Composite Loss Parameter for Imbalanced Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face challenges in achieving high accuracy and reliability due to class-imbalanced training datasets, particularly in detecting anomalies like rare diseases, where imbalanced data leads to inaccurate predictions and inefficient training processes.
Innovation Solution
The implementation of a combined anomaly detection machine learning model and classification prediction machine learning model using a composite loss parameter, which weights the normal prediction loss and global classification loss to improve predictive performance and reliability, especially in imbalanced datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a standalone anomaly detection machine learning model is used, then the model can detect anomalies, but the accuracy and reliability are low due to class-imbalanced training datasets
Solution Approach 1:
The patent combines an anomaly detection machine learning model with a classification prediction machine learning model into a hybrid system. The anomaly detection model generates anomaly scores for data objects, while the classification prediction model uses these scores along with feature data to predict classification labels. This merging of two separate modeling approaches allows the system to leverage both anomaly detection capabilities and classification accuracy, resolving the contradiction between reliability and precision in imbalanced datasets.
Solution Approach 2:
The patent transforms the output of the anomaly detection model (anomaly scores) into a useful input feature for the classification prediction model. By changing the parameter representation from raw anomaly scores to weighted contribution in the classification loss function, the system improves prediction accuracy while maintaining the reliability benefits of anomaly detection. This parameter transformation enables effective utilization of anomaly information in the context of classification tasks.
2Productivity
If traditional training methods are used with imbalanced datasets, then training is simpler, but the training process is inefficient and produces inaccurate predictions
Solution Approach 1:
The patent implements a feedback mechanism where the anomaly detection model's predictions (anomaly scores) are fed into the classification prediction model as additional features. This feedback loop allows the classification model to learn from the anomaly detection patterns while being trained on imbalanced data. The weighted loss function provides feedback during training to emphasize important samples, improving both training efficiency and prediction accuracy simultaneously.
Solution Approach 2:
The anomaly scores generated by the anomaly detection model serve as an intermediary between the raw data and the classification prediction. Instead of directly training the classification model on imbalanced data, the system uses anomaly scores as a mediator that captures important patterns. This intermediary representation enables more efficient training and improves prediction accuracy by providing the classification model with pre-processed, information-rich features.
3Ease of manufacture
If class-imbalanced training data is used, then the model can be trained on available data, but the predictions for rare anomalies become inaccurate
Solution Approach 1:
The patent applies preliminary action by first training the anomaly detection model on the available imbalanced data to learn anomaly patterns. This preliminary training step prepares the system by generating anomaly scores that capture rare patterns. Then, these scores are used as features in the classification prediction model, which is trained with a weighted loss function that gives appropriate emphasis to rare anomaly cases. This two-stage preliminary action enables effective utilization of imbalanced training data while improving reliability for rare anomaly detection.
Data Source
AI summary
Various embodiments of the present disclosure provide methods, apparatus, systems, computing devices, computing entities, and/or the like for improving machine learning model training based on receiving labeled training data objects, generating a normal prediction loss parameter, generating a global classification loss parameter, generating a composite loss parameter, and initiating the performance of one or more prediction-based operations.