The rapid advance in lifelog data which is captured through smartphones has shown the need for sophisticated tools to better understand human’s behaviour from physical activity. Analysing lifelog data is challenging and complex task due to its nonstationary nature. Designing proper human activity recognition (HAR) techniques is necessary to obtain a significant interpretation for human behaviour. In respond, we deigned a spectral graph wavelet transform (SGWT) based machine learning model for HAR. Each row of human activity data is transferred into an undirected graph and then, SGWT is applied. The SGWT coefficients are investigated and used to classify physical human activity. We employed the principal component analysis model to reduce feature dimensionality. The extracted features are sent to several classification models such as a support vector machine (SVM), stacking, Boosting, least support vector machine (LS-SVM), k-nearest (KNN). The proposed model is evaluated using a publicly available dataset. The stacking classifier obtained 0.97% accuracy. The proposed model for HAR obtained an outstanding classification rate compared to the previous models. The results proved that the proposed model is a useful tool for analysing human behaviour.