07-社会网络分析与算法研究.pdfVIP

社会网络中的链接预测问题 链接预测 ¢  网络中的链接预测(Link Prediction)问题是指通过已 知的网络节点以及网络结构等信息预测网络中尚未产 生连边的两个节点之间产生链接的可能性。 ¢  这种预测既包含了对未知链接（existed yet unknown links ）的预测 ，也包含了对未来链接（future links ）的预测。该问题的研究在理论和应用两个方面都具 有重要的意义和价值。 TEMPORAL LINK PREDICTION APPLICATIONS ¢  To suggest interactions or collaborations that haven’t yet been utilized within an organization ¢  To monitor terrorist networks - to deduce possible interaction between terrorists (without direct evidence) ¢  Used in Facebook and LinkedIn to suggest friends ¢  Reduce experimental costs (missing links) Discovery of links/interactions of biological networks costs much. Instead of blindly checking all possible links, to predict in advance and focus on the most likely existing links can sharply reduce costs if the predictions are accurate enough. METHODS FOR LINK PREDICTION ¢  All the methods assign a connection weight score(x,y) to pairs of nodes x, y, based on the input graph, and then produce a ranked list in decreasing order of score(x, y). —  Can be viewed as computing a measure of proximity or “similarity” between nodes x and y. ¢  Existing link prediction models can be categorized into two large classes: unsupervised learning models and supervised learning models. ¢  Unsupervised learning models compute scores for pairs of nodes based on the link structure, and do not involve any learning: node-similarity based methods and path -dependent methods. ¢  Supervised learning models make predictions on the unobserved relations by learning a supervised model from the observable data. —  Typical supervised models include feature-based classification methods, probabilistic graphical models, and matrix factorization based latent factor models. GRAPH DISTANCE COMMON NEIGHBORS E ¢  Graph distance: (Negated) length of shortest path bet