Analysis of the origin of differences in position among different fingerprinting datasets

Positioning plays a key role when dealing with location based systems (LBS). Global Navigation Satellite Systems (GNSS) are the default solution when outdoors. However, those systems fail when working indoors. Nowadays there is still not a standard solution for indoor positioning. A very common technique is fingerprinting, whose main idea is that intensity of the signals (RSSI) is different in different points of the location. These systems need, therefore, a database of positions linked to a set of RSSI. However the same algorithm can give different results in different locations, what makes it difficult to compare them. To overcome this problem, several public datasets have been created to test different positioning algorithms in the same conditions. Nevertheless, the same algorithm can give very different results in different datasets. Why the same algorithm gives different results in different datasets? This paper proposes several features to characterise fingerprinting datasets to find the reasons for different results in different datasets. Fingerprinting with the $k \mathbf{n n}$ algorithm is used. We find that the main reason for these differences is in the subsets created for validation.