US 12,462,151 B2
Generating new machine learning models based on combinations of historical feature-extraction rules and historical machine-learning models
Haichun Chen, Sunnyvale, CA (US)
Assigned to ADOBE INC., San Jose, CA (US)
Filed by Adobe Inc., San Jose, GA (US)
Filed on May 22, 2018, as Appl. No. 15/986,043.
Prior Publication US 2019/0362222 A1, Nov. 28, 2019
Int. Cl. G06N 3/08 (2023.01); G06F 11/34 (2006.01); G06F 18/22 (2023.01); G06N 3/048 (2023.01); G06N 20/00 (2019.01)
CPC G06N 3/08 (2013.01) [G06F 11/3476 (2013.01); G06F 18/22 (2023.01); G06N 3/048 (2023.01); G06N 20/00 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A method for generating machine-learning models, wherein the method includes one or more processing devices performing operations comprising:
receiving, from a client device, a request to generate a machine-learning (ML) model, wherein the request identifies a test interaction dataset and a task for the ML model;
identifying characteristics of the test interaction dataset by applying, to the test interaction dataset, a neural network, wherein the characteristics comprise one or more of metadata, datatypes, and data distributions in interaction datasets, wherein the neural network is trained to recognize one or more of the metadata, the datatypes, and the data distributions in the interaction datasets, and wherein the neural network is separate from the ML model;
searching historical interaction datasets that match the characteristics identified by the neural network;
selecting, based on the historical interaction datasets that match the characteristics of the test interaction dataset:
(i) a plurality of historical ML models that were previously applied to the historical interaction datasets, and
(ii) historical feature-extraction rules that were previously used to extract portions of the historical interaction datasets as inputs to the plurality of historical ML models;
generating an output ML model from a combination of (i) a feature-extraction rule from the historical feature-extraction rules and (ii) an ML model from the plurality of historical ML models, wherein the output ML model is generated based on the combination satisfying a performance metric included in the request; and
providing the client device with access to the output ML model.