Peer-to-peer (P2P) file sharing systems are now at the origin of most of Internet traffic. Improving the performance of the query mechanism of such systems has generated a lot of interest both in industry and academia. In a P2P system, peers are connected to a subset of other peers with which they can communicate. Each peer maintains a cache and makes available its contents to the rest of the system. Connecting peers sharing similar interest in the context of a given application has recently been identified as a sound basis to improve the search efficiency. Nevertheless, capturing such interest-based (or semantic) proximity patterns is a difficult task. Most of current approaches measure this proximity between peers as the overlap between their cache contents. Given the well-known popularity patterns of peer-to-peer file sharing systems, the overlap between cache contents of two peers may not reflect accurately their semantic proximity. In this paper we propose {\sc ProxSem}, a refined proximity measure taking into account peer generosity and file popularity. We evaluated the proposed solution by simulation against a real peer-to-peer file sharing system (eDonkey) workload and results show the effectiveness of the proposed approach. While peers generosity can easily be computed locally, file popularity may require a global knowledge of the system. We also propose in this paper an epidemic algorithm to compute in a fully decentralised fashion an estimation of files popularity.