US 12,175,520 B2
System, method, and non-transitory machine-readable information storage medium for recommendations of items and controlling an associated bias thereof
Priyanka Gupta, Noida (IN); Pankaj Malhotra, Noida (IN); Ankit Sharma, Noida (IN); Gautam Shroff, Noida (IN); and Lovekesh Vig, Noida (IN)
Assigned to TATA CONSULTANCY SERVICES LIMITED, Mumbai (IN)
Filed by Tata Consultancy Services Limited, Mumbai (IN)
Filed on Jul. 20, 2022, as Appl. No. 17/813,741.
Claims priority of application No. 202121049708 (IN), filed on Oct. 29, 2021.
Prior Publication US 2023/0169569 A1, Jun. 1, 2023
Int. Cl. G06Q 30/00 (2023.01); G06Q 30/0601 (2023.01)
CPC G06Q 30/0631 (2013.01) 12 Claims
OG exemplary drawing
 
1. A processor implemented method, comprising:
receiving, via one or more hardware processors, a training session browsing history of a user from a user device, wherein the session browsing history comprises information on one or more items, and a session comprising a sequence of clicks on the one or more items; performing, via the one or more hardware processors, deconfounding training of a neural network (NN) model using (i) a plurality of causal graphs obtained based on domain knowledge, (ii) the training session browsing history, and (iii) a catalogue of items to obtain a trained NN model, wherein the step of performing deconfounding training of the neural network (NN) model comprises:
creating an embedding look up matrix for the one or more items in the catalogue of items;
normalizing the embedding look up matrix to obtain one or more normalized item embeddings;
modelling one or more session embeddings based on the sequence of clicks on the one or more items to obtain one or more modelled session embeddings;
dividing the one or more normalized item embeddings and the one or more modelled session embeddings into one or more corresponding groups;
normalizing each group from the one or more corresponding groups to obtain one or more normalized item embeddings groups and one or more normalized session embeddings groups;
performing a comparison of (i) each normalized item embeddings group, and (ii) each normalized session embedding group to determine a similarity therein;
obtaining a logit for each item from the one or more items based on the determined similarity;
applying a softmax function on the logit obtained for each item from the one or more items to obtain a relevance score for each item from the one or more items;
computing one or more cross entropy losses corresponding to one or more values of one or more weights of the neural network model based on the obtained relevance score for each item from the one or more items; and
training, via an optimizer, the neural network model using the one or more cross entropy losses and updating the one or more weights of the trained neural network model;
applying, via the one or more hardware processors, the trained NN model to a test session browsing history comprising information corresponding to a sequence of items, to obtain a causal inference derived from a test output associated therein;
identifying, via the one or more hardware processors, a total effect associated with one or more items on the test session browsing history based on the causal inference;
removing, via the one or more hardware processors, an indirect effect from the total effect;
upon removing the indirect effect, obtaining, via the one or more hardware processors, a logit for each item comprised in the catalogue of items (212);
applying, via the one or more hardware processors, a softmax function on the logit obtained for each item from the catalogue of items to obtain a relevance score for each item from the catalogue of items (214); and
recommending, via the one or more hardware processors, at least a subset of items from the catalogue of items based on the relevance score (216).