AI Model Reliability via Shapley Value Feature Influence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence systems lack reliable methods to evaluate the trustworthiness of their predictions, particularly due to issues like outliers and high sensitivity to specific features, which can lead to unreliable model outputs.
Innovation Solution
The proposed solution involves assessing the reliability of AI models by determining the influence of features on predictions using methods like Shapley values, outlier detection, and influence sensitivity plots, and mitigating issues through techniques such as feature influence restriction and L2 norm calculations to reduce overreliance on few features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If AI models are trained to make accurate predictions, then prediction accuracy is improved, but the models become highly sensitive to specific features and vulnerable to outliers, reducing reliability
Solution Approach 1:
The system computes Shapley values to measure feature influence and uses this feedback to identify and mitigate outliers. The reliability assessment feedback loop continuously evaluates prediction reliability and adjusts model behavior based on feature influence metrics, resolving the contradiction between accuracy and reliability.
Solution Approach 2:
The system changes the parameter of feature influence measurement by computing Shapley values for each feature and comparing them against threshold values. This parameter transformation allows the system to identify problematic features and adjust their influence, thereby improving reliability while maintaining accuracy.
2Speed
If the model relies heavily on a few key features to make predictions, then prediction speed is improved, but the model becomes overly sensitive to those features and less reliable
Solution Approach 1:
The system transforms the prediction process by introducing Shapley value computations that measure the actual influence of each feature. This parameter change allows the system to identify when a few features are exerting excessive influence and to mitigate their impact, balancing speed with reliability.
Solution Approach 2:
The system introduces Shapley values as an intermediary mechanism between features and predictions. These values act as mediators that quantify feature influence and enable the system to adjust the weight of individual features, preventing over-reliance on a few key features while maintaining prediction efficiency.
3Stability of the object's composition
If the model considers all features equally in predictions, then robustness is improved, but computational complexity and processing time increase
Solution Approach 1:
The system segments the feature space by computing Shapley values for each individual feature and categorizing them based on their influence threshold. This segmentation allows the model to treat high-influence and low-influence features differently, achieving robustness without considering all features equally, thereby reducing computational complexity.
Solution Approach 2:
The system applies local quality by assigning different weights or levels of consideration to different features based on their computed Shapley values. Features with influence above the threshold receive special attention and mitigation, while others are treated differently, creating a non-uniform but computationally efficient approach to robustness.
4Reliability
If the model is made more complex to handle outliers and feature sensitivity, then reliability is improved, but ease of operation and interpretability decrease
Solution Approach 1:
The system uses Shapley value computations as a feedback mechanism that provides clear, quantifiable information about feature influence. This feedback is presented in an interpretable format that allows users to understand which features are causing reliability issues and how the model is adjusting its predictions, maintaining ease of operation while improving reliability.
Solution Approach 2:
The system introduces Shapley values as an intermediary that bridges the gap between complex model internals and user interpretation. These values provide a clear, mathematically sound measure of feature influence that is both computationally rigorous and easily interpretable, allowing users to understand model behavior without sacrificing reliability.
Data Source
AI summary
Computer accesses training dataset with plurality of datapoints, each datapoint having input vector of feature values and output value. Training dataset is for training machine learning engine to predict the output value based on the input vector of feature values. The computer stores the training dataset as a two-dimensional vector with rows representing datapoints and columns representing features. The computer computes, for each feature value, a QII (quantitative input influence) value measuring a degree of influence that the feature exerts on the output value. For each datapoint from at least a subset of the plurality of datapoints, the computer (i) determines whether the QII value for each feature value in the input vector is within a predefined range, and (ii) upon determining that the QII value for a given feature value in the input vector is not within the predefined range: adjusts the training dataset or the machine learning engine.