Improving Machine Learning-Based Wi-Fi Fingerprint Technique with Feature Selection and Grid Search Methods

This paper presents the enhancement of the WiFi-based fingerprint technique for an indoor positioning system applied to a real experimental area. In typical Wi-Fi-based fingerprint techniques, classification algorithms such as k-nearest neighbor (k-NN), decision tree (DT), and random forest (RF) are used for position estimation. However, these algorithms do not perform well with high-dimensional and large datasets. They also face limitations related to overfitting and uninformative features in datasets. This paper overcomes these challenges by deploying feature selection based on filter methods. This process removes uninformative features from the datasets before feeding them to construct a training model. The grid search method is also employed to perform hyperparameter tuning, which is used to construct the outperforming models. The experimental setup involved collecting received signal strength indicators (RSSIs) from access points (APs) in the real indoor environment to create the radio map. The accuracy performance of the proposed methods was tested by employing the feature selection method on the radio map dataset and using the grid search method to find optimized hyperparameters for constructing the training models based on the k-NN, DT, and RF algorithms. The accuracy results were compared with those of non-feature selection and default hyperparameters used for the three algorithms. The computational results demonstrate that the RF algorithm outperforms the other two algorithms. Furthermore, performance is improved when using grid search and feature selection as proposed.