Getting the Message?

Abstract

In many of today's online applications that facilitate data exploration, results from information filters such as recommender systems are displayed alongside traditional search tools. However, the effect of prediction algorithms on users who are performing open-ended data exploration tasks through a search interface is not well understood. This paper describes a study of three interface variations of a tool for analyzing commuter traffic anomalies in the San Francisco Bay Area. The system supports novel interaction between a prediction algorithm and a human analyst, and is designed to explore the boundaries, limitations and synergies of both. The degree of explanation of underlying data and algorithmic process was varied experimentally across each interface. The experiment (N=197) was performed to assess the impact of algorithm transparency/explanation on data analysis tasks in terms of search success, general insight into the underlying data set and user experience. Results show that 1) presence of recommendations in the user interface produced a significant improvement in recall of anomalies, 2) participants were able to detect anomalies in the data that were missed by the algorithm, 3) participants who used the prediction algorithm performed significantly better when estimating quantities in the data, and 4) participants in the most explanatory condition were the least biased by the algorithm's predictions when estimating quantities.

Algorithmic Anxiety and Coping Strategies of Airbnb Hosts

“…Researchers have also studied the effects of explanation of specific algorithms in other contexts such as recommendation systems [10,40,44], ranking [13], scoring [26] and personalization [16]. However, research on the impact of algorithmic awareness on microentrepreneurs in sharing economy platforms is relatively scarce.…”

On Selective, Mutable and Dialogic XAI: a Review of What Users Say about Different Types of Interactive Explanations

“…In addition, 15/27 of these studies are in real-world application areas, including areas that may be considered sensitive-4 in legal and civil, 2 in healthcare, and 3 in business and finance. However, it is possible that the empirical studies included in our corpus targeted non-expert users for practical reasons, such as to solicit platform workers like those on Amazon MTurk [22,38,42,44,93,94,98,124]. Nevertheless, some of these studies are primarily aimed at making the XAI systems more transparent and more accessible to a non-expert audience [7,111,116,119,127].…”

“…We also leveraged this body of work in Section 5.3. Other context-specific comparisons were made between an interactive explanation and other explanation types [42,98,115,124], other interactive systems [89,127], other AI models [94], other interactivity types [38] or random baselines [50], among others. Some of these user-based evaluations were within-subject experiments [30,35,111].…”

“…Morrison et al [81] emphasized the usability of Compare interactive features to support human cognitive processes, finding that "comparison is much easier than classification for a person". Khurana et al [98] demonstrated qualitatively that linear interactivity was perceived as useful. Furthermore, Springer and Whittaker [111] highlight the need for progressive disclosure of model information in order to prevent users from seeing their expectations violated and distrusting the system when it is correct.…”