Created in collaboration with Ken Duthie

This project utilizes the Turicreate Python library to create an item-content similarity recommender. To train and test the recommender, a local library of 10,005 songs was used. Audio metrics and metadata information for each song was then extracted using librosa and tiny tag respectively. For a full list of features, please see below.

Features:

• duration of album on which the song appeared
• song duration
• year released
• genre
• audio offset
• bitrate
• sample rate
• tempo
• tuning
• zero crossing rate
• spectral centroid
• spectral bandwidth
• spectral rolloff
• chroma frequency

Visuals (just for fun)

• A D3 visualization was created with the concept that user's should be able to see how their recommended artists are connected.
• A simple tableau story using treemaps was created to visualize the recommendation reach artists have in total and per song. This gives us a sense of who the most recommended artists are and therefore have the most reach in the database.