In my previous post on this topic, I introduced a problem – how to understand the work that explicit genre categorisations are made to do by people uploading tracks to the SoundCloud audio-sharing website – and a potential solution – identifying the three categories most frequently used by each individual in a sample and studying regularities in the ways in which pairs of categories tend to pop up within the same group of three. I also presented some partial and preliminary findings in the form of a matrix comparing co-occurrences of the five genre categories most frequently used by people within an initial sample. And I either glossed over or left unmentioned a slew of problems, some of which we’ve been more successful in addressing than others at present (because these are only blog posts, and we haven’t finished the research yet). The biggest problem is the sample itself: the analysis was done on the basis of a snowball sample, when a random sample would be more appropriate. Hence the provisionality of all this. The analysis will be redone soon on the basis of a sample that will enable us to make more robust claims, but in the meantime I wanted to share our thought processes and working methods with the world because – quite apart from anything else – I’m excited about the patterns that are emerging.
Ah, those patterns. In part i, I pointed out the clearest one: the third most-common genre category in our sample co-occurs very rarely with the first and second, but very frequently with the eighth. That is, there are many individuals in our sample among whose top three most frequently-used genre categories are to be found both ‘house’ (the most common) and ‘techno’ (the second most common), and even more among whose top three are to be found both ‘hip-hop’ (the third most common) and ‘rap’ (the eighth most common), but very few among whose top three are to be found both ‘hip-hop’ and ‘house’ or both ‘hip-hop’ and ‘techno’. Picking further patterns out by hand in a 50×50 matrix is likely to be unreliable, and there are a number of approaches we could take to automating the process. For example, we could compare the actually-occurring frequencies with what we would expect given the same total frequency for each individual category and completely random associations between the categories. That’s the sort of approach that corpus linguists take in identifying systematic relationships between pairs of lexical items such as ‘butter’ and ‘parsnips’. And in fact, it was one of the first things we tried. But while that approach can give a relatively robust measure of the strength of the association between any two categories, it’s of relatively limited use in helping the researcher to get a sense of how larger numbers of categories might interrelate as a system.
So we tried something else. This was to visualise the whole matrix as a graph, with categories represented by nodes connected together by edges representing co-occurrences in individual users’ top threes.
Once that first step has been taken, further options become available. First, one may lay out the resulting graph using a force-directed algorithm in order to place connected nodes closer together and unconnected nodes further apart. Second, one may draw thicker and thinner lines between pairs of nodes depending on how many times the categories they represent appear together. Third, one may resize each node to reflect the frequency with which it is connected to other nodes (its weighted degree). Fourth, one may use a community detection algorithm to identify densely-connected groups of nodes within the network, and then colour-code those groups.
So if there is no-one with (say) both ‘classical’ and ‘dubstep’ among his or her top three categories, the nodes representing those categories will be unconnected and distant from one another, but if there is a small handful of people with both these categories in the top three, those nodes will be connected by a skinny line, and if there are a great many such people, they will be connected by a thick line, and probably located close together in the visualised graph. Moreover, if ‘classical’ appears with many other categories but ‘dubstep’ does not, then the node representing the ‘classical’ category will be larger than the node representing the ‘dubstep’ genre. And if ‘classical’, ‘dubstep’, and ‘grindcore’ are all densely connected to one another through co-occurrence in many different SoundCloud users’ top threes, then they are likely to be detected as a community and given a single colour to distinguish them from other communities.
A pretty picture, but what does it tell us? First, it tells us to be cautiously optimistic about the potential of the method described above in exploring systematic relationships between genre categories on SoundCloud, because the communities it presents map onto intuitive groupings recognisable from the real world. If ‘classical’, ‘dubstep’, and ‘grindcore’ had really appeared as a community, for example, then we would have been forced to consider the possibility that our sample was extremely unrepresentative, or that there was something wrong with our chosen community detection algorithm, or even that we were barking up the wrong methodological tree by exploring genre in this way, because our cultural knowledge tells us that grindcore, classical music, and dubstep have little or nothing to do with one another and are very unlikely to be produced by the same people. What we see in the above is, however, a community dominated by musical genres with a historical association with black performers – for want of a better word, let’s call it the ‘urban’ community – and a community dominated by forms of electronic music designed for dancing to in clubs or at raves – for want of a better word, let’s call it the ‘EDM’ community – as well as a somewhat loosely-connected ‘community’ of… everything that is neither urban nor EDM, including guitar music (e.g. ‘rock’), un-danceable electronic music (e.g. ‘ambient’), and acoustic music (e.g. ‘classical’). (There are some interesting anomalies, but most of these are easily explained. For example, ‘instrumental’ is probably found within the ‘urban’ community because of the role it appears to play – at least within our sample – in setting hip-hop apart from EDM in the absence of its most obvious distinguishing feature, i.e. the rapped vocal.)
This doesn’t really qualify as a discovery, because – in common with record company executives everywhere – we already knew that urban music and EDM existed as socially real categories. And it doesn’t quite answer the problem with which I began part i, i.e. of how to distinguish electronic music from everything else, since electronic music extends beyond the EDM community in the graph above (as already noted, ambient is a form of electronic music even if it isn’t danceable; moreover, some of the supposedly ‘urban’ genres above, such as jungle, drum & bass, and dubstep, are both electronic and extremely danceable). But it suggests an approach to the study of genres as emic categories, and to studying them quantitatively as well as qualitatively.
So the key question now (apart from will these patterns recur in a random sample?) is of how to use groupings such as the above to inform other work on this project. For example, we could take a network of ‘follow’ relationships (such as the one explored in Anna’s first post on quantitative data collection) and partition it according to whether the individuals concerned primarily gave the tracks they uploaded genre categories from the ‘urban’ community or from the ‘EDM’ community. We’ve already done something like this with regard to geographical location, and found a tendency for producers identifying as based within the same city to follow one another, hinting at the continued importance of local music scenes in fostering appreciation of value, even where such appreciation is expressed via the internet. The challenge will be to approach an understanding of how genre and location interact in structuring music makers’ recognitions of value in one another’s work. Is an uploader of trance in Leeds more likely to ‘follow’ an uploader of house in Ibiza or an uploader of RnB in his or her hometown, for example? And why (not)?
Right now, we simply don’t know. Which is a good place to be.