In order to achieve effective integration and accurate fusion of multi-source Point of Interest (POI) data, a Mutually-Nearest Method considering Distance and Category (MNMDC) was proposed. Firstly, for spatial attributes, standardized weight algorithm was used to calculate the spatial similarity of the object to be fused, and the fusion set was obtained. Secondly, for non-spatial attributes, Jaro-Winkle algorithm was used to eliminate some objects with consistent categories by a low threshold, and exclude some objects with inconsistent categories by a high threshold. Finally, non-spatial Jaro-Winkle algorithm with distance constraint, category consistency constraint and high threshold was used to find out the missing objects in the spatial algorithm. The experimental results show that the average accuracy reaches 93.3%, compared with Combined Normal Weight and Title-similatity algorithm (COM-NWT) and the grid correction methods, the accuracy of MNMDC method in seven different groups of coincidence degree data, the average accuracy increases by 2.7 percentage points and 1.6 percentage points, the average recall increases by 2.3 and 1.4 percentage points. The MNMDC method allows more accurate fusion of POI data during actual fusion.

Due to the sparseness of user's historical location data collected by Global Positioning System (GPS) devices, the capability of location prediction model based on single user data was limited. Therefore, a new user location prediction model based on Author Topic Model (ATM) and Radiation Model (RM) was proposed. In the time dimension, the user group that similar to the target user was discovered by using ATM, and the target state of the user group at the prediction time was determined. In the spatial dimension, the RM algorithm was used to calculate the probabilities of target user's candidate location in the target state, and the user's target predictive location could be achieved by comparing the probability value of each candidate location to determine the location where the target user might occur. The experimental results show that the average prediction accuracy of the model is 61.49%, which is nearly 28 percentage points higher than that of the Markov model based on variable order. The proposed model can obtain higher prediction accuracy under the condition of small amount of single user data.