Teacher-Student Model Interpreter for Siamese Network Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Interpreting the behavior of complex Siamese Networks used for predicting correlations between data is inefficient, making it difficult to identify important features for improving prediction accuracy and aligning with human knowledge.
Innovation Solution
A teacher-student model is employed, where a Siamese Network acts as the teacher model and a simpler interpreter model, such as Random Forest, generates global and local feature importance scores to interpret the teacher model's behavior, focusing on specific features for retraining and interactive review of correlations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a Siamese Network is used to predict correlations between data, then prediction accuracy is improved, but interpretability of the model behavior deteriorates
Solution Approach 1:
The patent introduces an interpreter model as an intermediary component that sits between the Siamese Network and the user/analyst. This interpreter model generates feature importance scores and explanations for the correlation predictions, making the black-box Siamese Network interpretable without modifying its predictive functionality. The interpreter acts as a mediator that translates complex model outputs into human-understandable feature importance metrics.
Solution Approach 2:
The patent segments the interpretation task into two distinct models: the Siamese Network for prediction and the interpreter model for explanation. This segmentation allows each component to specialize - the Siamese Network focuses on accurate correlation prediction while the interpreter model focuses on generating interpretable feature importance scores, thereby resolving the contradiction between accuracy and interpretability.
2Reliability
If a Siamese Network is used for correlation prediction, then prediction capability is improved, but ease of identifying important features deteriorates
Solution Approach 1:
The interpreter model serves as an intermediary that automatically extracts and ranks important features from the Siamese Network's predictions. Instead of requiring users to manually analyze complex model internals, the interpreter generates feature importance scores that directly indicate which features drive the correlation predictions, thereby easing feature identification while maintaining prediction reliability.
Solution Approach 2:
The system implements feedback by using the interpreter model to continuously analyze and provide feature importance information back to users. This feedback loop enables users to understand which features are most influential in each prediction, making it easy to identify important features without compromising the Siamese Network's prediction capability.
3Productivity
If global feature importance scores are generated using an interpreter model, then feature identification efficiency is improved, but model complexity increases
Solution Approach 1:
The patent uses a simpler Random Forest model as a copy or approximation of the Siamese Network's decision-making process. The Random Forest interpreter model is trained to replicate the behavior of the Siamese Network and generate feature importance scores. This copying approach enables efficient feature identification using a simpler model structure, thereby improving productivity while managing system complexity through the use of well-established, computationally efficient algorithms.
Data Source
AI summary
Systems and methods are provided for interpreting a correlation model that predicts a correlation between a pair of data corresponding to a pair of incident tickets using an interpreter model. The correlation model includes a Siamese Network including a plurality of neural networks. The interpreter model, trained by using training data, represents a student model (a glass-box model) while the correlation model, trained using the training data, represents a more complex teacher model (a black-box mode) of a teacher-student model. The present disclosure generates global feature importance scores based on the trained interpreter model, which indicates a degree of influence of a feature compared to other features in incident data in determining correlations, to generate additional training data emphasizing influential features and to retrain the correlation model. The present disclosure further determines local feature importance scores based on the trained interpreter model for confirming an accuracy of predicting correlations.