Artist2vec was born out of a desire to gain a more meaningful understanding of genres. As more musicians seek to identify with more than one genre, I wanted to create a system that could:
In the beginning, I started with the intent to predict new favorite artists based on previous listening history.
I started by creating an artist network spanning the most popular artists in 5 genres. I then labeled the artists I had known in the network and gathered all the neighbors of my known artists, ranking them based on how likely they were connected to known artists.
I back-tested the model against 7 years of data I have meticulously collected since I was 15. In my testing, I found some interesting results. Obviously the artists I am familiar with have grown over time, but as time passes, this model fits closer and closer to my taste.
Once I finished the core functionality, I started brainstorming ideas for how to visualize it effectively. That's when I discovered Graph Embeddings. In my case, they provide an easy way to turn each artist and their connections into a vector, using different dimensionality reduction techniques to create meaningful visual representations of the musical landscape.