US 12,444,503 B2
Methods and systems for estimating causal effects from knowledge graphs
Rory McGrath, County Kildare (IE); Luca Costabello, Newbridge (IE); and Christophe Gueret, Dublin (IE)
Assigned to Accenture Global Solutions Limited, Dublin (IE)
Filed by Accenture Global Solutions Limited, Dublin (IE)
Filed on Apr. 28, 2022, as Appl. No. 17/731,590.
Claims priority of provisional application 63/188,689, filed on May 14, 2021.
Prior Publication US 2022/0367051 A1, Nov. 17, 2022
Int. Cl. G06N 20/00 (2019.01); G06N 5/02 (2023.01); G16H 50/20 (2018.01)
CPC G16H 50/20 (2018.01) [G06N 5/02 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computing device for estimating causal effects from knowledge graphs, the computing device comprising:
a network interface circuitry configured to obtain intervention application data and obtain subject history data stored in a database for a candidate group of subjects; and
an intervention circuitry configured to execute an intervention stack, wherein:
at a classification tier of the intervention stack, based on the intervention application data, the intervention circuitry is configured to divide the candidate group into a reception subgroup that received an intervention and a rejected subgroup that did not receive the intervention;
at a knowledge tier of the intervention stack, for each subject within the candidate group, based on the subject history data, the intervention circuitry is configured to map a covariate value set onto a knowledge graph with an embedding neural network, the covariate value set having a modeled functional relationship to selection for intervention;
at a matrix tier of the intervention stack, the intervention circuitry is configured to:
for each subject in the reception subgroup, translate the covariate value sets for the subjects within the reception subgroup into a reception matrix with a feature neural network,
for each subject in the rejection subgroup, translate the covariate value sets for the subjects within the rejection subgroup into a rejection matrix with the feature neural network,
compile reception matrices into an intervention matrix, and
compile rejection matrices into a non-intervention matrix; and
at a neural tier of the intervention stack, the intervention circuitry is configured to compare the reception subgroup to the rejection subgroup to determine a differential intervention effect by:
for each subject within the candidate group, applying the corresponding reception matrix or the corresponding rejection matrix to a likelihood neural network trained to determine an intervention application likelihood for the subject based on the covariate value set,
to obtain an intervention effect, applying the intervention matrix to an effect neural network trained to return an effect for subjects with same intervention application status,
to obtain a non-intervention effect, applying the non-intervention matrix to the effect neural network, and
subtracting the non-intervention effect from the intervention effect with accounting for the intervention application likelihood for the subjects, to obtain the differential intervention effect.