[论文解读] Predicting Graph Categories from Structural Properties

James P. Canning, Emma E. Ingram+6 more|arXiv (Cornell University)|May 7, 2018

Complex Network Analysis Techniques参考文献 47被引用 6

一句话总结

该论文提出一种仅基于结构属性预测复杂网络领域类别的方法，在真实与合成网络的组合数据集上，使用随机森林分类器实现了96.6%的准确率。结果表明，不同的结构特征指纹可实现高精度分类，且合成图几乎可完全根据其生成模型进行预测。

ABSTRACT

This paper has been withdrawn from arXiv.org due to a disagreement among the authors related to several peer-review comments received prior to submission on arXiv.org. Even though the current version of this paper is withdrawn, there was no disagreement between authors on the novel work in this paper. One specific issue was the discussion of related work by Ikehara \& Clauset (found on page 8 of the previously posted version). Peer-review comments on a similar version made ALL authors aware that the discussion misrepresented their work prior to submission to arXiv.org. However, some authors choose to post to arXiv a minimally updated version without the consent of all authors or properly addressing this attribution issue. ================ Original Paper Abstract: Complex networks are often categorized according to the underlying phenomena that they represent such as molecular interactions, re-tweets, and brain activity. In this work, we investigate the problem of predicting the category (domain) of arbitrary networks. This includes complex networks from different domains as well as synthetically generated graphs from five different network models. A classification accuracy of $96.6\%$ is achieved using a random forest classifier with both real and synthetic networks. This work makes two important findings. First, our results indicate that complex networks from various domains have distinct structural properties that allow us to predict with high accuracy the category of a new previously unseen network. Second, synthetic graphs are trivial to classify as the classification model can predict with near-certainty the network model used to generate it. Overall, the results demonstrate that networks drawn from different domains (and network models) are trivial to distinguish using only a handful of simple structural properties.

研究动机与目标

• 探究是否可仅基于结构属性对来自不同领域的复杂网络进行分类。
• 评估机器学习模型在仅依靠结构特征的情况下区分现实世界网络与合成网络的性能。
• 确定合成网络模型是否留下可检测的结构痕迹，从而实现准确分类。
• 评估结构特征在多样化网络领域与模型之间的泛化能力。

提出的方法

• 使用从每个网络中提取的14种简单结构属性，训练随机森林分类器。
• 数据集包含来自五个领域的现实网络以及由五个不同网络模型生成的合成图。
• 对特征进行归一化，并用于在真实与合成网络的组合数据集上训练分类器。
• 通过10折交叉验证评估模型性能，以确保稳健性与泛化能力。
• 在先前未见过的网络上测试分类器，以评估预测准确率。
• 本研究仅关注结构特征，避免使用领域特定或语义信息。

实验结果

研究问题

• RQ1 是否可仅基于其结构属性准确分类来自不同领域的复杂网络？
• RQ2 机器学习模型仅使用结构特征时，能多好地区分现实世界网络与合成网络？
• RQ3 合成网络模型在多大程度上留下独特的结构特征，从而可被识别？
• RQ4 是否存在某些结构属性在多样化领域中对网络类别具有最强预测力？

主要发现

• 使用仅结构属性的随机森林分类器在区分网络类别时达到了96.6%的分类准确率。
• 合成图被近乎确定地分类，表明每种网络模型均产生独特且可识别的结构模式。
• 结果证实，不同领域的网络表现出独特的结构指纹，足以实现高精度分类。
• 本研究表明，简单结构特征在多样化网络类型（包括现实与合成图）之间具有高度区分性。
• 高准确率表明，仅依靠结构属性即可区分网络类别，即使缺乏领域特定知识亦可。