Proxy Feature Mapping for Interpretable ML Network Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning-based network assurance systems lack explainability, making it difficult for administrators to determine appropriate corrective actions and leading to distrust due to the inability to understand why anomalies or failures are detected.
Innovation Solution
A service identifies probable cause attributes of a machine learning model's inference and generates a second model using these attributes as input features, providing interpretability by mapping input features to their causal counterparts, allowing for the creation of an interpretable model that aids in understanding the decision-making process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a machine learning model is used for network assurance, then the system can detect anomalies and predict failures, but the model lacks explainability making it difficult for administrators to understand why decisions are made
Solution Approach 1:
The patent introduces an intermediary service that sits between the black-box machine learning model and the network administrator. This service translates the model's internal feature representations into human-understandable proxy features by mapping them to known network attributes and relationships, thereby preserving the model's predictive accuracy while providing explainable insights into its decision-making process
Solution Approach 2:
The system creates a simplified copy or representation of the complex model's internal state through proxy features. Instead of presenting the administrator with the raw, incomprehensible feature vectors from the neural network, the system generates a simplified version that captures the essential reasoning in terms of familiar network concepts, making the model's logic transparent without sacrificing its predictive power
2Measurement precision
If complex machine learning models are deployed, then predictive accuracy improves, but the complexity of the system increases making it harder to interpret
Solution Approach 1:
The patent segments the complex prediction system into two distinct components: the black-box predictive model that maintains high accuracy, and the interpretability service that provides explanations. This segmentation allows each component to specialize - the model focuses on accurate prediction while the service focuses on generating understandable proxy features, thereby reducing the overall interpretability complexity without sacrificing predictive performance
3Ease of operation
If administrators need to understand model decisions, then interpretability is improved, but this may compromise the accuracy of the model
Solution Approach 1:
The interpretability service acts as an intermediary that does not replace the accurate black-box model but rather translates its outputs. By mapping internal features to proxy features representing known network relationships, the service provides interpretable explanations that reflect the model's actual reasoning, ensuring that ease of understanding does not compromise the model's predictive accuracy
Solution Approach 2:
The system substitutes the need for direct inspection of complex model internals with a different mechanism - the proxy feature mapping approach. Instead of requiring administrators to understand complex neural network mechanics, the system replaces this with a more accessible mechanism that expresses model decisions in terms of familiar network concepts and relationships, maintaining accuracy while improving interpretability
Data Source
AI summary
In one embodiment, a service identifies a set of attributes associated with a first machine learning model trained to make an inference about a computer network. The service obtains labels for each of the set of attributes, each label indicating whether its corresponding attribute is a probable cause of the inference. The service maps input features of the first machine learning model to those attributes in the set of attributes that were labeled as probable causes of the inference. The service generates a second machine learning model in part by using the mapped attributes to form a set of input features for the second machine learning model, whereby the input features of the first machine learning model and the input features of the second machine learning model differ.