ML Prediction Explanation Using Feature, Subgraph, and Node Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing interpretability technologies fail to provide comprehensive explanations for predictions made by two-stage machine-learning pipelines that utilize both data graph embeddings and other features, neglecting the interrelationships and interactions between these components and lacking clarity in identifying influential nodal features.
Innovation Solution
A method involving local search spaces for model, subgraph, and nodal features to determine a candidate explanation that satisfies an explanation rule, considering embeddings and other features, and transmitting the explanation to a user interface.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If existing interpretability technologies are used to explain predictions, then the explanation process is simple, but the explanations fail to account for interrelationships between embeddings and features, resulting in incomplete interpretability
Solution Approach 1:
The patent segments the explanation task into three distinct local search spaces: (1) model-feature explanation component that identifies important model features, (2) subgraph explanation component that identifies important subgraphs from the data graph, and (3) nodal-feature explanation component that identifies important nodal features. This segmentation allows each component to be optimized independently while collectively providing comprehensive explanations that account for interrelationships between embeddings and features.
Solution Approach 2:
The patent introduces a multi-dimensional explanation framework that operates across different levels of abstraction: the model-feature level (what features influence predictions), the subgraph level (which structural patterns are important), and the nodal-feature level (specific feature values at individual nodes). This dimensional expansion enables comprehensive interpretability by capturing interactions between embeddings and features at multiple granularities simultaneously.
2Loss of information
If comprehensive explanations accounting for all features and embeddings are provided, then interpretability is improved, but the computational complexity and time required for explanation increases significantly
Solution Approach 1:
The patent implements partial action by focusing explanation efforts on the most relevant portions of the model and data graph. The local search spaces are designed to identify only the subset of features, subgraphs, and nodal features that have the greatest influence on predictions, rather than analyzing all possible components. This selective approach maintains explanation completeness for the influential factors while significantly reducing computational overhead.
Solution Approach 2:
The patent applies local quality by creating specialized search spaces and explanation methods tailored to each component type: model features require different analysis than subgraph structures or nodal features. Each local search space is optimized for its specific data type and relationship patterns, enabling efficient computation of relevant explanations without the overhead of a monolithic explanation system.
3Measurement precision
If the explanation system analyzes all model features, subgraphs, and nodal features, then the accuracy of explanation is improved, but the device complexity and computational resources required increase
Solution Approach 1:
The patent segments the complex explanation task into three manageable local search spaces, each handling a specific aspect: model-feature explanations, subgraph explanations, and nodal-feature explanations. This segmentation reduces the complexity of individual components while maintaining overall explanation accuracy by ensuring each segment is optimized for its specific purpose.
Solution Approach 2:
The patent implements dynamic adaptation in the explanation system by allowing the navigation process to adaptively explore the local search spaces based on the specific prediction being explained. The system dynamically adjusts which features, subgraphs, and nodal features to examine based on their relevance to the current prediction, maintaining high explanation accuracy while avoiding unnecessary analysis of irrelevant components.
Data Source
AI summary
A computing platform may be configured to: (i) perform a plurality of explanation runs for explaining a prediction rendered by a machine learning model, where each explanation run involves (1) identifying (a) a first local search space for a model-feature explanation component, (b) a second local search space for a subgraph explanation component, and (c) a third local search space for a nodal-feature explanation component, (2) navigating the first, second, and third local search spaces to determine a respective candidate explanation that satisfies an explanation rule, and (3) determining a respective explanation score for the respective candidate explanation; and (ii) based on the respective explanation scores for the respective candidate explanations, determine a given explanation for the prediction including a given subset of a set of model features, a given subgraph of a data graph, and a given subset of a set of nodal features.