K-Neares Neighbors (KNN) and its variant weighted KNN (WKNN) have been explored for years in both academy and industry to provide stable and reliable performance in WiFi-based indoor positioning systems. Such algorithms estimate the location of a given point based on the locality information from the selected nearest WiFi neighbors according to some distance metrics calculated from the combination of WiFi received signal strength (RSS). However, such a process does not consider the relational information among the given point, WiFi neighbors, and the WiFi access points (WAPs). Therefore, this study proposes a novel Deep Neighborhood Learning (DNL). The proposed DNL approach converts the WiFi neighborhood to heterogeneous graphs, and utilizes deep graph learning to extract better representation of the WiFi neighborhood to improve the positioning accuracy. Experiments on 3 real industrial datasets collected from 3 mega shopping malls on 26 floors have shown that the proposed approach can reduce the mean absolute positioning error by 10% to 50% in most of the cases. Specially, the proposed approach sharply reduces the root mean squared positioning error and 95\% percentile positioning error, being more robust to the outliers than conventional KNN and WKNN.