Actionable Behaviour Discovery in Heterogeneous Social Graphs

点击翻译按钮获取中文摘要

Today, organizations and companies have accumulated online user activity data on unprecedented scales. Such data has become ubiquitous everywhere, ranging from online news agencies to e-commerce retailers. Mining and understanding user behaviour have crucial roles in making better business decisions, improving the quality of service, and enhancing user experience in many domains. Thus, there is a growing demand for leveraging the full potential of available data to generate practical and actionable intelligence. In the real world, users are connected (e.g., friendship in social media) and interact with different objects (e.g., news, products). These objects are usually in the form of unstructured text, which is hard to mine/analysis, and have relationships with other objects, carrying rich semantic meaning. For instance, a user may like a news article because it contains two celebrities involved in a particular event. Moreover, users and the environment (e.g., relationships) keep changing over time. Traditional approaches for user behaviour modelling overlook these complexities by pruning involved objects, ignoring their rich semantic relationships, and focusing on static data. Moreover, the simplicity of current model outputs (e.g., subscription prediction) does not lead to actionable intelligence (e.g., reasons behind subscription) in many scenarios. The long-term goal of the research program is to develop an end-to-end framework for building and inferring actionable knowledge from heterogeneous and dynamic user activity data. The framework is built upon the notion of heterogeneous social graphs, graphs with different types of nodes and links, including users, objects and their relationships. The goal is to develop theories, techniques, and tools that make the mining process simple and flexible in order to be applied to a wide range of applications in heterogeneous social graphs. The short-term objectives of the proposed research program are identified as follows. We seek to mine user behaviour in text enriched heterogeneous social graphs. Furthermore, we study the models which capture temporal dynamics in dynamic heterogeneous social graphs. In particular, we focus on developing actionable intelligence discovery in heterogeneous social graphs by designing transparent, explainable models. The anticipated outcomes of our research are: (i) a framework for building and mining of large-scale dynamic heterogeneous graphs, built from text enriched user activity data, (ii) novel algorithms, techniques, and tools that address the lack of actionability in user behaviour discovery techniques. The proposed program is of broad industrial interest as it makes sense of online user activity data, and consequently, paves the way for numerous impactful applications in a wide range of social, economic, and environmental issues. It contributes to training the demanding High Qualified Personnel (HQP) and boosts innovation in products/services across Canada.