Generalized Additive Model for Interpretable Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models for anomaly detection often require large amounts of labeled data and are not interpretable, limiting their effectiveness in distinguishing anomalies from normal data, especially in fields like healthcare and manufacturing.
Innovation Solution
A generalized additive model (GAM) trained using unsupervised and semi-supervised learning techniques with an anomaly detection partial identification (AD PID) loss function, incorporating differentiable neural decision trees to generate interpretable anomaly scores and explanations, allowing for accurate anomaly detection with limited labeled data and handling noisy or heterogeneous features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning is used to train anomaly detection models, then model accuracy can be improved, but large amounts of labeled data are required which limits effectiveness when labeled data is scarce
Solution Approach 1:
The patent introduces an intermediary approach by combining unsupervised learning (which does not require labeled data) with semi-supervised learning (which uses limited labeled data). The GAM model serves as a mediator that can be trained on unlabeled data to capture normal patterns, then refined with limited labeled anomaly examples, thus bridging the gap between completely unsupervised and fully supervised methods.
Solution Approach 2:
The patent changes the training parameters by using different loss functions for different training phases. During unsupervised training, it uses reconstruction-based loss to learn normal data distribution. During semi-supervised training, it switches to classification-based loss functions that leverage limited labeled data, thereby adapting the model's learning objective to match data availability.
2Reliability
If complex machine learning models are used for anomaly detection, then detection capability can be improved, but model interpretability deteriorates resulting in black-box models
Solution Approach 1:
The patent segments the model into interpretable components by using Generalized Additive Models that decompose predictions into individual feature contributions. Each feature's impact on the anomaly score can be separately analyzed and visualized, allowing users to understand which specific features drive anomaly detection decisions rather than treating the model as an opaque black box.
Solution Approach 2:
The patent introduces explanation mechanisms as intermediaries between the model's internal computations and human understanding. By adding feature importance scores, partial dependence plots, and other interpretability tools, the system mediates between complex model operations and human-comprehensible explanations, maintaining both accuracy and transparency.
3Measurement precision
If models are trained to detect anomalies, then anomaly detection performance can be improved, but handling noisy or heterogeneous features becomes difficult
Solution Approach 1:
The patent applies local quality by allowing different features to have different treatment and weighting within the GAM framework. Each feature can have its own smooth function that adapts to its specific characteristics, and feature importance can vary locally across different data regions. This allows the model to handle heterogeneous features with different noise levels and distributions appropriately.
Solution Approach 2:
The patent changes parameters dynamically to handle noisy features by using robust loss functions and adaptive smoothing parameters. The model can adjust the degree of smoothing for different features based on their noise characteristics, and use weighted loss functions that downweight potentially noisy or heterogeneous features during training.
Data Source
AI summary
Aspects of the disclosure provide for interpretable anomaly detection using a generalized additive model (GAM) trained using unsupervised and supervised learning techniques. A GAM is adapted to detect anomalies using an anomaly detection partial identification (AD PID) loss function for handling noisy or heterogeneous features in model input. A semi-supervised data interpretable anomaly detection (DIAD) system can generate more accurate results over models trained for anomaly detection using strictly unsupervised techniques. In addition, output from the DIAD system includes explanations, for example as graphs or plots, of relatively important input features that contribute to the model output by different factors, providing interpretable results from which the DIAD system can be improved upon.