rare event, from input training data (4) including a first initial set (6) of data associated with an absence of occurrence of said rare event, and a second initial set (8) of data associated with a presence of occurrence of said rare event. The device implements modules for determining (20), by an automatic learning method, at least one subset of variables associated with a risk of occurrence of said rare event higher than a risk threshold, generating (22) a third set of data by pseudo-random variations on said subset of data to provide an augmented database of training data.