MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202621063941 A) filed by Anish Sachin Katariya; Kuldeep Vaydande; Prajakta Pawar; Shagupta M. Mulla; Aditya Bhandare; Rizwan Shaikh; Rohit Yeole; Siddhika Zanje; and Soniya Warade on May 20, 2026, for An Enhanced Binary Variant Of Qihpa For Efficient Feature Selection In Classification Task.
Inventors include Anish Sachin Katariya; Kuldeep Vaydande; Prajakta Pawar; Shagupta M. Mulla; Aditya Bhandare; Rizwan Shaikh; Rohit Yeole; Siddhika Zanje; and Soniya Warade.
The application for the patent was published on July 17, 2026, under issue no. 29/2026.
Abstract: In machine learning, feature selection is very relevant in analyzing data of high dimensions in which there are redundant features that reduce the quality of the model and enhance the cost of computation. The proposed paper is also presenting a new metaheuristic, the Binary Quadratic Interpolated Hybrid Pathfinder Algorithm, as a wrapper feature selection algorithm to classification problems. The BQIHPA extends continual QIHPA to discrete binary search spaces by the means of a sigmoid transfer function that would facilitate the potential efficient exploration of good feature subset combinations without a tradeoff between the quality of the result. To this end, the BQIHPA performance was measured and compared with three binary optimization methods, which include Binary Particle Swarm Optimization, Binary Grey Wolf Optimizer and Binary Whale Optimization algorithms on three datasets i.e., LIBRAS Movement (90 features), Parkinson Disease Detection (22 features) and Sonar Rock vs. Mine (60 features) with the aim of exploring robustness in various dimensionalities. Precision of classification, rate of reduction of features, rate of convergence, execution of a computationally process and memory consumption were observed. Equally, the suggested BQIHPA performed flawlessly in terms of classification accuracy at an average of 83.57, average feature reduction rate at 64.1, and they also were computationally lightweight and much faster than highly complicated baselines like BWO and BGWO by an average of 5.2 times. The specialized algorithms are optimally applied to a particular set of data, including 53.3% feature reduction on high-dimensional Libras by BGWO and 91.53% on low- dimensional Parkinson by BWO, although no single baseline can be effective across all of the measures applied. The usefulness of BQIHPA is that it has a strong balanced operation without the need to modify problem-specific parameters. In order to substantiate such a statement, the ablation experiment proves that each of the parts contributes greatly to algorithmic performance-a deletion of an element results in the corresponding 8-24% accuracy loss. This makes BQIHPA a useful general-purpose generalizable solution to automated feature selection pipelines in which reliable consistent performance should be a priority, over incremental improvements in accuracy on particular problem types
Disclaimer: Curated by HT Syndication.
