Diversifying Music Recommendations Houssam Nassif
[email protected] Kemal Oral Cansizlar
[email protected] Mitchell Goodman
[email protected] Amazon, Seattle, USA S. V. N. Vishwanathan
[email protected] Amazon, Palo Alto, USA, and University of California, Santa Cruz, USA Abstract It is common for music recommendations to be rendered in list form, which makes it easy for users to peruse on We compare submodular and Jaccard meth- desktop, mobile or voice command devices. Naively rank- ods to diversify Amazon Music recommenda- ing recommended songs by their personalized score results tions. Submodularity significantly improves rec- in lower user satisfaction because similar songs get recom- ommendation quality and user engagement. Un- mended in a row. Duplication leads to stale user experi- like the Jaccard method, our submodular ap- ence, and to lost opportunities for music content providers proach incorporates item relevance score within wanting to showcase their content selection breadth. This its optimization function, and produces a relevant impact is amplified on devices with limited interaction ca- and uniformly diverse set. pabilities. For example, smart phones have a limited screen real estate, and it is usually more onerous to navigate be- tween screens or even scroll down the page (see Figure1). 1. Motivation In fact, other factors besides accuracy contribute towards With the rise of digital music streaming and distribution, recommendation quality. Such factors include diversity, and with online music stores and streaming stations dom- novelty, and serendipity, which complement and often inating the industry, automatic music recommendation is contradict accuracy (Zhang et al., 2012). Since we also becoming an increasingly relevant problem.