Abstract:

Though many online sites claim to make learning "how to play piano easy”, the reality is rather different. With a few exceptions that provide in-depth feedback from *human* teachers in a paid model, most platforms offer but a simple library of songs with a fixed and manually labelled grade (”easy”, ”intermediate”, ”hard”). These systems thus (1) lack any personalization, (2) do not account for the high multi-dimensional nature of piano music complexity and (3) do not guide the students through their curriculum -- elements so crucial for motivating students in the long run.

In this work, a recommender system is presented that suggests classical piano pieces to a particular student based on his/her history of played pieces. It is trained automatically by recording the path of real student's curricula. In other words, it learns from human teachers how to make sensible recommendations. Compared to systems commonly found in e-commerce, the recommendation context in education is considerably more complex. To start, the ultimate goal of these systems is "learning" rather than sales. Second, users have very different learning speeds and modelling what a student masters and knows is harder than modelling ”taste”. Also, educational entities are not simply a ”bag of unordered items”, but a highly multi-dimensional collection of interacting resources. In our use case, specifically, there is a (non-total) order imposed on songs which mirrors the scaffolding of skills and concepts that need to be acquired in order to proceed. For this reason, standard Collaborative Filtering clearly does not suffice and a preprocessing step is needed.

The main goal of the presented Recommender System is to optimize students' motivation by keeping them in a state of "flow" where learning is optimized. Flow occurs when a person perceives the challenges and the skills brought to it, as both balanced and (slightly) above average difficulty. For example, if an exercise is too hard, a student may become anxious or frustrated which not only hampers learning but is also detrimental to motivation. If, on the other hand, the exercises that are presented are too easy, a student may become bored and lose interest. Finding this balance is one of the challenges of our tutor.

To this end, the curricula of students are analysed on-the-fly and aligned to the student's curriculum, in order to reformulate the educational recommendation as a "traditional" (e.g. movie recommendation) problem, so that the existing state-of-the-art Recommender System technology can be reused.

The recommendations made by our system have been compared to human suggestions in a blind test performed by piano teachers. Preliminary evidence suggests that the quality of suggestions was similar, and that teachers had trouble identifying which recommendations were made by a real teacher and which by a computer.

Keywords:

Recommender systems, music education, motivation, theory of flow, tutoring, adaptive learning.