Leveraging Knowledge Graph for Intelligent Business Information Search


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date6 Jul 2020


Business information search refers to the process by which a person identifies the most appropriate item(s) to meet the information need in business domains. Despite continuous improvements in search algorithms over the past few decades, search remains iterative and time-consuming in some areas such as literature review, patent review, or target screening in investment activities. Motivated by the recent debate on whether artificial intelligence (AI) will replace humans, this dissertation aims to realize AI’s full potential in search by shifting more human efforts to machine work. In particular, we leverage the emerging technology of knowledge graph, which is a multi-relational graph comprising entities of different types and complex relations among them, to develop innovative search algorithms to facilitate business information search. We investigate into the problems of textual information search and target selection in investment activities with three studies.

The first study examines the problem of textual information search such as patent search and academic literature search. To improve the search productivity of knowledge workers in complex information search, we advocate a new search paradigm called Search By Example (SBE) that strives to reduce human involvement while delivering good search results in fewer iterations as a complement to the traditional paradigm of Search By Keyword (SBK). To demonstrate the unique value of SBE, we develop a proof-of-concept SBE tool with innovative use of existing and new search techniques and conduct a randomized laboratory experiment in the context of academic literature search. The research results show a significant improvement of search productivity in terms of search accuracy per unit cost of search effort using the SBE tool in comparison with the SBK tool. Findings suggest that the SBE tool is more effective than the SBK tool in the given research setting. Furthermore, a survey conducted after the experiment reveals that users are positive about their experience with the SBE tool and are willing to adopt it when available. We further rationalize the value of the SBE paradigm via the information foraging theory and suggest future research directions.

The second and third studies investigate into the problem of target screening for investors, which is also an important search task in the business context. Venture capital has been growing rapidly in the past decades and gained massive attention in academic literature. As a critical first step for effective venture capital investment, the selection of appropriate startup companies is challenging due to the large number of potential targets and the high uncertainty of potential financial or strategic benefits. In the second study, we tackle this business problem through data science approaches and develop an innovative framework based on knowledge graph to predict future investment behaviors of venture capitalists (VCs). We first build a domain-specific knowledge graph to describe the characteristics of VCs, startups, and employees and the complex interactions among them. Then, we exploit the advanced knowledge graph embedding (KGE) approach to model entities and relations in the graph into low-dimensional vector spaces. Based on the entity and relation embeddings, we estimate the plausibility of the existence of an investment link between a focal VC and a startup and perform predictions. Computational experiments on the CrunchBase dataset suggest that the proposed framework achieves better performances than the baseline methods, namely item-based collaborative filtering, singular value decomposition, and feature-based model.

The third study examines the measure of social proximity, which is one the most important determinants of investment decisions. Existing social proximity measures are mainly based on unimodal or uniplex networks and cannot effectively support applications in heterogeneous networks. Hence, we develop a novel social proximity measure named “Entity Proximity” through a KGE approach, which models different entities and their relations within a graph in continuous vector spaces. Compared with a number of existing measures, entity proximity not only provides a finer-grained assessment of social proximity but also is able to incorporate different types of relations and entities at the same time. We validate the proposed measure in the business context of venture capital investment. Results show that entity proximity has the highest predictive power on investment decisions than existing measures. Further, we demonstrate that entity proximity can integrate social cohesion and structural equivalence in a traditional unimodal and uniplex network through correlation and VC co-investment analyses.

    Research areas

  • search by keyword, search by example, search productivity, knowledge worker, artificial intelligence, knowledge graph, graph embedding approach, venture capital, social proximity, network analysis, information search, FinTech, structural equivalence, social cohesion, paradigm innovation