Music to Wallet

Visualizations

• Goal is to create a dashboard so that the user could easily visualize the song attributes and its effects on the popularity score

• This was the initial step in our project as we wanted to see which attribute would have the most impact on the popularity score

• Dashboard at the bottom of the "Visualizations.ipynb" has 4 tabs

1. Popularity by Year and Genres - See the recency bias effect between year and popularity
2. Genre Attributes on Popularity - Broad 3D visualizations on popularity to attributes
3. Song Attributes on Popularity - Heat maps to show the overall trend between popularity and song attributes
4. Customizing Attributes - Allows the user to play around with visualizing the data/ grouped by popularity and year

Modeling

Preparation

• Songs created before 1960 had very low popularity score so removed from dataset

• Only columns which clearly respresnted the songs were used

• We splited data into 14 input X values to predict 1 output y value

• Used Standard scaler to scale and transfrom X

Process

• Initailized Neural Networks with 4 hidden layes

• Compiled & Fitted Model

Analyze

• Model statistic shows a p-value less that 0.05

• Plot show loss getting lesser in Trainning

• Plot show loss between Training & Testing

• Actual popularity vs Predited popularity from model is closer

Model Comparison - Classification Types

• Ran our test/train data for four classification type models:

  1. Deep Learning
  2. Gradient Boosting
  3. Decision Trees
  4. Random Forest

• Ran summary statistics to receive an accuracy score for each model

• Most accurate model was Random Forest

Model Comparison - Scaler Types

• Similarly to the step above, ran our test/train data for four Scaler Types:

  1. MinMax
  2. Standard
  3. RobustScaler
  4. QuantileTransformer

• Ran summary statistics to receive an accuracy score for each scaler

• Most accurate model was StandardScaler

Relationship between Popularity Rank and Revenue

• Collected data manually by choosing a sample of 5 songs from each popularity score category
• Spotify pay-per-stream: up to $0.00437
• Apple pay-per-stream: up to $0.00735

Data Observations

The following are some interesting observations that can only be explained by Spotify's secret algorithm for popularity scores.

• Some songs that have well-known artists but a lesser play count still seem to receive higher popularity scores. Ex: Wabash Cannonball by Johnny Cash - play count 521,601. Other songs with popularity score 40 had 1.6-2.4M plays

• Holiday Music has a much greater # of plays than other songs with the same popularity score. Ex: Let it Snow! Let it Snow! Let it Snow! had 306M plays while others with the same popularity score of 50 had around 6-11M plays

• Multiple songs with the same title (cover music, featuring other artists.) This was most prevalent with classical music and seasonal or holiday music

• Multiples of the same songs repeated in different albums/collections - many have different play counts

Predictions

• Created a revenue prediction widget to utilize our model

• Key audio features can be adjusted by sliders

• Deep learning model is executed by a button press

• The popularity score is converted to a dollar amount and displayed

• Musicians and record producers could utilize this tool to fine tune audio features and maximize streaming revenue from Spotify

  Revenue Prediction Widget