Prince Song Recommender

Building a song recommender using the Million Song Dataset

Telvis Calhoun
technicalelvis.com

Goals

Analyze the Million Song Dataset generated by The Echo Nest and Labrosa
Create a song recommender using metadata from the Million Song Dataset.
Create a song recommender demo using R and Shiny for songs by US recording artist 'Prince'

Getting and Cleaning the Data

The entire dataset is 280GB and Stored in Amazon S3 in HDF5 format.
We extracted some HDF5 fields to a serialized R file containing the features.

df <- readRDS("data/songs.rds")
names(df)

##  [1] "artist_name"               "title"                    
##  [3] "release"                   "song_hotttnesss"          
##  [5] "tempo"                     "loudness"                 
##  [7] "energy"                    "danceability"             
##  [9] "duration"                  "artist_familiarity"       
## [11] "artist_hotttnesss"         "artist_latitude"          
## [13] "artist_location"           "artist_longitude"         
## [15] "end_of_fade_in"            "key"                      
## [17] "key_confidence"            "song_id"                  
## [19] "start_of_fade_out"         "time_signature"           
## [21] "time_signature_confidence" "track_id"

Feature Selection

plot of chunk unnamed-chunk-3

plot of chunk unnamed-chunk-4

Recommender

We use the song_hotttnesss and loudness as numeric features to model songs in a 2-dimensional space.
A user can query the recommender by selecting values for song_hotttnesss and loudness
The recommender calculates the 5 nearest songs in the 2D space using a euclidean distance metric.

plot of chunk unnamed-chunk-5

Shiny Application: Prince Song Recommender

We built a prototype that uses 25 of the 1 million songs.
The 25 songs are recorded by US recording artist 'Prince'
The 25 Songs have valid (non-NA, non-zero) values for both song_hotttnesss and loudness
Click here to try the Shiny App that demonstrates the recommender!