Unraveling the impact of Audio Features on Spotify Popularity Score

Goals:

• Explore relationship between audio features and Spotify popularity score for tracks.
• Identify audio features that are more influential to Spotify popularity score.

A.Data Collection

Source: We used the Spotify for Developers API to gather information on Spotify music. Here are the links to the docomentation of Spotify for Developers: https://developer.spotify.com/documentation/web-api.

Process:

• Search for Playlists:
  We utilized Spotify's Search API (https://developer.spotify.com/documentation/web-api/reference/search) to search for playlists containing English songs. By setting the query to "English songs,"
  we ensured that the data would be diverse and representative of different music genres.
  We retrieved 50 playlists in total. This step is handled in code/data/get_playlists.py.

• Retrieve Playlist Items:
  Using the Get Playlist Items API (https://developer.spotify.com/documentation/web-api/reference/get-playlists-tracks),
  we extracted all the songs from these 50 playlists. This allowed us to gather basic information for each song,
  such as ID, name, release date, artists, popularity, and more. This process resulted in an initial dataset containing 5,668 tracks. This step is handled in code/data/get_tracks.py.

• Extract Audio Features:
  With the track IDs obtained from the previous step, we then used the Get Track's Audio Features API (https://developer.spotify.com/documentation/web-api/reference/get-audio-features)
  to retrieve detailed audio features for each track. These features included attributes such as loudness, energy, danceability, and more,
  providing a deeper understanding of the songs' characteristics. After dropping the duplicates and NA,
  the dataset contains data on 4,929 tracks. We saved the data into a CSV file named spotify_data.csv. This step is handled in code/data/get_spotify_data.py.

Execution method: To execute the code and get the tracks' information and audio featires (the data that you can use for further analysis), you should first get a Spotify API Client ID and Client Secret. Please follow the steps below to obtain them. Then set them as environment variables, and our code in get_spotify_data.py can help you get your access token. The steps are as follows:

• Go to the Spotify for Developers Dashboard (https://developer.spotify.com/dashboard/) and log in or create a Spotify developer account.
• Create a new application, and you will receive a Client ID and Client Secret.
• Execute the following commands in the terminal:

  git clone git@github.com:ClaireLu0608/eco395m_midterm_project.git
  cd eco395m_midterm_project
  pip install -r requirements.txt
  cd code
  cd data

• Set your Client ID and Client Secret in the .env file. Then you can run codes in get_spotify_data.py and produce your own data.

  python3 get_spotify_data.py

Results you will get: A CSV file named spotify_data.csv, here.

B.Data Overview

documentation:

1. id: The Spotify ID for the album
2. name:The name of the album
3. release date:The date the album was first released
4. artists: The artists of the album
5. duration (ms):The track length in milliseconds
6. popularity:The popularity of the track. The value will be between 0 and 100.
7. danceability:How suitable a track is for dancing based on a combination of musical elements.
8. energy:A measure from 0.0 to 1.0 and represents a perceptual measure of intensity and activity.
9. key:The key the track is in.
10. loudness:The overall loudness of a track in decibels (dB).
11. mode:Mode indicates the modality of a track, the type of scale from which its melodic content is derived.
12. speechiness:Speechiness detects the presence of spoken words in a track.
13. acousticness:A confidence measure from 0.0 to 1.0 of whether the track is acoustic.
14. instrumentalness:Predicts whether a track contains no vocals.
15. liveness:Detects the presence of an audience in the recording.
16. valence:A measure from 0.0 to 1.0 describing the musical positiveness conveyed by a track.
17. tempo:The overall estimated tempo of a track in beats per minute (BPM).

distribution:

