Abstract:

Because of randomness, the robustness of Label Propagation Algorithm (LPA) is severely hampered. To improve the robustness, a LPA based on potential function of data field (LPAP) was proposed. The potential of every node was calculated, and local extreme potential was searched. Only the node with extreme potential was labeled initially, and the label was updated according to the sum potential of its neighbors with equal label during iteration. When there were no nodes changing its label, iteration stopped. The experimental results show that the average distinct community partition of LPAP is 4.0% of that of LPA, 12.9% of that of Balanced Propagation Algorithm (BPA), and the average Variation of Information (VOI) of LPAP is 45.1% of that of LPA, 73.3% of that of BPA. LPAP is significantly more robust, and is suitable for community detection in large network.

Key words: Community detection, label propagation algorithm, data field, potential function, robustness

摘要：

针对标签传播算法(LPA)存在大量随机性、算法稳定性差的问题，提出了基于数据场势函数的标签传播算法(LPAP)。该算法计算所有节点的势值，搜索势值极值点。初始化时仅赋予势值极值点以标签，迭代过程中根据邻接节点中相同标签节点势值之和更新标签，所有节点标签不再改变时迭代结束。实验结果表明：该算法得到的社区划分方式平均是LPA的4.0%，是平衡传播算法(BPA)的12.9%；信息变化参数平均是LPA的45.1%，是BPA的73.3%。具有更好的稳定性，适用于大型网络的社区发现。

关键词: 社区发现, 标签传播, 数据场, 势函数, 稳定性