Activity Prediction
Diane J. Cook, Narayanan Chatapuram Krishnan · 2015
This chapter introduces a compression-based approach to perform sequence prediction. It considers three alternative methods for predicting the timing of activity occurrences, which include time series-based activity forecasting, probabilistic graph-based activity prediction, and induction of activity timing rules. Sequence prediction has also been applied to challenges outside the realm of activity learning, including biological sequence analysis, speech modeling, text analysis, and music generation. Forecasting has been used in time series analysis to predict future values of a target variable given observed past and current values. Probabilistic graphs such as naive Bayes graphs, hidden Markov models (HMMs), and conditional random fields can be used to calculate a probability distribution over activity labels given observed information such as sensor events. The chapter summarizes some methods for prediction evaluation: leave n-at-the-end-out testing, Jaccard index, mean absolute error (MAE), mean squared error (MSE), and normalized root mean squared error (NRMSE).