Link prediction is the problem of predicting the location of either unknown or fake links from uncertain structural information of a network. Link prediction algorithms are useful in gaining insight into different network structures from partial observations of exemplars. However, existing link prediction algorithms only focus on regular complex networks and are overly dependent on either the closed triangular structure of networks or the so-called preferential attachment phenomenon. The performance of these algorithms on highly sparse or treelike networks is poor. In this letter, we proposed a method that is based on the network heterogeneity. We test our algorithms for three real large sparse networks: a metropolitan water distribution network, a Twitter network, and a sexual contact network. We find that our method is effective and performs better than traditional algorithms, especially for the Twitter network. We further argue that heterogeneity is the most obvious defining pattern for complex networks, while other statistical properties failed to be predicted. Moreover, preferential attachment based link prediction performed poorly and hence we infer that preferential attachment is not a plausible model for the genesis of many networks. We also suggest that heterogeneity is an important mechanism for online information propagation.