Quadratic Discriminant Analysis Based on Graphical Lasso for Activity Recognition

Abstract

New sensing technology are widely used in ambient and wearable sensors for human activity recognition. In this paper, we apply quadratic discriminant analysis based on graphical lasso (QDAGL) to the Opportunity activity recognition dataset. The Opportunity activity recognition dataset is a versatile human activity recognition dataset recorded in a sensor-rich environment. Quadratic discriminant analysis is a simple nonlinear discriminant analysis based on maximum likelihood estimation. Graphical lasso can estimate a covariance matrix and sparse inverse covariance matrix remarkably fast, using a coordinate descent algorithm for a single lasso problem. Because of the high dimension feature of the Opportunity dataset, the accuracy and speed of Quadratic discriminant analysis for it are not well. In this paper, we use graphical lasso to estimate the covariance matrix and inverse covariance matrix on the Opportunity activity recognition dataset. The experiments demonstrate that QDAGL can perform better and faster than QDA.