Neural Network for Accurate SHAP Value Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models fail to provide accurate explanations for computer alerts due to the assumption of feature independence in SHAP values, leading to high computational complexity and fidelity issues in practical applications like cyber-security and fraud prevention.
Innovation Solution
A novel artificial neural network architecture is trained to compute conditional expectations, allowing for accurate predictions and SHAP value generation without assuming feature independence, using binary masks to handle missing data and reduce computational overhead during model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If sampling-based SHAP approach is used to compute conditional expectations, then model agnosticism is maintained, but computational complexity increases exponentially and processing power requirements become exorbitant
Solution Approach 1:
The patent pre-computes and stores feature marginals during the training phase, transforming the complex sampling-based conditional expectation computation into simple lookups during inference. This preliminary action eliminates the need for expensive sampling at prediction time while maintaining model-agnostic SHAP value computation.
Solution Approach 2:
The patent creates a copy of the training data distribution in the form of pre-computed feature marginals. Instead of performing expensive sampling operations during inference, the system uses these stored copies to efficiently compute conditional expectations, dramatically reducing computational complexity while preserving the sampling-based SHAP approach.
2Device complexity
If feature independence assumption is made in SHAP values, then computational complexity is reduced, but fidelity of explanations deteriorates and results become inaccurate in real datasets
Solution Approach 1:
The patent pre-computes feature marginals during training, enabling efficient computation of conditional expectations without assuming feature independence. This preliminary computation stores the necessary statistical information to accurately compute SHAP values for dependent features, maintaining explanation fidelity while avoiding the exponential computational complexity of exact methods.
Solution Approach 2:
The patent changes the computational parameters by pre-computing and storing feature marginals during training. This parameter transformation allows the system to efficiently compute conditional expectations for dependent features during inference, achieving both low computational complexity and high explanation fidelity without making the feature independence assumption.
3Loss of information
If conventional SHAP systems are used to provide visibility into alert causes, then interpretability is improved, but the assumption of feature independence leads to inaccurate recommendations in practical applications
Solution Approach 1:
The patent pre-computes feature marginals during the training phase, storing the statistical distribution of each feature. This preliminary action enables the system to accurately compute conditional expectations for dependent features during inference, providing both visibility into alert causes and accurate recommendations without assuming feature independence.
Solution Approach 2:
The patent creates a copy of the training data distribution in the form of pre-computed feature marginals. This copy allows the system to efficiently and accurately compute SHAP values for features with dependencies, maintaining both interpretability and accuracy in practical applications like cyber-security and fraud prevention.
Data Source
AI summary
Methods and systems are disclosed for generating recommendations for causes of computer alerts that are automatically detected by a machine learning algorithm are described. For example, the system may receive a first feature vector with an unknown alert status, wherein the first feature vector represents values corresponding to a plurality of computer states in a first computer system. The system may input the first feature vector into an artificial neural network, wherein the artificial neural network is trained to detect a known alert status based on a set of training data comprising labeled feature vectors corresponding to the known alert status, and wherein the artificial neural network is trained to detect conditional expectations the plurality of computer states in an inputted feature vector.