Explainable Dataset Similarity via Feature Pair Contributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data analysis methods, such as Earth Mover's Distance (EMD) and its adaptations, fail to provide transparent and explainable similarity results between datasets, particularly in AI systems, which is crucial for commercial deployments requiring interpretability.
Innovation Solution
A computer-implemented method that accesses pairs of feature weights from query and reference datasets, discovers flows to maximize similarity, computes pair contributions, and ranks contributive elements to explain the similarity results, using linear-complexity optimization and linear algebra primitives on graphics processing units, allowing for explainable classification and visualization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If EMD/WMD methods are used to compute similarity between datasets, then measurement precision of similarity is improved, but explainability of the decision is lost (black box problem)
Solution Approach 1:
The patent segments the overall similarity computation into individual pair contributions between features from the first dataset and features from the second dataset. Each pair contribution can be independently computed and explained, allowing the system to maintain measurement precision while providing interpretability through localized explanations of which feature pairs drive the similarity.
Solution Approach 2:
The patent introduces pair contributions as intermediary elements that mediate between the overall similarity metric and the underlying feature comparisons. These pair contributions serve as explainable intermediaries that capture the essence of why two datasets are similar without requiring the full complexity of the black box neural network to be interpreted.
2Measurement precision
If traditional EMD computation is used, then similarity measurement is achieved, but computational complexity becomes intractable for large datasets
Solution Approach 1:
The patent divides the computational task into segmenting datasets into features and computing pairwise contributions between them. This segmentation allows for more efficient computation by breaking down the intractable overall EMD computation into manageable pair-wise operations that can be parallelized and optimized.
Solution Approach 2:
The patent computes pair contributions for all feature pairs, which may be more than strictly necessary for a final similarity score. However, this partial/excessive computation of individual pair contributions enables both accurate overall similarity measurement and provides explanatory information about which specific feature pairs contribute most to the similarity, while still being computationally feasible through efficient algorithms.
3Loss of information
If explainable contributive elements are computed, then transparency of AI decisions is improved, but additional computation beyond basic EMD is required
Solution Approach 1:
The patent merges the computation of similarity measurement and explainability into a unified framework. The same pair contributions that are used to compute the overall similarity metric are also used to provide explanations, eliminating the need for separate explanation computations and thus maintaining productivity while achieving transparency.
Solution Approach 2:
The pair contributions serve dual purposes: they both compute the overall similarity and provide the explanatory information needed for transparency. The system essentially serves itself by using the same computational elements for both measurement and explanation, rather than requiring separate dedicated explanation mechanisms.
Data Source
AI summary
In an approach for construing similarities between datasets, a processor accesses a pair of sets of feature weights, wherein the sets of feature weights include a query dataset and comprises first weights associated to first features and a reference dataset and comprises second weights associated to second features. Based on similarities between the first features and the second features, a processor discovers flows from the first features to the second features, wherein the flows maximize an overall similarity between the pair of sets of feature weights. Based on the similarities and the flows, a processor computes pair contributions to the overall similarity in order to obtain contributive elements, wherein the pair contributions are contributions of pairs joining the first features to the second features. A processor ranks the contributive elements to obtain respective ranks. A processor returns a result comprising the contributive elements and indications to the respective ranks.