Intellectual Property India Publishes Patent Application for Method And System For Removing Learned Feature Interactions From Trained Additive Machine Learning Models

MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202641126816 A) filed by Vellore Institute Of Technology on October 24, 2026, for Method And System For Removing Learned Feature Interactions From Trained Additive Machine Learning Models.

Inventors include Ramalingam Murugan; and Rajeswari C.

The application for the patent was published on October 30, 2026, under issue no. 44/2026.

Abstract: ABSTRACT METHOD AND SYSTEM FOR REMOVING LEARNED FEATURE INTERACTIONS FROM TRAINED ADDITIVE MACHINE LEARNING MODELS A method and system remove learned information from a trained additive machine learning model (108) in response to an unlearning request. The model computes a predicted output (214) as a sum of a bias term (210), main-effect functions (204) each associated with an input feature, and interaction-effect functions (206) for input-feature combinations, and is modified without retraining. A target interaction-effect function (208) of constituent input features is designated and purified by iteratively projecting the function onto smooth one-dimensional functions of each constituent input feature and reassigning the projected components to corresponding main-effect functions, until a purified target interaction effect (222) orthogonal to any smooth function of each constituent input feature alone is obtained, the predicted output remaining unchanged. The purified effect is excised by direct subtraction from the sum without gradient-based retraining, yielding an excised model (128) retaining the main-effect and remaining interaction-effect functions. (Fig. 2)

Disclaimer: Curated by HT Syndication.