SDIPR: A Generic P2P Semantic Distance Inference Approach Based on Spring Relaxation System

Abstract

Large-scale P2P applications can benefit from the ability to predict semantic distances to other peers without having to contact them first. In this paper, we propose a generic semantic distance inference scheme, SDIPR, in P2P network, which, through measuring semantic distances from peers to a handful of other peers and the current coordinates of those peers, assigns synthetic coordinates to each peer, and can approximately predict the semantic distance between any two peers. Specifically, in our paper, the semantic distance between peers is quantitatively characterized through vector space model based on peers’ semantic profiles and weighted with the number of documents in each peer. Then, we adopt the spring relaxation method, mimicking the physical massspring system, to simulate the semantic embedding procedure, which can find minimal energy configuration corresponding to relatively accurate semantic embedding. Simulation results show that a 3-dimension Euclidean model can embed these peers with relatively high accuracy. Moreover, we compare SDIPR with the influential existing Locality Sensitive Hashing (LSH) based multi-dimension indexing approaches, and show that SDIPR performs much better than one index scheme (that is, one group of hash functions with same dimensions as our approach), and are comparable with LSH-based schemes using three indexes (total 9 dimensions).