Taking down malicious webshops: designing Explainable AI against growing e-commerce fraud

Abstract

The amount of internet fraud in the Netherlands is increasing to the extent where the Landelijk Meldpunt Internet Oplichting (LMIO) cannot keep up with the number of reports [2, 39, 42]. As a result, the process of evaluating webshops requires a complete redesign. Odekerken and Bex [40] designed a solution to the problem: the WEbsite Evaluation Tool (WEET). This tool utilizes Explainable AI (XAI) to extract bona fide and mala fide features of webshops and presents them to analysts of the LMIO, thereby removing manual searching and decreasing the evaluation processing time. Yet, the accuracy of XAI systems is generally low [68] and its success strongly relies on the strength of the user interface [10]. This thesis aims to (i) help the LMIO fight internet fraud and (ii) provide more insights in the academic literature by providing design guidelines for XAI interfaces. A singular fully functional demo has been created based on principles of human decision-making and usability design principles, which endured five iterative tests by the developers of WEET. Then, the performance of the demo was tested by conducting a cognitive walkthrough and evaluating various non-fictional webshops on three analysts of the LMIO. The results show that applying usability guidelines increase usability in an XAI system and applying the use of heuristics in addition to the 80/20 rule allows for fast and accurate decisions. Designers of XAI systems are encouraged to apply these rules in other contexts and verify its generalizability among other XAI contexts.