High-Dimensional Rare-Event Detection With Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing modeling techniques, such as support vector machines and decision trees, are poorly suited for predicting extremely rare events like subscriber churn, as they suffer from class imbalance and overfitting, especially when event rates drop below 1%, leading to inaccurate predictions.
Innovation Solution
A method using neural networks and deep learning to identify key performance indicators, construct a linear model, and apply OOT validation datasets to improve prediction accuracy and reduce overfitting, while integrating social media data for enhanced insights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional modeling techniques (SVM, decision trees) are used for predicting extremely rare events, then the model structure is simple and easy to implement, but the prediction accuracy deteriorates due to class imbalance and overfitting when event rates drop below 1%
Solution Approach 1:
The patent segments the prediction process into multiple stages: first using a neural network to identify and segment the data into training and validation sets, then using a linear model for the actual prediction. This segmentation allows the complex neural network to handle class imbalance during data preparation while the simpler linear model performs the prediction, resolving the contradiction between accuracy and complexity.
Solution Approach 2:
The patent introduces an intermediary validation dataset as a mediator between the training data and the final prediction. This validation set acts as a buffer that allows the model to be trained on imbalanced data while being validated on balanced data, preventing overfitting and improving prediction accuracy without requiring the entire model to be overly complex.
2Reliability
If the event rate is reduced by up to 20 times to maintain class balance, then the prediction reliability improves, but the quantity of data available for modeling decreases
Solution Approach 1:
The patent applies preliminary action by creating the validation dataset before the actual prediction process. The neural network is trained on the full dataset including all rare events, and the validation set is prepared in advance to represent the balanced distribution. This allows the model to learn from the complete data quantity while being validated on a balanced subset, resolving the contradiction between reliability and data quantity.
3Measurement precision
If neural networks are used to select key performance indicators, then the measurement precision for identifying important features improves, but the computational time and complexity increase
Solution Approach 1:
The patent applies partial action by using the neural network only for the specific task of selecting key performance indicators rather than training it on the entire prediction problem. The neural network processes the data to identify important features, and then a simpler linear model is used for the actual prediction. This partial application of the complex neural network approach achieves high feature selection precision while limiting the computational time to only the necessary feature identification phase.
Data Source
AI summary
Methods and systems for using machine learning to identify extremely rare events in high-dimensional space are disclosed. A method includes: identifying, by a computing device, a plurality of derived attributes using an external data source; selecting, by the computing device, a plurality of key performance indicators from the plurality of derived attributes using a neural network and based on an extremely rare event being modeled; constructing, by the computing device, a linear model using the plurality of key performance indicators; and predicting, by the computing device, occurrences of the extremely rare event using the linear model.