Two-Phase Combined Model to Improve the Accuracy of Indoor Location Fingerprinting

Abstract

Wi-Fi Fingerprinting based Indoor Positioning System (IPS) aims to help locate and navigate users inside buildings. It has become a popular research topic in recent years. For the most parts, authors use the traditional machine learning algorithms to enhance the accuracy of locationing. Their methods involve using a standalone algorithm or a combination of different algorithms in only one phase, producing results with an acceptable accuracy. In this paper, we present a different approach applying a machine learning model that combines many algorithms in two phases, and propose a feature reduction method. Specifically, our research is focused on the combination of different regression and classification algorithms including K-Nearest Neighbor (KNN), Support Vector Machines (SVM), Random Forest (RF), Extra Tree Regressor (extraTree), Light Gradient Boosting Machine (LGBM), Logistic Regression (LR) and Linear Regression (LiR) to create a new data set and models that can be used in the training phase. These proposed models are tested on the UJIIndoorLoc 1 dataset. Our experimental results show a prediction accuracy of 98.73% by floor, and an estimated accuracy of 99.62% and 99.52% respectively by longitude and latitude. When compared with the results of the models in which we use independent algorithms, and of other researches that have different models using the same algorithms and on the same dataset, most of our results are better.