Inside the Box

• We have got a total of 2209 unique religious songs, which means there are exactly 127 doubles in the set. [1]

• The non-religious list makes up a total of 2310 unique non-religious songs, surprisingly enough there are no doubles here. [1]

• The balance without removed duplicates equals 48 songs in favour of religious music. After data processing the balance alters to -101 in favour of non-religious. The difference in balanced based on songs and actual measure lies in the fact that sometimes not as many songs are gathered from a playlist a there should be. This might be due to a specific spotifyr implementation, because sometimes the amount of songs which are in a playlist according to Spotify differs slightly from the actual amount retrieved.

A few important notes

First of all, it is hard to tell whether all non-religious songs are really non-religious (considering that it also contains songs with titles as pray and faithfully) (1). Moreover, the data might be biased with pop-music (2). Furthermore, one could count for more duplicates based on title if one is willing to argue for a comparison on completely different kinds of music instead of counting different covers of the same song. (3). In addition, the quality of the complements is debatable. To keep it objective I used mainly meta-data like the title of a playlist to keep the musical features separate and measurable. One could also choose to complement based on a few musical features and look for differences on other features. However, there are still some vague matches. Religious and Non-religious seem intuitively good complements where Buddhist Meditation Songs and Top allertijden look semantically further separated (4). Balance is measured before removing duplicates. For higher accuracy balance should be computed after knowing how many songs from each playlist are gathered (but it can only count for one playlist). For balance the following measure is used (5):

balance = songs_religious - songs_non-religious

[1] Using the distinct function from R.

Features Everywhere

Using a cluster analysis we can get some insight in how our data is ordered. It looks like there are at least two big clusters for both the individual musical features and the individual songs in the corpus. Loudness and c01 are very closely related which is not surprising because they should represent the same measurement: the first Spotify timbre feature is loudness. However we see that energy and timbre feature c02 are also closely related to those. It seems like there are some smaller groups having different properties on those features which indicates an unusual culture or religion. Some keys are also closely clustered together which indicates the use of the same keys in musical pieces with also other similar features. Maybe we can have a closer look at those ‘general’ features?

Strange Spread

These visualisations show the spread of each musical variable. De circle with cross shows where the mean lies. There are already some interesting findings visible like the difference in danceability of religious and non religious music. On average religious music is less danceable but more acoustic compared to non-religious music. Speechiness and duration are quite the same with exceptions to many outliers. On most properties Religious music has a very wide spread. Tempo is the same on average, however there is more spread in religious music.

Many Mode’s

Using spotify’s key_mode property one is able to visualize how the keys are distributed for different religions. Religious songs tend to use a bit more of C major. Jewish music is using many different keys while Christian music follows the sorting order of the whole data set more general. Buddhist and Hindu music also seem quite diverse. However the total difference in religious and non-religious music is not that big. Maybe we can get more from a more emotional view of music?

Emotions Everywhere

On the arousal-valence plane religious and non-religious music have kind of the same spread while there is a big difference for individual religions. Buddhist music tends to be quite sleepy while most Jewish music tends to be very happy. Christian and Islamic music have a wider spread like non-religious music, where the former is in the more angry corner of the plane. The different musical modes do not have much influence on the emotional placement. These results point to a research direction which focuses more on different aspects of music of different religions. For now it seems that there is indeed a difference between religious and non-religious music. However it might also be that the differences are just the result of comparing western music with non-western music. To test that hypothesis Christian and Islamic songs have to be compared more thoroughly with the non-religious data set.

Predict Type

Using Random Forest on the whole dataset with 10-fold cross-validation we get an accuracy of 0.86 together with a kappa and j-index of both 0.73, which is not bad at all.

c11 tends to be very descriptive in the difference between religious and non-religious music. Religious music has more of c11 on average compared to non-religious music. Besides, non-religious music seems to have a bit more of c06.

Also, in this particular corpus religious music seems tho consist of songs of a longer duration compared to non religious music. Further investigation would be needed to verify whether longer pieces of music seem to be of a religious kind on average in general. It might be that religious music has to say more about the fundamentals of life compared to non-religious music which might be more biased with only empirical knowledge but that would be large guessing for the moment.

Predict Religion

Again using a Random Forest classifier we obtain an accuracy of 0.79 on the whole dataset with 5-fold cross-validation. The kappa and j-index this time are 0.69 and 0.56 respectively.

Loudness, energy, the higher order timbre feature c11 and acousticness tend to be the most descriptive to discriminate between different religions. Because energy and loudness seemed to be describing somewhat the same features (as also indicated by the clustering analysis) the other three components are used to visualize the different religions.

Mostly the Buddhist, Hindu and a bit of the ‘Religious’ music tend to have a lower loudness compared to the other religions. The non-religious music scores relatively low on timbre feature c11 as we saw when predicting religious music. However it seems that it is mostly the Buddhist music scoring high on that part.

Predicting the Unknown

So what does our ‘Religious’ data set contain according to our classifier? A random forest classifier trained on all other religions predicts that most of the data is actually Buddhist, Christian or Jewish music. Where the last two sound plausible, the first one might be biased by the extreme outliers the Buddhist music has shown and therefore the algorithm is trained to get it right if it predicts sounds that are a bit unusual as Buddhist. However, on the loudness scale there are indeed ‘religious’ songs very close to Buddhist music as visible on the last tab.