Computational methods for predicting human behaviour in smart environments

This systematic literature review presents the computational methods of human behaviour prediction research from Pentland and Liu’s seminal 1999 paper on human behaviour prediction to the latest research to date. The PRISMA framework for systematic reviews was used as the review methodology to structure this information aggregation. This review provides a high-level summary of the field with key areas identified for new research. The results show that there are frequently used datasets for training predictive models: MavHome, MavLab, LIARA, CASAS, PlaceLab, and REDD. Accuracies in the range of 43.9% to 100% for predictions of varying complexity. Common data structures for modelling behavioural data: Vectors, tables, trees, Markov models, and graphs. Algorithms that fall into three distinct categories: Machine Learning (NN, RL, LSTM), Probabilistic Graphical Models (namely Bayesian and Markov variants), and Statistical and Trend Analysis (ARIMA, Prophet). Additionally, we document other notably useful algorithms that fall outside of these three main categories including Jaro-Winkler and Levenshtein distances. Opportunities identified for further research include the use of audio as the data source for behaviour prediction methods, and applying times-series prediction machine learning algorithms (RNN, LSTM) to the smart home problem space.