Toward Advanced Indoor Mobility Models Through Location-Centric Analysis: Spatio-Temporal Density Dynamics

Abstract

Building's density, as its number of nodes at a specific period, is a significant parameter that affects mobile and smart applications performances and evaluations. Consequently, the buildings' temporal density predictions and their nodes spatial distribution modeling have to follow real-world scenarios to provide a realistic evaluation. However, there is lack of real-world building-level density studies that examine these aspects thoroughly. As a result, this work is a data-driven study that investigates the temporal density predictability and spatial density distributions of more than 100 real buildings with ten different categories, over 150 days across three semesters. The study covers the buildings nodes' temporal modeling and predictions, and their spatial distributions in the building. Seasonal predictive models are utilized to predict hour-by-hour density for a variable length of consequent periods using training data with different lengths. The models include Seasonal Naive, Holt-Winters' seasonal additive, TBATS, and ARIMA-seasonal. The results show that the Seasonal Naive model is often selected as the best predictive model when training phase covers a shorter period. For example, Seasonal Naive predicted with the least error in 73%, 63% and 57% of cases in summer, spring, and fall respectively when using only one week to predict its consecutive five weeks with mean normalized error 25% on average. However, when using five weeks of data to predict the sixth week, the TBATS model predicted with the least error in 60%, 54% and 43% of cases in fall, spring and summer respectively with mean absolute error 19% on average. When investigating the spatial density distributions, power law, log-logistic and lognormal distributions are usually selected as the first best-fit distributions for 82%, 65%, 62% of buildings in the summer, spring and fall respectively.