User Profile-aware Daily Activity Prediction
Alexandros Zamichos, Maria Tsourma, Stavros Papadopoulos, Anastasios Drosou, Dimitrios K. Tzovaras · 2023
Human activity Prediction mechanisms are a challenging issue discussed within literature. Daily human activities are complex and are constituted of multiple actions, where each one provides important information about a person’s lifetime and routine. These mechanisms use this information as a baseline for the prediction of humans’ next activity, by decomposing and examining the sequences of all the activities a person has performed using information that has been collected by sensors. This paper proposes a Human Activity Prediction module based on a Recurrent Neural Networks (RNN) - Long Short-Term Memory (LSTM) network, for predicting the next activity of a user based on the information included in the user’s personalized activity profile. The activity profile is constituted of sequences of historical activities performed by an individual along with their demographic characteristics. Aside from the method, this paper presents also its three points evaluations, by comparing (a) its results with a Markov Chain model, (b) the effect that the number of activities included in a sequence has on the accuracy of prediction and (c) by evaluating the introduction of characteristics as input to the network alongside with the insertion of historical sequences or not.