Developing robust methods for the measurement of online social media populations

点击翻译按钮获取中文摘要

As human activities move online, governments, companies, and organizations want to provide optimized experiences to users of online services. Moreover, they see an opportunity to track online and offline human behavior through online user activity. All this requires knowing details about those users: e.g., their age, gender, interests, and general geographic location. Users may not provide such details for a variety of reasons. This motivates the need for methods that infer user attributes from their online content. Research in the area of demographic inference has shown that many such attributes can be accurately inferred from the textual and behavioral records left by users on a platform. The problem is that, while, researchers have made attempts to infer demographic features such as age, political orientation, income, and education, the accuracy of such methods leaves much to be desired. Few viable inference systems exist for actual use outside of research. In this initiative, we will tackle two of the major opportunities for improving the performance of demographic inference methods. First, we will develop methods for learning the nature of the relationship (e.g., family, friendship, in-person, fan) that underpin links in the online social graph. This information is both valuable in itself and also as an input for other analysis and classification tasks. Despite this, no work to date has been done on this specific version of relationship learning. Second, we will develop rich models that can accurately assess the topics users engage with. A major limitation of topical features used in existing demographic inference methods is that they are derived from the dominant language used on the platform. This inherently limits the richness of the resulting topic model. We will develop generalized topic models learned from community-curated discussion platforms such as StackOverflow and Reddit. In recent years, demand for new approaches to demographic inference have generated much research without many results. Here we return to the fundamentals of the field - constructing new and informative user features. Our work will provide generalized tools that improve the inference of a many user attributes. In addition, our work will make direct contributions to important and open problems in the fields of natural language processing and network science.