Estimation of the qualities of overlay links is useful for optimizing overlay networks on the Internet. Existing estimation methods requires sending large quantities of probe packets between two nodes, and the software for measurements have to be executed at both of the end nodes. Accurate measurements require many probe packets to be sent, and other communication can be disrupted by significantly increased network traffic. In this paper, we propose a link quality estimation method based on supervised learning from the previous observation of other similar links. Our method does not need to exchange probe packets, estimation can be quickly made to know qualities of many overlay links without wasting bandwidth and processing time on many nodes. We conducted evaluation of our method on PlanetLab, and our method showed better performance on path latency estimation than estimating results from geographical distance between the two end nodes.