python3 distribution.py

1. The distribution of duration(ms), danceability, energy, loudness, liveness,valence and tempo are more closely follows a normal distribution
2. Duration: The distribution of duration shows a somewhat unimodal pattern with most songs having durations between 200,000 and 400,000 milliseconds (or roughly 3 to 6 minutes). The peak occurs around 300,000 ms (5 minutes).
3. Danceability: Danceability appears to have a normal distribution, centered around 0.6. This suggests that most songs have moderate danceability, with fewer songs on the extreme ends (low or high).
4. Energy: The energy feature follows a right-skewed distribution, with a large number of songs having energy levels between 0.6 and 0.8. This suggests that many songs have relatively high energy.
5. Key: The distribution of key shows a somewhat uniform pattern, meaning songs are evenly spread across different keys, though certain keys appear more frequently than others.
6. Loudness: Loudness has a left-skewed distribution, with most songs clustered between -10 dB and 0 dB, indicating relatively high loudness for the majority of the tracks.
7. Mode: The number of songs have mode 1 is more than songs which have mode 0.
8. Speechiness: Speechiness is heavily skewed to the left, meaning most songs have very low speech content. Only a few songs show higher speechiness values.
9. Acousticness: Many songs having very low acoustic characteristics. This indicates that most songs are electronically produced.
10. Instrumentalness: Most songs include vocals and are not purely instrumental.
11. Liveness: Most songs have a low live performance aspect, though a small number of songs feature higher liveness values, possibly indicating live recordings.
12. Valence: Valence is fairly evenly distributed, with songs spread across the full range of valence from 0 to 1, though there is a slight peak around 0.5, indicating that many songs have a neutral emotional tone.
13. Tempo: The tempo distribution is multimodal, with several peaks indicating that songs tend to cluster around certain common tempo ranges (such as 60-75 bpm, 120-140 bpm).
14. Year: The distribution of the release year shows a steep rise from the 1960s onwards, with a noticeable peak around 2020. This suggests that the dataset contains a larger number of recent tracks.

C.Correlation Analysis

Correlation:

• Determine the correlation coefficients between each music feature.

  Since some of the features don't follow normal distribution,so Spearman correlation is more suitable for correlation analysis.

  python3 spearman_correlation.py

  

loudness and energy: correlation coefficient is positive high(0.7) acousticness and energy: correlation coefficient is negative high(-0.56) valence and danceability: correlation coefficient is relatively high(0.43)

• Check for Multicollinearity

  Variance Inflation Factor (VIF): A common measure used to quantify how much the variance of the estimated regression coefficients is increased due to multicollinearity.

Feature	VIF
duration (ms)	13.100347
danceability	19.692102
energy	22.023305
key	3.173974
loudness	9.762100
mode	2.279409
speechiness	2.218682
acousticness	2.702970
instrumentalness	1.095804
liveness	2.869287
valence	7.627239
tempo	17.631970

• Conclusion: duration, danceability, energy, loudness, tempo and valence have VIF>5, which means these feature may effect the durability of regression model. According to the correlation result, duration, danceability, loudness have higher correlation with popularity, so we can delete energy and valence.

D.Models and Result

Variables: From previous correlation results, we have left with duration_ms, speechiness, acousticness, instrumentalness, danceability, liveness, loudness, tempo, key, and mode as features and popularity as our y label. "Mode" is the only binary variable and all other variables are continous.

Data Cleaning: For model interpretation and to avoid overfitting, we excluded data points with a popularity score below 5. We applied three different data transformation methods: Log Transformation for skewed features, Standard Scaling for features that followed a normal distribution, and Min-Max Scaling for features without an obvious distribution pattern. The cleaned version is in a csv file here in the artifacts folder by running the following command:

python3 code/cleaning/data_cleaning.py

Models: To explore which features are more relevant and influential to the popularity score, we used three models for feature selection.

• Random Forest: We used impurity-based feature importance from the Random Forest model to identify which features contributed the most to reducing impurity when splitting the data at various nodes. A 5-fold cross-validation was conducted on the training data to select the best parameters, which were then used to fit the model and obtain the feature importance scores. The results below are displayed in descending order of importance.
• Lasso Regression: We also utilized Lasso regression for feature selection, as it shrinks some variable coefficients to 0 by adding a penalty term to the loss function. A 5-fold cross-validation was performed on the training data to select the optimal alpha from a range of values. We then obtained the coefficients from the optimal model. The results below show the selected features in color and the unselected features with coefficients of 0
• OLS Regression: Lastly, we employed OLS regression to obtain the estimated coefficients. Using the OLS regression function from the Statsmodels package allowed us to perform statistical inference, such as calculating p-values for the coefficients, which helped identify influential features. The summary of the OLS regression fitting is shown below.

Results: We concluded that Mode, Tempo, instrumentalness, and acoutisticness are the less influential features to popularity score, since at least two of the three models excluded them from the subset of important features.

