US 12,189,725 B2
Apparatus, method, and computer program product for multi-class classification using adapted two-level binary classification models
Waad Subber, Niskayuna, NY (US); and Ankit Singh, Apex, NC (US)
Assigned to HONEYWELL INTERNATIONAL INC., Charlotte, NC (US)
Filed by Honeywell International Inc., Charlotte, NC (US)
Filed on Jun. 12, 2023, as Appl. No. 18/332,905.
Prior Publication US 2024/0411839 A1, Dec. 12, 2024
Int. Cl. G06F 16/906 (2019.01); G06F 18/2431 (2023.01)
CPC G06F 18/2431 (2023.01) [G06F 16/906 (2019.01)] 16 Claims
OG exemplary drawing
 
1. An apparatus comprising at least one processor and at least one non-transitory memory comprising program code stored thereon, wherein the at least one non-transitory memory and the program code are configured to, with the at least one processor, cause the apparatus to at least:
transform an original training data set corresponding to a multi-class classification task into at least two adapted training data sets each corresponding to a binary classification task, the at least two adapted training data sets including a coarse-level adapted training data set and at least one fine-level adapted training data set, wherein a coarse-level instance of a binary classification model configured to perform the binary classification task is trained based at least in part on the coarse-level adapted training data set, and at least one fine-level instance of the binary classification model is trained based at least in part on, respectively, each at least one fine-level adapted training data set, wherein the original training data set comprises a multi-class classification of training objects each into an original classification set of at least three original classification sets, and wherein the coarse-level adapted training data set comprises a binary classification of the training objects from the original training data set each into one of:
a majority classification set corresponding to a particular original classification set of the at least three original classification sets into which a largest proportion of the training objects are classified in the original training data set relative to other original classification sets of the at least three original classification sets; and
a non-majority classification set representing a combination of the other original classification sets of the at least three original classification sets other than the particular original classification set to which the majority classification set corresponds;
transform binary output data generated with respect to an input data set using the trained coarse-level instance of the binary classification model and the trained at least one fine-level instance of the binary classification model into multi-class output data corresponding to the multiclass classification task; and
cause performance of at least one enterprise management operation based at least in part on the multi-class output data.