Data science and usefulness in domains of human action

Abstract

Rapid advancements in computational science and artificial intelligence are transforming virtually every industry. In many situations however, uninterpretable modelling techniques are challenging to implement, and raise significant governance, operational and ethical challenges. In this thesis, we focus on three domains where there is limited data availability, significant complexity, and human decision makers. Rather than focusing on increasing data collection in these domains, we focus on developing useful models based on readily available data, describing a model’s usefulness as being based on five characteristics of interest: performance, scalability, comprehensibility, justifiability and actionability. In Chapter 3, we model Australian Rules Football as spatial systems rather than individual possession events. Several methods are introduced to disentangle relative team performance and the functioning of sub-systems to evaluate historical games and predict future performance. Next, Chapter 4 explores startup transformation pathways in South Australia. Working with limited data in the South Australian startup ecosystem to map startup capitalisation, we follow 151 startup journeys over an eight year period to develop an approach to support policy-makers to understand ecosystem transformation, with a focus on grant interventions and private capitalisation events. In Chapter 5, we explore creativity and the writers room, working alongside the TV series Aftertaste to evaluate the limits and potential for natural language processing to support the creative process. By approaching the intersection of creativity and data analysis from the direction of usefulness, we are able to evaluate an existing method for story arc generation and rethink the approach to make it a more useful tool to support creative development. Finally, we conclude in Chapter 6 by discussing our results and the role of data science modelling in these three domains. This includes a summary of results across the three domains, the relationship between governance and data science projects, and areas for further research in each direct domain. This work presents advances in each of the three domains explored, presenting new practical approaches as well as revealing significant new areas for further research. In addition, the work demonstrates the viability of usefulness characteristics for data science research, with positive implications for governance, research, and development of complementary techniques to uninterpretable artificial intelligence and machines learning methods.