Awesome Music Analysis

Spotify adds a bunch of music statistics to each song. I’ll be using these statistics to find out how different our music tastes are, and where they are different (if at all…, we do share a playlist after all). I’ll be looking mainly at the following features:

Finding Awesome Clusters

So in which musical clusters can we divide our Awesome Mixtape #1 playlist? To answer this question I will be using the K-Means clustering method. The basic idea behind k-means clustering consists of defining clusters so that the total within-cluster variation is minimized.

The drawback is that we have to specify the number of clusters we want the data to be devided into. In order to determine the optimal number of k I emply both the elbow method and silhouette method as seen below. Although the silhouette method suggests k = 2 as the optimal number, this will not provide me with much more detail. Therefor I’ll guide my choice based on the elbow method, where at k = 5 the line starts bending somewhat.

# Centering and Scaling is necessary for the k-means to work properly
cluster_df <- mixtape %>% 
    select_if(.predicate = is.numeric) %>%
    map_df(scale)

set.seed(123)

gridExtra::grid.arrange(
fviz_nbclust(cluster_df, kmeans, method = "wss") + 
    theme(plot.title = element_text(hjust = 0.5),
          plot.subtitle = element_text(hjust =0.5)) +
    labs(subtitle = "Elbow Method"),
fviz_nbclust(cluster_df, kmeans, method = "silhouette") +
    theme(plot.title = element_text(hjust = 0.5),
          plot.subtitle = element_text(hjust =0.5))+
    labs(subtitle = "Silhouette Method"),
ncol = 2)

Creating final clusters with k = 5.

final_clusters <- kmeans(cluster_df, 5, nstart = 25)

mixtape_clusters <- mixtape %>% 
    mutate(cluster = final_clusters$cluster) %>%
    select(cluster, artist_name, track_name, danceability,
           energy, valence, loudness, speechiness,
           acousticness, instrumentalness,
           liveness, tempo, duration_ms)

Now that everything is done, we can start looking at the clusters, and see what kind of distinction the algorithm made.

mixtape_clusters %>% 
    group_by(cluster) %>%
    select(-artist_name, -track_name) %>%
    summarize_all(mean) %>%
    mutate_all(round, 3) %>%
    kable(format = "html") %>%
    kable_styling(bootstrap_options = c("hover", "striped", 
                                        "responsive", "condensed"), 
                  full_width = T,
                  position = "left") %>%
    scroll_box(width = "100%")

Although the distinction is hard to tell in this way, I see the following patterns:

1. Cluster 1 - Fun, happy, dancable and energetic songs.
2. Cluster 2 - Angry, up-tempo songs
3. Cluster 3 - Instrumental, acoustic songs
4. Cluster 4 - High tempo and energetic instrumental songs
5. Cluster 5 - Far more likely to be live performance songs

Let’s try to see if this fits with some songs for each cluster!

Awesome Conclusion

As a first impression, the songs fit the descriptions I made quite well. I do think that I put too much emphasis on energy. For example cluster 2 does have quite negatively loaded songs, but aren’t necessarily energetic songs. Cluster 3 seems to fit the acousticness value quite well. The addition of Xavier Rudd, Dermot Kennedy and Luke Sital-Singh confirm the slow acoustic nature of this cluster. In Cluster 5 we see only one live performed song, yet still one more than in the other clusters. So perhaps if we had a bigger sample of songs we’d see more added live songs in this cluster.

In the end, I feel like the clusters made some nice distinctions. Perhaps in the future I could add more metadata or even sentiment of the songs by analyzing the lyrics of a song with the Genius API.