Encouraging Attention and Exploration in a Hybrid Recommender System for Libraries of Unfamiliar Music

Abstract

There are few studies of user interaction with music libraries comprising solely of unfamiliar music, despite such music being represented in national music information centre collections. We aim to develop a system that encourages exploration of such a library. This study investigates the influence of 69 users’ pre-existing musical genre and feature preferences on their ongoing continuous real-time psychological affect responses during listening and the acoustic features of the music on their liking and familiarity ratings for unfamiliar art music (the collection of the Australian Music Centre) during a sequential hybrid recommender-guided interaction. We successfully mitigated the unfavorable starting conditions (no prior item ratings or participants’ item choices) by using each participant’s pre-listening music preferences, translated into acoustic features and linked to item view count from the Australian Music Centre database, to choose their seed item. We found that first item liking/familiarity ratings were on average higher than the subsequent 15 items and comparable with the maximal values at the end of listeners’ sequential responses, showing acoustic features to be useful predictors of responses. We required users to give a continuous response indication of their perception of the affect expressed as they listened to 30-second excerpts of music, with our system successfully providing either a “similar” or “dissimilar” next item, according to—and confirming—the utility of the items’ acoustic features, but chosen from the affective responses of the preceding item. We also developed predictive statistical time series analysis models of liking and familiarity, using music preferences and preceding ratings. Our analyses suggest our users were at the starting low end of the commonly observed inverted-U relationship between exposure and both liking and perceived familiarity, which were closely related. Overall, our hybrid recommender worked well under extreme conditions, with 53 unique items from 100 chosen as “seed” items, suggesting future enhancement of our approach can productively encourage exploration of libraries of unfamiliar music.

Continuous affect responses to a large diverse set of unfamiliar music: Bayesian time-series and cluster analyses.

“…But our particular interest here was whether any of the prior expressed participant preferences could predict the response mechanisms they used. In the recommender system development and analysis during which the present data set was obtained (Taylor & Dean, 2019, 2021, we indeed found that prior expressed preferences for certain musical features (such as "bass," see Method) were useful for making successful recommendations within our set of unfamiliar pieces. We also found then that diversity of taste among a set of genre preferences was effective in predicting the range of "unusualness" within our musical excerpts that an individual would find acceptable (and give a relatively high liking score).…”

“…Acoustic features were extracted using a combination of standard Max objects, for example, fft ∼ and fzero ∼, third-party externals such as zsa.descriptors (Malt & Jourdan, 2008), Alex Harker externals (Harker, 2017), and CNMAT external objects (from the University of California, Berkeley). A complete list of acoustic features, together with an explanation of their unit output is shown in Table 1 (these are described in more detail in Taylor & Dean, 2019).…”

“…Participants were those described in (Taylor & Dean, 2019). Sixty-nine first-year psychology undergraduate students were recruited via our university online participation system and received course credit for participation.…”

“…Overall, participants reported pop music as the most likeable and familiar; country music as the most disliked, and experimental music as the most unfamiliar. For more detailed information on the preexperiment questionnaire and summary music preference analyses, see Taylor & Dean (2019). Among musical features, they rated rhythm and melody as the most important, followed by bass, noise, loudness, and brightness.…”

“…This work has not previously been presented outside our institute but is based on a dataset that has been used in part in our prior publication (Taylor & Dean, 2019). This work was supported in part by a grant from the Australian Research Council to the first author.…”