US 11,954,590 B2
Artificial intelligence job recommendation neural network machine learning training based on embedding technologies and actual and synthetic job transition latent information
Haiyan Luo, San Jose, CA (US); Shichuan Ma, Dublin, CA (US); Anand Joseph Bernard Selvaraj, Fremont, CA (US); and Yu Sun, Austin, TX (US)
Assigned to Indeed, Inc., Austin, TX (US)
Filed by INDEED, INC., Austin, TX (US)
Filed on Jul. 31, 2020, as Appl. No. 16/944,972.
Claims priority of provisional application 62/882,340, filed on Aug. 2, 2019.
Prior Publication US 2021/0034975 A1, Feb. 4, 2021
Int. Cl. G06N 3/08 (2023.01); G06N 3/04 (2023.01); G06Q 10/105 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/04 (2013.01); G06Q 10/105 (2013.01)] 25 Claims
OG exemplary drawing
 
1. A method of artificial intelligence job recommendation generation with machine learning training based on embedding technologies and actual and synthetic job position related training data, the method comprising:
generating a collective directed graph G having vertices for each job position and edges for each job position-to-job position transition derived from actual job data scraped from multiple job seeker resumes stored by an online job search site;
determining probabilistic job position-to-job position transitions between vertices of the collective directed graph G to expose latent job position-to-job position transition trends;
generating synthetic training data for machine learning training using subsets of the vertices and edges of the collective directed graph G, wherein the synthetic training data includes:
positive synthetic training data generated by sampling the vertices and the edges of the collective directed graph G to collect pairs of the vertices connected by ones of the edges that follow a positive transition order;
first negative synthetic training data generated by randomly creating, from samples of the positive synthetic training data, new vertices for the collective directed graph G that are unreachable from the vertices and the edges that are based on the actual job data; and
second negative synthetic training data generated by performing a reverse random walk from each of multiple vertices of the collective directed graph G;
applying at least one machine learning process to the collective directed graph G to embed the collective directed graph G and the synthetic training data as vectors in a vector space that preserves asymmetric job position-to-job position transitions included in the collective directed graph G and the synthetic training data;
optimizing values of job position nodes in the vector space;
receiving job position-to-job position transition data for a job seeker using the online job search site;
classifying the job position-to-job position transition data for the job seeker with the optimized, job-to-job transition vector space to predict one or more job transitions for the job seeker; and
providing at least a subset of the predicted one or more job transitions to a job recommender engine of the online job search site to at least assist the job recommender engine in generating, within the online job search site, one or more job recommendations for the job seeker.