PDF Publication Title:
Text from PDF Page: 006
G Model POETIC 1239 No. of Pages 13 wide range of music genres, while the decision to restrict the results to the English language is suggested by the fact that the search term itself is in English. This initial set consists of 500 videos. Note that this is the maximum number of items that the YouTube API v. 2.0 allows one to retrieve when searching videos by keyword.5 From this list, we retained only those videos classified as ‘music’ content according to the related parameter included in the meta-data. This way we obtained a ‘seeding sample’ of 333 music videos. In the second phase, we ‘crawled’ the YouTube related videos algorithm in order to build a more consistent sample and to derive a relational dataset for the network analysis. For each video in the seeding sample, we queried the API for the 25 related videos that it allows one to collect, thus expanding our data with new videos and the links among them. After further filtering non-music content, we reiterated the last procedure on the video list we had just obtained. Therefore, we ended up with a network consisting of 22,141 nodes, linked by 70,582 edges.6 This procedure is essentially an iterative aggregation of a set of ego-networks, which recalls well-established snowball sampling techniques in classical social network analysis (Browne, 2005). Although the sample does not allow for the generalisation of quantitative findings such as the distribution of categories it is nevertheless adequate in providing an exploratory account of how music videos are matched according to the logic of the platform. This is particularly pertinent given the fact that the seeding sample of 333 music videos, which served as a basis to collect the larger dataset, has proven to be heterogeneous in terms of content according to preliminary inspection. In order to analyse the data, we first applied network analysis techniques with the purpose of exploring the relations that exist among music videos in the sample. The concept of ‘network’ is increasingly adopted in the study of popular music (e.g. Holt, 2007; Webb, 2007), and, consequently, network analysis is also employed more often in this field (e.g. Crossley, 2008). Taking inspiration from existing research that looked at community detection on social media platforms (e.g. Van Meeteren, Poorthuis, & Dugundi, 2010), we concentrated on the so-called ‘community structure’ of the network. More concretely, we applied a community detection algorithm7 in order to identify internally connected clusters of videos within the overall network.8 This allowed us to disentangle the various bottom-up associations of music content, which are in part related to the spontaneous behaviour of music consumers. Next, we conducted a content analysis of video titles in order to identify and code the most recurrent words in each cluster (see Krippendorff, 2013) to interpret different cluster associations from a qualitative perspective. In practical terms, we classified the clusters according to the distribution of the observed keywords. This allowed us to explore the interplay between the platform’s affordances and users’ aggregated practices in defining the associational logic underlying the related music videos. With regards to potential ethical issues concerning user privacy in the collection of digital data (see Boyd & Crawford, 2012), we started from the assumption that the content at our disposal is public, and, moreover, focused on YouTube item- based network and metadata instead of specific users and their activities. This allowed us to avoid potential issues of privacy and ethics, since individual user data are not included in this study. 4. Results This section is organised as follows. First, we show how the videos cluster in a relatively small number of aggregations. To make sense of these clusters, we inductively describe and name each cluster by analysing the 10 words most frequently used in the titles and the most central content for each cluster. Second, we position the clusters on a bi-dimensional map on the basis of the different categories of words occurring in the video titles. These are coded according to their reference to a plurality of dimensions (all of which are defined below): ‘reception’, ‘milieu’, ‘venue/radio/label’, ‘song/album/tour’, ‘genre’, ‘artist’ and ‘cross-genre’. Finally, we present the networked associations among different clusters. 4.1. Exploring the network of related music videos The overall network is made of 22,141 videos uploaded by 9994 users. Videos are linked by 70,582 relational ties that are drawn whenever two videos are related by the algorithm (see Fig. 1). The graph is weighted on the basis of connections between pairs of videos occurring more than once. Weights between ties range from 1 to 85, signalling a redundancy of connections between videos, which gives substantial meaning to the analysis of their clustering. Although the relations generated by the algorithm are not necessarily symmetrical, we decided not to treat the graph as a directed network, since we are interested in the overall logic of aggregation (and particularly in cluster detection), rather than the specific structure of the network itself. Therefore, adding directionality to the ties would unnecessarily complicate the analysis. 5 For a discussion on the limits of YouTube Data API v. 2.0, see: https://code.google.com/p/gdata-issues/issues/detail?id=428 [Accessed: 27 December 2015]. 6 It is important to stress that by using the YouTube APIs, we did not incur in biases due to the machine’s search history on the platform that may impact on the algorithmic selection of recommended videos (Bendersky et al., 2014: 2), since data collection does not require any sort of authentication method. 7 More specifically, the Louvain method (Blondel et al., 2008). 8 There are different terms to refer to the components resulting from a community detection, like clusters, modules and communities. We decided on ‘cluster’ because the term ‘module’ has quite a different connotation, while the term community would be too ambiguous in a sociological paper. It is important to stress that this analysis has nothing to do with the popular statistical method of cluster analysis, nor with other network analysis measures such as the clustering coefficient. M. Airoldi et al. / Poetics xxx (2015) xxx–xxx 5 Please cite this article in press as: M. Airoldi, et al., Follow the algorithm: An exploratory investigation of music on YouTube, Poetics (2016), http://dx.doi.org/10.1016/j.poetic.2016.05.001PDF Image | Follow the algorithm: An exploratory investigation of music on YouTube
PDF Search Title:
Follow the algorithm: An exploratory investigation of music on YouTubeOriginal File Name Searched:
AiroldiBeraldoGandini2016Preprint.pdfDIY PDF Search: Google It | Yahoo | Bing
Cruise Ship Reviews | Luxury Resort | Jet | Yacht | and Travel Tech More Info
Cruising Review Topics and Articles More Info
Software based on Filemaker for the travel industry More Info
The Burgenstock Resort: Reviews on CruisingReview website... More Info
Resort Reviews: World Class resorts... More Info
The Riffelalp Resort: Reviews on CruisingReview website... More Info
| CONTACT TEL: 608-238-6001 Email: greg@cruisingreview.com | RSS | AMP |