Abstract
Mobility data is of great significance because of the recurring events in daily behavior. These events are consistent and have a hidden structure. Identifying the hidden structures in such data can enhance indoor location predictability and provide interesting features. In this study, we present a methodology to predict and analyze the human mobility routines in indoor environments. We consider a structured indoor environment (i.e., university campuses) and generate an indoor Wi-Fi received signal strength (RSS) dataset. Based on this Wi-Fi RSS dataset, we synthesized the location information pertaining to 400 students from 10 batches. Afterward, we used the K-nearest neighbor (KNN) algorithm to localize each student in the dataset. Then, we build the students' mobility data by considering three indoor locations: the lecture room, laboratory, and cafeteria. We represented the repeated structure in the students' mobility data by using the principal components analysis (PCA). PCA extracts the significant information from the dataset and represents this information by a set of new orthogonal vectors termed as principal components (PC). The first PC explains the largest portion of the dataset. In this context, the top four PC's is used to describe the characteristics of the entire indoor mobility space and named as eigenlocations. This study provides three main contributions. First, we approximate the student mobility behavior over a day by using the weighted sum of the students' primary eigenlocations. Second, we show how eigenlocation scheme perform in terms of inferring the student affiliations and estimating the friendship. Finally, we demonstrate the performance of proposed eigenlocation method for the synthetic indoor localization data by considering arbitrary levels of localization errors that are resulted from the used indoor positioning systems. Using the proposed eigenlocation scheme, we were able to reconstruct and predict the students' locations over a specific day with an accuracy of 84%. Additionally, we approximately obtained 99% inference accuracy for batch affiliations and friendship estimation.