Exploring explainability in big data policing: Identifying feature importance in decision-making models

• In Event: WG-PLACE 2: Towards an evidence-based model for big data policing (BIGDATPOL)

Abstract

The adoption of new data-driven methods has been transforming both public and private sectors. Within the context of law enforcement, the utilisation of big data and big data analytics is fundamentally altering decision-making processes at operational and strategic levels. Specifically, place-based big data policing, i.e. using big data sources to predict when and where specific crime types are likely to occur, has broadened discretionary powers to a diverse array of stakeholders involved in its application lifecycle, including software developers, analytical translators, and traditional actors like police officers.
The engagement of multiple stakeholders underscores the critical need for explainability, ensuring that all involved parties comprehend the rationale behind decisions. Understanding the "why" behind actions serves as an additional safeguard for oversight, as it keeps key users informed about the underlying reasoning behind their tasks. Moreover, explainability addresses concerns regarding the necessary justification, especially pertinent in high-risk applications such as place-based big data policing.
Drawing upon the evolving empirical literature on eXplainable Artificial Intelligence (XAI), this paper tests various techniques to assess the feature importance in big data policing models. The study aims to present an overview comparing the trade-offs among selected techniques and evaluating the implications of feature importance at both a local (specific observations) and global (overall model) level. By documenting the trade-offs inherent in these techniques, this research pivots a context-specific approach for explaining the decision-making processes and output of big data policing models.
This presentation enhances our understanding of the complexities surrounding big data policing and facilitate the development of transparent and accountable decision-making processes within law enforcement.

Authors

• Naomi Theinert, Ghent University

• Robin Khalfa, Ghent University

• Wim Hardyns, Ghent University