Taking inspiration from the network representation of real-life actors co-staring in movies (see Bacon number) the central question for this Tiny Use Case (TUC) was can we find patterns in the networks of co-appearing characters that are specific to Japanese media mixes (explained below). The short answer is we couldn’t, but read on to learn about the interesting things we found in the process of trying.
Theorizing the media mix
The term media mix is often referenced in the discourse around Japanese popular visual media. Its exact meaning is somewhat hard to pin down, as the concept itself originates not in theoretical texts, but rather marketing discourse and later business practices, which themselves have evolved over time. In his book Anime’s Media Mix: Franchising Toys and Characters in Japan (2012, University Of Minnesota Press), Marc Steinberg provides a genealogy of the term tracing the roots of its meaning, as it is understood today, back to Tezuka Osamu’s business model of relying on the revenue generated by character merchandising for the production of the pioneering anime series Astro Boy. However, in his view, it was the business model championed by Kadokawa Books, initially relying on the simultaneous marketing of the film-soundtrack-novel triad and then later featuring the more otaku-market focused ensemble of manga-anime-video game-light novel-etc, that helped firmly position the meaning of the term as what it is understood to refer to today. Namely:
“Since the 1980s, the term media mix has been the most widely used word to describe the phenomenon of transmedia communication, specifically, the development of a particular media franchise across multiple media types, over a particular period of time.” (Steinberg 2012: 135)
But what makes the media mix specific to Japanese content production, and how is it different from media franchising practices in, for example, North American or European markets? As Olga Kopylova argues in her dissertation, the key points that characterize Japanese media mixes are the following:
-
- “media mix projects have usually been run by a single conglomerate (Kadokawa) or by the production committee”
- “often uses adaptation as a way to spread across media”
- “continuity is by no means a priority — the story often diverges in different works, creating multiple “parallel” worlds, and sometimes not just events, but the setting changes with the change of the medium”
- “franchises may include parodies and stories set in alternate reality that invoke radical changes in style”
(Kopylova 2016: 59, my emphases)
While none of these elements are strictly unique to the Japanese context, in combination they do seem to outline a somewhat different attitude towards the development of media franchises than from what is found, for example, in US pop culture. And although the extensive world and story building that characterizes often cited examples from this latter domain, like The Matrix or Star Wars, do have their Japanese counterparts in the likes of Gundam or Dragonball, it is perhaps the frequency and emphasis of these elements vis-a-vis each other that define tendencies more pronounced in the one or the other tradition.
Unfortunately for us, irrespective of how easy or hard it might be to pin down certain trends in media franchising characteristic of Japanese popular culture, or of how it is more dependent on the reiteration of the same story – either in adaptation or in multiple divergent retellings and expansions – rather than on the progression of a larger story arc, the above points in relation to what characterizes a media mix still point to a dead end for our present examination. The reason for this is that looking at groups of connected works from the perspective of the co-appearance networks of characters in them, it makes no difference to the structure of the network whether the same group of characters appeared in multiple works together because the story was adapted or retold in alternate universes and parody versions multiple times, or if it was a string of works expanding on an ongoing story arc featuring the same cast.
In this regard our starting question would seem to require no empirical analysis to prove it to be a poorly positioned inquiry purely based on what we would expect in light of the theoretical discussion around media mix. Nevertheless, this is in part what motivated us to undertake the present analysis. Would it be possible to find something different by examining a large number of networks of co-appearing characters compared to the understanding that was reached based on well-informed expert opinions and carefully selected examples?
Data and operationalization
While the previous three TUCs (see TUC 1, TUC 2 and TUC 3) all featured data from The Visual Novel Database (VNDB) and data from the Anime Characters Database (ACDB) was used in TUC 2, the present TUC finally offered an opportunity to explore data from AnimeClick. This database not only has data on a large number of Japanese works, but also includes a number of non-Japanese ones, providing us with examples for our attempted comparison.
In order to extract the co-appearance networks of characters from the data, we started from a given work, added all characters from that work to the first network. Then for all characters we checked for further works they appear in. If such further works existed, we added all characters for those works to the network as well. This process was iteratively repeated until there were no new works to be found that any of the characters in the network also appeared in. We then moved on to the next unprocessed work in the database to start creating the next network. When two characters co-starred in multiple works this information was also recorded as the edge weight between the nodes representing said characters in their respective network.
3592 separate networks of co-appearing characters were identified in this way. It is important to note that not all works have characters listed for their entries. These works lacking character information were not considered for the present examination.
