Abstract

Applications of artificial intelligence play an increasingly important role in diagnosing and treating cancer. The complexity of this dynamic research field requires scientists coming from various backgrounds to continuously collaborate and create multi-disciplinary teams. However, finding the potential collaborator(s) for high-quality research effectively and efficiently is considered a difficult task for many stakeholders. In this thesis, we address these problems through developing co-authorship predictive model not only to predict potential co-authorships but also to interpret and explain which factors play essential roles in selecting potential collaborations. Finding these factors may help policymakers and research organizations investigate drivers as well as constraints in order to facilitate fruitful research collaboration and the formation of strong research teams.

The thesis has two research objectives. The first one is to characterize and map the recent research landscape in the field of artificial intelligence applications for cancer diagnosis and treatment. The second one is to explore the driving factors for different co-authorship patterns of researchers working in this field. We first used NLP techniques to characterize the evolution of artificial intelligence in the cancer research area. We observed great interest of researchers who have been gradually moving from conventional to advanced learning techniques. We then employed complex network analysis with co-authorship as a proxy for collaboration and constructed several co-authorship networks of researchers working in this field. We extracted different structure-based and attribute-based metrics related to the individual authors and their collaboration patterns. Finally, we used machine learning and interpretability techniques to predict collaboration links of co-authorship patterns and to interpret the machine learning models. We were able to successfully predict future co-authorship links for various collaboration patterns, namely new co-authorships, persistent co-authorships and discontinued co-authorships. In general, our results show that common neighbors-based and discipline similarity factors have a positive impact on the appearance of co-authorship links. We conclude that using machine learning models and interpretability techniques is a useful and effective way to predict potential co-authorships and derive driving factors for collaboration.