USPTO Patent Grant: Recommending ML Models Using Dataset Properties

ChangeBridge: Patent Grants - AI & Computing (G06N)

Published March 24th, 2026

Detected March 25th, 2026

Summary

The USPTO has granted patent US12585995B2 to International Business Machines Corporation for a method of recommending machine learning models based on dataset properties. The patent describes a system that extracts metadata from datasets used to train ML models, stores this information in a catalog, and uses it to identify the most suitable model for new datasets.

View original document View source feed page

What changed

USPTO Patent Grant US12585995B2, filed by International Business Machines Corporation, details a novel method for recommending machine learning (ML) models. The patent focuses on capturing dataset properties as metadata, storing this information in a model catalog, and using similarity scores to identify the most appropriate ML model for a new dataset. This approach aims to streamline the process of selecting and applying ML models by leveraging the characteristics of the data they were trained on.

This patent grant is primarily an intellectual property matter and does not impose direct regulatory obligations on companies. However, it signifies innovation in AI and ML model management. Companies developing or utilizing ML models, particularly those in the technology sector, may find the patented methodology relevant to their internal processes for model selection and deployment. The patent was filed on September 28, 2022, and granted on March 24, 2026, with 20 claims.

Source document (simplified)

← USPTO Patent Grants

Capturing data properties to recommend machine learning models for datasets

Grant US12585995B2 Kind: B2 Mar 24, 2026

Assignee

International Business Machines Corporation

Inventors

Manjit Singh Sodhi, Suja Mohandas, Nitin Gupta, Kalapriya Kannan, Prerna Agarwal

Abstract

Recommending machine learning models is provided. The method comprises training machine learning models, wherein each machine learning model is trained with a unique respective dataset. Metadata associated with each machine learning model is extracted, wherein the metadata includes properties of the respective dataset used to train the machine learning model. The machine learning models and metadata are stored in a model catalog. Upon receiving a new dataset, similarity scores are calculated between the new dataset and the machine learning models in the model catalog according to the properties of the datasets in the metadata of the machine learning models. A closest match machine learning model is identified from the model catalog for the new dataset according to similarity score. Responsive to a determination that the closest match machine learning model exceeds a similarity threshold, predictions for the new dataset are generated with the closest match machine learning model.