Toward Pioneering Sensors and Features Using Large Language Models in Human Activity Recognition

TL;DR

This paper introduces a novel feature engineering approach for human activity recognition using Large Language Models, reducing sensor requirements while maintaining high accuracy.

Contribution

It proposes a method leveraging LLMs to automatically discover new sensor locations and features, reducing reliance on human expertise.

Findings

Achieved comparable accuracy with fewer sensors

Demonstrated efficient feature extraction using LLMs

Validated on Opportunity Dataset

Abstract

In this paper, we propose a feature pioneering method using Large Language Models (LLMs). In the proposed method, we use Chat-GPT 1 to find new sensor locations and new features. Then we evaluate the machine learning model which uses the found features using Opportunity Dataset [ 4 , 9]. In current machine learning, humans make features, for this engineers visit real sites and have discussions with experts and veteran workers. However, this method has the problem that the quality of the features depends on the engineer. In order to solve this problem, we propose a way to make new features using LLMs. As a result, we obtain almost the same level of accuracy as the proposed model which used fewer sensors and the model uses all sensors in the dataset. This indicates that the proposed method is able to extract important features efficiently.