CAREER: Mining and Exploring Heterogeneous Information Networks with Social Factors

异质的社会信息网络，如在线社交网络、在线论坛和数字政府，是数据分析的宝贵来源。然而，目前的信息网络研究大多忽略了所涉及的社会因素，将人与人之间的互动简单地视为图中的节点和链接。该项目提供了一种系统的方法来分析这种涉及人的因素相关问题的网络，认识到不同类型的链接与特定问题具有不同的相关性。例如，“导师”链接可能更适合推荐某人申请某一份工作，而不是去看某部电影。该项目确定了五个基本研究问题并提供了解决方案：(1)预测缺失用户和链接特征；(2)识别个性特征；(3)角色检测；(4)社交活动预测；(5)推荐系统。基本的方法是提供概率模型，该模型可以(1)结合来自领域专家的有限标签或启发式指导，(2)自动地在复杂的异质信息网络中为目标问题选择最关键的信息。例如，对于年龄组预测的用户配置问题，通过定义给定不同类型链路上的网络结构和强度的可能标签配置的概率，设计了一个概率模型。该学习算法将通过不同类型的链路传播来自少数用户的标签，并根据该链路类型上标签的配置概率来学习每种链路类型的强度。直觉是，如果“同学”链接类型将两个年龄相近的用户聚集在一起，则算法需要为连接的两个同班用户分配相同的年龄组标签，并为“同学”链接类型赋予更高的强度权重。该项目将开发一个基于Spark和GraphX的集成网络挖掘系统，以支持大规模网络上提出的算法。该系统将被用作一种研究工具，用于探索有效的近似，并保证所提出的算法的质量。

Heterogeneous social information networks, such as online social networks, online forums, and digital government, are valuable sources for data analysis. However, most of the current information network studies ignore the social factors involved and treat people and their interactions simply as nodes and links in graphs. This project provides a systematic approach for analyzing such networks that addresses human factor-related questions, recognizing that different types of links have different relevance to a particular question. For example, a "mentor" link might be much more relevant to recommending someone to apply for a particular job rather than see a certain movie. This project identifies five fundamental research problems and provides solutions to these problems in heterogeneous social information networks: (1) predicting missing user and link characteristics, (2) identifying personality traits, (3) role detection, (4) prediction of social activities, and (5) recommender systems. Together these provide a way to include social understanding in analysis of networks.The basic approach is to provide probabilistic models that can (1) incorporate guidance in terms of either limited labels or heuristics from domain experts, and (2) automatically select the most critical information in complicated heterogeneous information networks for the target problem. For example, for the user profiling problem of age group prediction, a probabilistic model is designed via defining the probability of a possible label configuration given the network structure and strengths on different types of links. The derived learning algorithm will propagate the labels from only a few users via different types of links, and the strength of each link type will be learned according to the configuration probability of labels on that link type. The intuition is that if the "classmates" link type brings two users with similar age together, the algorithm needs to assign the same age group label to the two connected users that are classmates and assigns a higher strength weight to the "classmates" link type. The project will develop an integrated network mining system based on Spark and GraphX, to support the proposed algorithms on large-scale networks. This system will be used as a research vehicle for exploring efficient approximations with quality guarantees for the proposed algorithms.