[Mini ML Project] Predicting Song Likeness from Spotify Playlist

Introduction

In this article, we will be building a small machine learning project to predict whether user will like a song or not based on the songs present in his Spotify playlist. Here the dataset which will be used can be created using steps used in our previous article on Scraping Spotify data. This dataset in the CSV format consists of all audio and technical information about the tracks of a single playlist.

Also Read – Tutorial – How to use Spotipy API to scrape Spotify Data

So it’s time that we should start off with this tutorial.

Initially required libraries are loaded for performing different operations.

In the below code, a JSON file is loaded. This file contains the URI(Uniform resource indicator) of the Spotify playlists. Information about these playlists is fetched using the URI. Along with this, like attribute is used to denote whether user likes the playlist or not. These details have already been covered in the previous article, where we created a dataset. You can look at the article from here.

Loading dataset

As already mentioned, you can build as large dataset as you want. Here in this article, I will be using two CSV files, where one of them contains the list of audio tracks that I like and another CSV file contains the list of songs that I dislike. Now we will be loading these CSV files in our jupyter notebook for further processing.

Classification of tracks for identifying songs likeness

As we are looking to classify songs on the basis of user’s liking, therefore we need to classify them individually for training purposes.

For this the playlist which had like attribute as true in the playlists_like_dislike.json file, all the songs of that playlist will be classified as liked songs and when like attribute is false, then all the songs will be classified as disliked..

Using the below code, the song which user likes is denoted through ‘1’ and song which is disliked is denoted through ‘0’. The song_like attribute will be added as a column in our dataframe.

The values of song_like column will help in training and building of the machine learning model, which will be subsequently used for predicting whether a user will like the songs of a playlist.

As you would scroll sideways in both the dataframes shown above, you would find song_like column.

We can also check the number of rows and columns present in a dataframe. Since we have to operate over these two different dataframes, it is better if we concatenate these two dataframes into a single dataframe. I have used append function for this.

As it can be seen, the songs dataframe has 174 rows i.e. rows of first and second CSV files are successfully added.

Sometimes there can be a scenario where duplicate values are present in the dataframe, to remove the duplicate rows, we will be using the duplicate function.

As it can be seen, the row count has been same. Thus there are no duplicate values.

In this project where we are predicting whether user will like a song or not, so for this purpose we don’t require the information contained in columns like id, first_artist and all_artists, therefore we will drop these columns from the dataframe.

Now we will be creating another dataframe where we will drop the title column as well. This is because while building a machine learning model, we don’t want our model to encounter values that are in the form of string, thus we are dropping it.

We didn’t dropped title column in the previous dataframe because we will require the title attribute to showcase our results.

Exploring data through visualization

As a part of machine learning model building process, we must get familiar with our data. For this purpose, we will perform data exploration through visualizing various attributes present in the dataset.

Distribution Plots

With the help of distribution plots, we will analyze both the songs which are liked and disliked by the user. Here four different attributes i.e. danceability, energy, key, and loudness are depicted. It is evident from the plots that danceability and energy are almost similar for both sets of songs. In terms of keys, there are various values and songs that are liked, have more loudness as compared to songs disliked.

Similarly, various other attributes like mode, acousticness, tempo and technical audio analysis features like bars, segments and sections are visualized.

Building Machine Learning Model

We can see how these four dataframes look and the information stored in it.

Now scikit-learn library is loaded, Logistic Regression is imported for building logistic regression classifier.

Using the code shown below, model is created and fitted over the data.

Finally, the model is checked to analyze its performance with the help of a confusion matrix and classification report. These two features help in calculating the accuracy of machine learning models.

Here Precision parameter is ratio of correctly predicted positive observations to the total predicted positive observations and the Recall parameter is ratio of correctly predicted positive observations to all observations. F1 Score is the weighted average of Precision and Recall.

The weighted avg value in the f1-score determines the accuracy of the model. Therefore, the accuracy(ratio of correctly predicted observation to the total observations) of the logistic regression model is 74%.

K-Nearest Neighbors classifier is imported from scikit-learn. After this, the classifier is trained by fitting the model over the data.

As mentioned before, the weighted avg value of f1-score column will help in finding the accuracy of the model. So here the accuracy is 71%.

Thus, we will be using the logistic regression model for prediction purposes.

NOTE – You can try different machine learning models for achieving higher accuracy and for learning purposes as well.

Logistic Regression Model is used for predicting likes and dislikes of user.

Here we can see the prediction made by the Logistic Regression Model. So we have achieved our aim of this project.

Concluding the article, we learned how we can predict which songs users will like based on the playlist listed on Spotify with the help of the Machine Learning Classifier. We covered the steps for scraping data from Spotify using the Web API of Spotify i.e. Spotipy. Along with this, Data Visualization was covered in the article. Lastly, we got to know how we can evaluate the performance of a machine learning classifer with the help of confusion matrix and classification report.