Version française
Music recommendation is influenced by several factors that make research in this area particularly challenging. Schedl et al. (2018) list, for example, the short consumption time of music items (typically songs), the large number of new resources created each year, the importance of the order in which these items are consumed, the possibility of recommending items that the user already knows and wishes to listen to again, the importance of the context and purpose of music listening, etc.
This internship focuses on a very specific aspect of music recommendation: music discovery. Most users of music platforms hope to discover new music via their recommendation systems (Kamehkhosh, Bonnin and Jannach, 2020). More specifically, we will focus on two key factors of musical discovery: mental load and familiarity. Mental load arises in particular when a user listens to several unfamiliar songs in a row and reaches saturation point. He may then start listening to music he knows and enjoys, or stop listening to music at all. Music familiarity can be defined according to the concept of retention (Yonelinas, 2002), which considers that knowledge is forgotten according to an exponential decay (or other functions with a similar form), and that this decay diminishes as the number of repeated exposures increases (Rubin and Wenzel, 1996).
The goal of this internship will therefore be to develop various mathematical formalisations to estimate mental load and familiarity based on the listening history of music platform users. These formalisations will facilitate the creation of new indicators to be used as inputs for machine learning algorithms in music recommendation.
A first practical objective is to implement the entire experimental pipeline to enable the comparison of different discovery-oriented music recommenders. An initial basic version of the mental load and familiarity indicators will be developed. Subsequently, various alternatives of these indicators will be proposed, drawing more specifically from the literature in psychology and cognitive science.
Data from the Last.fm and Deezer platforms will be available to assess the relevance of the proposed explanatory variables. These data will include the listening history of tens of thousands of users over several years, as well as timestamps of the music they explicitly enjoyed.
Send a CV, cover letter and transcript of grades per e-mail to Geoffray Bonnin: bonnin [at] loria [dot] fr
Hockey, G, and J Robert. Operator functional state as a framework for the assessment of performance degradation. NATO science sub series in life and behavioural sciences, 2003.
Kamehkhosh, I, G Bonnin, and D Jannach. Effects of recommendations on the playlist creation behavior of users. User Modeling and User-Adapted Interaction, 2019.
Rubin, DC, and AE Wenzel. One hundred years of forgetting: A quantitative description of retention. Psychological review, 1996.
Schedl, M, H Zamani, CW Chen, Y Deldjoo, and M Elahi. Current challenges and visions in music recommender systems research. International Journal of Multimedia Information Retrieval, 2018.
Yonelinas, AP. The nature of recollection and familiarity: A review of 30 years of research. Journal of memory and language, 2002.