基于因果作用估计的网络节点分类方法研究

The class labels of nodes in networked data are not only affected by the influence of nodes' features, but also by the peer influence of the neighboring nodes, these two kinds of influences should be taken into account when classifying nodes. However, existing studies mainly use peer influence or the influence of nodes' features to classify nodes and fail to consider both effectively. Thus, developing classification methods that take both the peer influence and the influence of nodes' features into consideration becomes an essential and challenging research direction. Existing studies have indicated that there is a causal relationship between peer influence and nodes' class labels. Therefore, this proposal intends to study the node classification based on causal effect estimation. More specifically, our research contents include: (1) designing a causal effect estimation based approach to estimate the weights of peer influence and nodes' features to nodes' class labels; (2) studying the heterogeneous causal effect of peer influence and nodes' features on different subgroups in the network, and designing a grouping method to partition nodes into several subgroups, then the weights of peer influence and the nodes' features on each subgroup are estimated separately to improve accuracy; (3) using the weights of peer influence and nodes' features as parameters to construct a classification algorithm that takes both peer influence and nodes' features into consideration; (4) The outputs of the project will apply to real-world video-on-demand datasets from telecom network for the classification of user interest, and will design a prototype system to validate the effectiveness of the methods proposed in the project and provide in-depth studies and further improvement of these research contents. This proposal integrates causal effect estimation into node classification, providing new theoretical and scientific exploration for node classification research.

网络数据中节点类别的判断不仅受节点特征影响，也受到相邻节点的同伴影响，分类时应兼顾这两种影响。已有工作主要利用同伴影响或节点特征单一因素来分类。因此，研究兼顾节点特征和同伴影响的分类方法可以提高分类质量。研究表明同伴影响与节点类别之间是因果关系。基于此，本项目拟开展基于因果作用估计的网络节点分类方法研究。主要内容包括：（1）基于因果作用估计理论，研究节点特征和同伴影响对节点类别的影响权重的计算模型；（2）针对一些网络数据中同伴影响和节点特征对不同节点的因果作用存在异质性的问题，研究节点子群划分方法，对不同子群分别计算，获得更加准确的影响权重；（3）将影响权重作为参数，设计兼顾节点特征和同伴影响的节点分类算法；（4）以电信网络影视点播数据为例构建用户兴趣分类原型系统，验证提出的方法的有效性，完善和深化研究工作。本项目将因果作用估计用于网络节点分类，为网络节点分类研究提供新的理论和科学探索。

现实世界中大量数据以网络的形式组织。节点分类旨在面向网络中的节点进行类别预测学习，是网络数据分析的重要前提之一，因此研究节点分类方法具有重要的现实意义。然后，已有工作主要利用同伴影响或节点特征单一因素来分类。本项目以分析同伴影响和特征影响为切入点，研究网络节点分类问题。具体研究内容包括：相邻个体的同伴影响的因果作用研究，基于图模式匹配的子群发现，基于特征选择的多标签学习模型研究。结合工作基础和相关领域的发展趋势，研究了面向链接预测的知识图谱知识融合问题。在本项目的支持下，项目负责人发表了7篇论文，包括在数据挖掘领域顶级期刊ACM Transactions on Knowledge Discovery from Data论文1篇和顶级会议IEEE International Conference on Data Mining论文1篇，并且申请专利3项。本项目为网络和图数据的学习提供了新的概念和方法探索，为研发具有实用价值的网络和图数据分类方法做出了有意义的尝试。