Outdoor location is an important concern for mobile carriers to optimize wireless telecom network and coverage. When GPS information is not fully available, techniques such as fingerprinting can be developed to retrieve positioning. In this article, a novel fingerprint algorithm based on WKNN is proposed. This algorithm introduces additional weights computed with random forests and based on importance of each feature collected via measurement report. It also makes adaptive selection of reference points by discarding untrustworthy ones in order to refine precision of each user equipment positioning prediction. The whole methodology is implemented with Apache Spark framework and tested on real data flow from a major Chinese telecom carrier. Results show significant improvement of the methodology compared to both TA+AOA and classic WKNN fingerprinting methods.