Towards Explainable and Interpretable Musical Difficulty Estimation: A Parameter-efficient Approach

TL;DR

Addresses the problem of objectively estimating musical piece difficulty for education. Proposes RubricNet, a parameter-efficient white-box model that uses explainable descriptors and an ordinal optimization scheme to yield rubric-like explanations. Demonstrates on CIPI and MKD datasets that the approach achieves Acc-9 of 41.4% and MSE of $1.7$, with Acc-3 of 79.6% on MKD, outperforming several prior models. Highlights interpretable decision boundaries and descriptor contributions, and provides open-source code and interactive visualization to bridge MIR and music education.

Abstract

Estimating music piece difficulty is important for organizing educational music collections. This process could be partially automatized to facilitate the educator's role. Nevertheless, the decisions performed by prevalent deep-learning models are hardly understandable, which may impair the acceptance of such a technology in music education curricula. Our work employs explainable descriptors for difficulty estimation in symbolic music representations. Furthermore, through a novel parameter-efficient white-box model, we outperform previous efforts while delivering interpretable results. These comprehensible outcomes emulate the functionality of a rubric, a tool widely used in music education. Our approach, evaluated in piano repertoire categorized in 9 classes, achieved 41.4% accuracy independently, with a mean squared error (MSE) of 1.7, showing precise difficulty estimation. Through our baseline, we illustrate how building on top of past research can offer alternatives for music difficulty assessment which are explainable and interpretable. With this, we aim to promote a more effective communication between the Music Information Retrieval (MIR) community and the music education one.

Figures (8)

• Figure 1: To promote a more objective and transparent assessment, in our white-box model RubricNet, similarly as educational rubrics, scores (here difficulty) are dependent on descriptors' values. The Rubric Interpretability table displayed at the right is inspired by ustun2019jmlr
• Figure 2: Detailed RubricNet's architecture.
• Figure 3: Hierarchical clustering of features based on their average correlation distance within each difficulty class.
• Figure 4: Decision boundaries of the model between grades on $S_\text{agg}$ ($X$ axis) for all splits ($Y$ axis) on CIPI.
• Figure 5: Average relative contribution of descriptors ($Y$ axis) normalized between 0 and 1, across grades ($X$ axis).