US 12,217,143 B1
Resolving opaqueness of complex machine learning applications
Vijayan N. Nair, Matthews, NC (US); Agus Sudjianto, Charlotte, NC (US); Jie Chen, Fremont, CA (US); Kurt Schieding, Charlotte, NC (US); Linwei Hu, Charlotte, NC (US); Xiaoyu Liu, Newark, CA (US); and Joel Vaughan, Charlotte, NC (US)
Assigned to Wells Fargo Bank, N.A., San Francisco, CA (US)
Filed by Wells Fargo Bank, N.A., San Francisco, CA (US)
Filed on Mar. 5, 2019, as Appl. No. 16/293,252.
Application 16/293,252 is a continuation in part of application No. 16/179,073, filed on Nov. 2, 2018, granted, now 11,550,970.
Int. Cl. G06N 20/20 (2019.01); G06F 16/901 (2019.01); G06F 17/13 (2006.01); G06F 17/15 (2006.01); G06F 18/23213 (2023.01); G06N 3/047 (2023.01); G06N 3/048 (2023.01); G06T 11/20 (2006.01)
CPC G06N 20/20 (2019.01) [G06F 16/9027 (2019.01); G06F 17/13 (2013.01); G06F 17/15 (2013.01); G06F 18/23213 (2023.01); G06N 3/047 (2023.01); G06N 3/048 (2023.01); G06T 11/206 (2013.01)] 19 Claims
OG exemplary drawing
 
1. A computing system comprising:
a replicator that employs Locally Interpretable Models and Effects based on Supervised Partitioning (LIME-SUP) techniques on a plurality of machine learning-models, comprising at least one machine learning model with opacity issues that include a black box effect, to create a replicated semi-additive index data structure that provides local and global explainability for the black box effect,
wherein the replicator communicates with a library of interpretable machine learning models, and wherein the created replicated semi-additive index data structure is a neural network based formulation of an additive index model where a neural network of the neural network based formulation is a replicated structured neural network that is trained using mini-batch gradient-based methods, and wherein the black box effect is a non-visible operation, of the at least one machine learning model with the opacity issues, that lacks explainability;
a translator module that generates a score and a plurality of sub-scores based on the created replicated semi-additive index data structure, wherein generating the score involves ranking a relative effect of a disparity to a plurality of elements of the created replicated semi-additive index data structure, wherein the ranking comprises ranking relative importance of a decomposed effect of the disparity, wherein the score is indicative of a match status to the at least one machine learning model with the opacity issues; and
a graphical user interface that renders selected characteristics of the created replicated semi-additive index data structure to produce a rendering that depicts the local and global explainability for the black box effect of the at least one machine learning model with the opacity issues.