Feature selection on database optimization for Wi-Fi fingerprint indoor positioning

Indoor location-based services have become very popular, principally, because of its wide and valuable applications. On that context, Wi-fi fingerprinting based on the received signal strength indicator (RSSI) has become very popular, due the fact that RSSI values are easily acquired. On the Wi-fi fingerprint method, machine learning algorithms are trained on the constructed fingerprint database and then used on a new entry to give the indoor location based on its estimations. Choosing the correct machine learning algorithm is one of the main problems in the literature. However the database sizes used during the training phase is also one of the main concerns. In this paper, a proposed feature selection method used on the original UJIIndoorLoc database created a smaller version of it, with the 30 highest RSSIs after the APIDs responsible for then in descending order, and created even smaller database subsets. Both databases, the original UJIIndoor Loc database and ours, were split into smaller subsets that were used on the classification problem according the DESIP method proposed in [1]. Six machine learning algorithms were deployed for training and testing the two database subsets with the classification attributes modified for symbolic localization. The J48 with the AdaBoost iterative algorithm gave the best results on both database subsets. The minimized database subsets showed smaller elapsed time results for all the classifications that were done. The accuracy results show similar results for both database subsets, on building and floor classification. Although, on the region attribute, the database subset with 520 attributes got better accuracy results than the reduced one.