Collaborative Recommendations for Music Session Instrumentation : Contrasting Graph to ML Based Approaches

Sammanfattning: Digital music composers are required to become proficient with relevant tools necessary for music in their particular domain. The learning curve for acquiring the skills for creative music composing, relative to the respective tooling, can be steep. The topic of recommendation systems aims to help the user getting over this threshold by filtering out irrelevant material. Many of the state-ofthe-art recommendation systems focus on metrics that are easy to measure as opposed to focusing on metrics that reflect a natural next-step when creating and listening to music. There is an exaggerated focus on smaller sets of metrics, especially accuracy metrics, and how these can be optimized. At the same time, there is support for the need of using complementary metrics, such as novelty and catalogue coverage, for more diversified recommendations. This suggests that even though the need for complementary metrics is known, it is often overlooked. The majority of these state-of-the-art approaches utilizes recommendations based either on graphs or machine learned models and little elaboration on how these approaches will effect the metrics are shown. The used toolset for the conducted experiment is composed of sessions with digital instruments, where the recommendation systems aims to give recommendations on what instrument to pick as the next step in the session. The contribution of this thesis includes how the architecture of the recommendation system can be composed in order to have a more fine grained control over the optimization of different metrics. By using scoring from a linear combination of similarity, selfexciting events and a weighted graph different metrics can dynamically be given more space. By contrasting this graph based approach to a machine learned model this thesis shows how metrics are effected by the architecture, so that recommendation systems can be built for better transparency and more user control over metric optimization.