Link prediction exercises may prove particularly challenging with noisy and incomplete networks, such as criminal networks. Also, the link prediction effectiveness may vary across different relations within a social group. We address these issues by assessing the performance of different link prediction algorithms on a mafia organization. The analysis relies on an original dataset manually extracted from the judicial documents of operation “Montagna”, conducted by the Italian law enforcement agencies against individuals affiliated with the Sicilian Mafia. To run our analysis, we extracted two networks: one including meetings and one recording telephone calls among suspects, respectively. We conducted two experiments on these networks. First, we applied several link prediction algorithms and observed that link prediction algorithms leveraging the full graph topology (such as the Katz score) provide very accurate results even on very sparse networks. Second, we carried out extensive simulations to investigate how the noisy and incomplete nature of criminal networks may affect the accuracy of link prediction algorithms. The experimental findings suggest the soundness of link predictions is relatively high provided that only a limited amount of knowledge about connections is hidden or missing, and the unobserved edges follow some kind of generative law. The different results on the meeting and telephone call networks indicate that the specific features of a network should be taken into careful consideration.

对于嘈杂和不完整的网络（例如犯罪网络），链接预测练习可能会证明特别具有挑战性。同样，链接预测有效性可能会在一个社会群体中的不同关系之间变化。我们通过评估黑手党组织上不同链接预测算法的性能来解决这些问题。该分析依赖于原始数据集，该原始数据集是从意大利执法机构针对西西里黑手党成员个人进行的“蒙塔尼亚”行动司法文件中手动提取的。为了进行分析，我们提取了两个网络：一个网络包括会议，另一个网络记录犯罪嫌疑人之间的电话。我们在这些网络上进行了两个实验。第一，我们应用了几种链接预测算法，并观察到即使在非常稀疏的网络上，利用全图拓扑（例如Katz得分）的链接预测算法也可以提供非常准确的结果。其次，我们进行了广泛的模拟，以研究犯罪网络的嘈杂和不完整的性质如何影响链接预测算法的准确性。实验发现表明，只要隐藏或丢失了关于连接的有限知识，并且未观察到的边缘遵循某种生成定律，链接预测的可靠性就相对较高。会议和电话网络上的不同结果表明，应仔细考虑网络的特定功能。