5G mmWave technology can turn multipath into a friend, as multipath components become highly resolvable in the time and angle domains. Multipath signals have not only been used in the literature to position the user equipment (UE) but also to create a map of the surrounding environment. Yet, many multipath-based methods in the literature share a common assumption, which entails that multipath signals are caused by single-bounce reflections only, which is not usually the case. There are very few methods in the literature that accurately filters out higher-order reflections, which renders the exploitation of multipath signals challenging. This paper proposes an ensemble learning-based model for classifying signal paths based on their order of reflection using 5G channel parameters. The model is trained on a large dataset of 3.6 million observations obtained from a quasi-real ray-tracing based 5G simulator that utilizes 3D maps of real-world downtown environments. The trained model had a testing accuracy of 99.5%. A single-bounce reflection-based positioning method was used to validate the positioning error. The trained model enabled the positioning solution to maintain sub-30cm level accuracy 97% of the time.