| CPC G06N 20/00 (2019.01) [G06F 17/18 (2013.01); G06F 18/211 (2023.01); G06F 18/241 (2023.01); G06F 18/2431 (2023.01)] | 20 Claims |
|
1. A computer-implemented method of training a machine learning model, the computer-implemented method comprising:
electronically retrieving a plurality of feature vectors from a data set having a predetermined data imbalance, each feature vector comprising an electronic representation of a corresponding one of a plurality of multidimensional observations, each multidimensional observation uniquely associated with an observation value;
generating a multi-class data structure comprising a plurality of buckets by binning the observation values associated with the multidimensional observations, each bucket corresponding to a distinct range of the observation values and containing one or more of the plurality of multidimensional observations whose associated observation values lie within the distinct range;
using the plurality of feature vectors, training the machine learning model to classify feature vector inputs by assigning each feature vector input to one of the plurality of buckets based on feature values of each feature vector input, wherein the machine learning model is a deep neural network having one or more hidden layers and is executable by computer hardware;
mitigating a bias of the machine learning model as trained by performing a prediction adjustment on the machine learning model, wherein the bias arises from the predetermined data imbalance and the prediction adjustment includes:
executing the machine learning model as trained on a set of simulation feature vectors, wherein the machine learning model classifies the set of simulation feature vectors by, assigning each simulation feature vector to one of the plurality of buckets based on feature values of each simulation feature vector; and
determining for each bucket a regression value based on an aggregation of the simulation feature vectors assigned to the bucket by the classifying performed by the machine learning model;
wherein the machine learning model is configured to predict, in real time, regression values by classifying a subsequent feature vector input into a selected bucket and output the regression value of the selected bucket.
|