Execution method:

python3 code/models/random_forest_feature_importance.py

• After running the above command in the terminal, you will get the above image results in the images folder. Random Forest and Lasso will also return resulting Mean Square Error on testing data.
• Remember to change the name of the Python file to the method you want to use.

E.Case Study

Top Songs vs. Regular Songs feature analysis

Objective: To examine how the audio features of the top 10% of most popular songs (by some ranking metric) differ from other tracks.

Key Insights:

• The comparison is based on several audio features such as danceability, tempo, liveness, instrumentalness, loudness, speechiness, key, mode, and acousticness.
• Notable differences can be observed in loudness, key, and mode, where the top 10% songs show distinct characteristics compared to regular songs.
• Features like instrumentalness and liveness seem to have a negative correlation for top songs, while others like loudness are positively skewed.
• The key and mode features are also more prevalent in the top tracks, indicating potential influence on popularity.
• This result aligns with the feature importance findings, indicating that the differences among less important features are small.

Result:

Trends in Audio Features Over Time (1980-2024)

Objective: To identify patterns and trends in the evolution of audio features in songs over time, starting from 1980.

Key Insights:

• Features such as danceability, tempo, liveness, instrumentalness, loudness, and speechiness are analyzed across different decades. Some features, like loudness and speechiness, show consistent trends over time, potentially reflecting the changing production styles in the music industry.
• There’s a lot of variance in tempo and liveness over the years, suggesting shifts in genre popularity or production techniques during different periods.
• Danceability has remained fairly consistent, though slight variations suggest it may not be as critical a factor for songs’ success over time as other features.

Result:

F.Reproducibility

• Data: You can follow the guidance in Part A. Please note that you need to have the Client ID and Client Secret first.
• Models & Case Study: From Part D instructions, you will see two command lines. One is used to get a cleaned data csv from data_cleaning.py and the other is used to run different models and case study. You are able to adjust your desired methods as needed.

  G.Limitations

  Data Limitations:

The "popularity" feature in our dataset is calculated by Spotify using their own algorithm. This algorithm have factors in the total number of plays a track has received and how recent those plays are. However, we don't know the exact details of the algorithm, and we cannot access the actual play count or the number of people who have saved a track through the API. Additionally, some Audio Features, like Danceability and Energy, are actually scores that Spotify assigns to each track. We cannot determine the exact criteria for these scores or the detailed steps used to calculate them.

Model Limitations:

Besides doing feature selection, models are unable to perfectly predict popularity scores based solely on audio features. One of the main reasons is that Spotify calculates these scores using some features that are not publicly available, making it challenging to build accurate machine learning models. This limitation also introduces bias in feature selection and leads to some variability in the results, as the models may not be able to capture enough patterns using audio features only.

Feature Selection Bias:

The analysis focuses on a predefined set of audio features (e.g., danceability, tempo, loudness), which are curated by Spotify's algorithms. The accuracy and interpretation of these features might not be consistent across all music genres or eras. Some nuances, such as musical innovations or genre crossovers, might not be captured by these specific audio features.

Noise and Variability Over Time:

The temporal analysis shows a lot of variance in certain features like tempo and liveness. This noise could be due to random fluctuations in the dataset or genre shifts that were not fully addressed in the analysis. A more refined analysis could control for genre or other contextual factors to reduce noise and improve the reliability of the findings.

H.Further Improvements

Data Collection:

Some methods can be developed to obtain the actual play counts for each track and how many people have saved them. Additionally, we can retrieve the genre tags for each track and analyze which features influence the popularity of tracks within specific genres, making the analysis more targeted.

Genre-Specific Analysis:

A genre-based comparison could shed light on whether the trends observed in the top 10% songs are genre-specific or universal. Some audio features, like instrumentalness or tempo, may have different influences depending on whether the song is pop, rock, hip-hop, electronic, etc. Splitting the data by genre and conducting genre-specific case studies could lead to more accurate insights.

Prediction and Coefficient Estimation:

If additional features, such as the total number of plays per track and recent plays, can be obtained through the Developer API or web scraping, a prediction task can be performed to develop machine learning models for predicting popularity scores. Alternatively, if all confounding variables (i.e., all covariates that influence the popularity score) are presented, unbiased estimation of the coefficients using causal inference methods can be applied to assess the actual impact of each feature on the popularity score.