The effects of controllability and explainability in a social recommender system
- PDF / 1,352,695 Bytes
- 37 Pages / 439.37 x 666.142 pts Page_size
- 37 Downloads / 152 Views
The effects of controllability and explainability in a social recommender system Chun‑Hua Tsai1 · Peter Brusilovsky2 Received: 9 September 2019 / Accepted in revised form: 25 September 2020 © Springer Nature B.V. 2020
Abstract In recent years, researchers in the field of recommender systems have explored a range of advanced interfaces to improve user interactions with recommender systems. Some of the major research ideas explored in this new area include the explainability and controllability of recommendations. Controllability enables end users to participate in the recommendation process by providing various kinds of input. Explainability focuses on making the recommendation process and the reasons behind specific recommendation more clear to the users. While each of these approaches contributes to making traditional “black-box” recommendation more attractive and acceptable to end users, little is known about how these approaches work together. In this paper, we investigate the effects of adding user control and visual explanations in a specific context of an interactive hybrid social recommender system. We present Relevance Tuner+, a hybrid recommender system that allows the users to control the fusion of multiple recommender sources while also offering explanations of both the fusion process and each of the source recommendations. We also report the results of a controlled study (N = 50) that explores the impact of controllability and explainability in this context. Keywords User interface · Hybrid recommendation · Transparency · Controllability · Explainability
1 Introduction The growth in artificial intelligence (AI) technology has advanced many humanfacing applications. A recommender system is one of the best-known examples of applying the ideas of AI to a range of real-world applications. For instance, social
* Chun‑Hua Tsai [email protected] 1
The Pennsylvania State University, State College, PA, USA
2
University of Pittsburgh, Pittsburgh, PA, USA
13
Vol.:(0123456789)
C.-H. Tsai, P. Brusilovsky
recommendations (recommendations based on explicit or implicit social connections) have been widely adopted in many social media and e-commerce platforms. Social recommendations attempt to filter out “irrelevant” information so users can reduce the efforts of decision making, such as purchasing an item online or following a new friend on social media. Social recommender systems typically use largescale data from multiple data sources and process this data with complex intelligent inference methods; however, the “reasons” of offering specific recommendations usually stay in a black box, which frequently make the resulting recommendations “puzzling” to the users (Amershi et al. 2019). Herlocker et al. (2000) demonstrated that the user generally has little understanding about the mechanism behind these systems. In this situation, processing this output could produce user behavior that can be confusing, frustrating, or even dangerous in life-changing scenarios. Moreover, a lower-transparency system is kno
Data Loading...
