Keywords: Machine Learning, Spotify, API handling, Recommendation Engine, Random Forrest

This research is a follow-up to our previous research, where we introduced working with the Spotify API, to analyse different features that are available in the API. Here we related available features to associated emotions of songs (Emotional Quadrants), and this gave us an idea how (my) musical preferences is constructed. Furthermore we presented enhanced features for future analysis.

In this research I want to take it one step further and see if machine learning can help in predicting my preference.

For the readers of the previous post, for readability and providing context there is some duplication between both posts. 

In the past already when evaluating streams of my debut album: Somewhere in Between I worked with some regression algorithms to identify opportunities for future musical releases on The Data Artists Music label. This lead to some investing insights, however scored less on predictablility. This research can be seen as my second attempt to statistically model Spotify data. Only this time with the objective to model my musical preference. In general this is the similar approach Spotify would use for instance for their 'daily' suggestions or album suggestions in the Spotify App.

For this research the following research questions were formulated

In our previous post we gave an extensive overview of what data is available in Spotify' API and how we extracted it with the use of the R package SpotifyR.

With the use of R (language and environment for statistical computing and graphics) I will be using data science technique and methods like descriptive statistics, machine learning, regression and API handling. The insights can be used to contribute to a better understanding of musical preference and the underlying constructs of the music that preference is built on. 

As mentioned the main research data consists of modelling my musical preference with the use of Spotify data. For connecting with the Spotify API we will use the SpotifyR library by Thompson, Parry and Wolf (2017) in R. A great introduction for using this library is given by Mia Smith in 'What is Speechiness' (2018). 

With the use of the SpotifyR wrapper is relatively easy to extract artist, album and song information from the Spotify API. The API documentation from Spotify is available at the Spotify for developers page. There you can also use the console section if you want to (manually) query the API. Note that here you also need to register your app to receive an access token. You will need that token for every query. Below you can find an example of the Spotify API console. Where you can manually enter search fields and the response is given back in de console.

The SpotifyR library in combination with the Spotify API allows you to go through the (meta) data in a fairly straightforward manner. The image below gives you an idea of what of the possibilities are and which functions are available in the package. The overview is directly printed as part of the package from my R console. Here you can see that there are functions to call for album characteristics, artists etc.

Below, you can find the playlists of my liked and dislikes songs.

After setting up the Spotify API link  and going through the functions of the SpotifyR wrapper, we need to have a better understanding on what data (features) are available in the API. A simplified representation of the Spotify API is given below, here we see that we can retrieve data on the following "subject areas". (As distilled from Spotify API Console).

• Albums
  
• Artists
  
• Browse
  
• Episodes
  
• Follow
  
• Library
  
• Personalisation
  
• Player
  
• Playlists
  
• Search
  
• Tracks
  
• Shows
  
• Users Profile

For this research we will be looking at combining data from the following three subject areas, as we  want to investigate individual songs, the album they below to and by which artists they were written. 

1. Albums
   
2. Artists
   
3. Tracks
   

From the image above you can clearly see that on some features there are differences between the songs I like and the songs I dislike. The features being: (1) Release Year, (2)Danceability, (3)Valence, (4) Valence, (5)Liveliness. It seem I like releases more, that are somewhat older,  prefer songs that are more lively and enjoy less joyful songs (valence). Intuitively this makes sense.

From the model results we can see that the model with mtry equal to 18 produces the highest accuracy which is 0.86 - where mtry is number of variables available for splitting at each tree node.

Now let's have a closer look at the features that are important in our model. For this model: instrumentalness is the most important feature.

Now that we know that instrumentalness is the most important feature let's see what the Spotify documentation says about this feature, to give it some context.

"Predicts whether a track contains no vocals. “Ooh” and “aah” sounds are treated as instrumental in this context. Rap or spoken word tracks are clearly “vocal”. The closer the instrumentalness value is to 1.0, the greater likelihood the track contains no vocal content. Values above 0.5 are intended to represent instrumental tracks, but confidence is higher as the value approaches 1.0." (Spotify API documentation)

Random forest model produce a value called out-of-bag estimates (OOB) that can be used as a reliable estimate of its accuracy on unseen examples. from this plot it can be seen that Liked (green) error value is relatively higher than the Disliked (red).

For the evaluation of the model we use the confusion matrix. A confusion matrix is a table that is used to analyse the performance of a classification model on a set of test data. Meaning that we check if the algorithm that we found in our training set also performs well on our test set. 

• Out of the 739 (658+81) actual Disliked the model classified 658 songs correctly
  
• Out of the 646 (104+542) actual Liked songs the model classified 542 songs  correctly 
  

The most important predictor in our algorithm is  "instrumentalness", which relates to the amount of vocals vs instrumentalness of a song. We found that the more instrumental a song is, the greater the chance will be that I like it! 

So let't put this to one last test. What do you think how I would classify these two following songs:

In this research we only used one machine learning technique, namely Random Forrest,. For future research we might be able to improve model accuracy by adding and comparing different modelling techniques and comparing results. E.g. k-Nearest Neighbour or Logistic Regression. From a feature perspective it would be interesting to add subject related features into the analysis set, such as personality based features and personal preferences.The current model only incorporates song specific features.

With this research I hope to have made you enthusiastic about adding API's to your Data Science Toolkit. And in the process have simplified machine learnning for you. For me, as somewhat a dinosaur in the field - the endless possibilities in working with vasts amount of data never seems to stop an amaze me. 

At the same time I also see a managerial implication concerning the 'more or less data is better'. discussion. In my perspective this should alway be a balanced consideration, as data science initiatives can be fruitful, costly and fail miserably. In practice to manage the 'more or less data' discussion I look at goals of data science initiatives and the alignment with innovation horizons. Where different horizons (McKinsey’s Three Horizons of Growth Model) come with different levels of levels of experimentation room.

 In my research on the bushfires in Australia - you read more on this topic. (managerial implications section).

All the best,

Ditty