back to list

Project: Explaining Spotify recommendations through narrative visualizations

Description

Main research question of project: To what extent can Spotify recommendations be explained to a satisfactory extent through narrative visualizations.

Description: Artificial intelligence (AI) is becoming a more prominent part of our every day lives. However, a concern that comes with AI is that depending on the algorithm and implementation, it is not very intuitive and often it is difficult to understand how the system arrived at a given conclusion (Hoffman, Mueller, Klein, & Litman, n.d.). Hence, the area of explainable AI is becoming a popular domain in both human technology interaction research as well as computer science. Explainable AI strives to understand and build interpretations of the processes of AI and neural networks. The focus of most of this research has been on visual processing, such as AI for facial recognition (Biran and Cotton, 2017). In this research I am going to be looking at music, Spotify recommendation more specifically, a less studied field. In HTI some papers have looked into the recommendations made by Spotify based on personal characteristics (Liang and Willemsen, 2021). I want to take this research further, especially by employing a relatively new visualization type, narrative visualization. Narrative visualizations are visualizations that tell as story and have been shown to lead to higher interpretability and understanding on the content of the visualization for users that do not have a high level of visualization literacy (Ghidini et al., 2017).

Scientific relevance of the question: Being able to understand how AI makes decisions is becoming increasing important as machines are becoming smarter. A way to understand AI is by visualizing its decision processes. This research aims to bring a new type of visualizations, narrative visualizations, to the domain of interpretable AI. Additionally, this research contributes to the domain by looking at a challenging data type to visualize, sound. Being able to effectively visualize such data type would allow for higher interpretability of systems such as Spotify and can inspire further research into visualizing nonvisual types of data.

Societal/practical relevance of the question: First relevance of this question is that it aims to aid with interpretability of music recommender systems. This can explain how personal music characteristics are used in order to make accurate music recommendation. On top of that, this research will be making use of narrative visualizations, which have been shown to be easier to interpret. Hence, through this method I will be ensuring that not only is the recommendation explained through a visualization, but also that this visualization is understandable for more users. 

Details

Student

Yana Onushkina

Supervisor

Stef van den Elzen

Secondary supervisor

MC

Martijn Willemsen / Rianne Conijn (HTI department)

Link

Thesis