In order to analyse our results, we decided to create a rough taxonomy of the types of networks in our data. We used the NetworkX package to generate a large number of metrics for each network to be able to separate them by. Furthermore, we also employed NetworkX for visualizing our networks (all network images below were generated with the package). The network metrics were used in conjunction with the visualizations to help identify different types of networks.
A possible taxonomy of networks of co-appearing characters
The following main types of networks were identified for the purpose of our analysis:
-
- Networks with no bridges or bridge nodes
- Single node networks: 336
- Complete networks with uniform edge weights: 2490
- Complete networks with variable edge weights: 211
- Networks with bridges or potential bridge nodes
- Networks with bridges: 38
- Networks with no bridges, but with potential bridge nodes: 272
- Networks with no bridges and no potential bridge nodes: 216
- Networks with connections based on errors in the database (these were not analysed further): 29
- Networks with no bridges or bridge nodes
In network science bridges refer to edges that when removed lead to the network being separated into two independent networks (highlighted with green rectangles in the diagram below). And bridge nodes are nodes that when removed lead to the network becoming two or more separate networks (marked by a red circle in the diagram below).
Networks with no bridges or bridge nodes
Single node networks are the result of only one character being listed for a given work (or for multiple works in rare cases). Although it is possible for a work to feature only a single character, such works are quite rare, and thus the relatively high number of single node networks is a sign of the incompleteness of the data. This finding again highlights the importance of the missing or incomplete data problem found in these community compiled databases. As already discussed in TUC 2, this does not necessarily mean that the data cannot be used for research purposes. However, it is important to be aware of this problem and the limitations it imposes on both the approaches that can be used and the results that can be reached.
Complete networks with uniform edge weights are by far the most numerous making up almost 70% of all of the identified networks. These networks are called complete, as they contain all possible edges with every node being connected to every other node. Uniform edge weights means that each connection exists the same number of times, most often just once. Thus, these networks of co-appearing characters in the overwhelming majority of cases represent a single work with its characters not making an appearance in any other work. In 77 cases, however, there is more than one work listing the same set of characters. These networks with edge weights higher than one might be examples of media mixes. A typical network from this group looks like the one below.
Complete networks with variable edge weights feature a group of characters that appear in multiple works together, while other characters only appear in one or some of these works. The thicker edges on the diagram below indicate higher edge weights, thus the five characters connected by them are the ones who appear in multiple works together. These networks might also potentially represent media mixes.
Networks with bridges or potential bridge nodes
The three remaining group of networks, namely 1) networks with bridges, 2) networks with no bridges, but with potential bridge nodes and 3) networks with no bridges and no potential bridge nodes represent the configurations of co-appearing characters that do not conform to the orderly symmetric nature of the previous network types. Separating these networks into these subgroups was mainly for technical reasons as it was easier to check for errors in them this way, a process which we won’t go into the details of here. The diagram below offers a sample of what these network structures looked like.
A large number of these networks were composed of two or three works with a number of characters shared between them. A smaller but still substantial portion was made up of more complex networks of many works being interconnected in various configurations by their featured characters. These networks are potentially the most exciting, for example, featuring spin-offs (parts of the network being connected by one or two nodes, namely the character(s) featured in the spin-off, see below) or large story universes (large complex network structures, like the one in the bottom right above). Their interpretation, however, required manually checking their actual contents, as there was no way of knowing exactly what type of relationship (e.g. sequel, spin-off, cross-over, parody, etc.) the network configurations depicted. Thus, a sample of these more complex networks was examined in detail.
Even though our purely network based approach had reached its limit at this point in the analysis, it had served as an excellent filter. This filtering function helped us in both ordering our finds into the above described groups, and in identifying these more interesting networks that would require manual checking and interpretation.
Clues for future work
Although it would be impossible to provide any conclusive results based on only a sample of the available data, a couple of interesting points did stand out, which could be potential leads for future research along a similar approach.
First, the configuration we decided to call the “knotty star”, which features a number of clusters branching off from the central cluster of characters typically connected by one or few characters (see diagram below), seems to be more prevalent in non-Japanese franchises. Out of four such networks that were found, only one was Japanese. Even though this is a very small number for making even tentative claims, the fact that all other configuration types featured mostly Japanese networks with only a small portion representing non-Japanese works, does add some further weight to the significance of this distribution, as small as it might be.
Contentwise the “knotty star” represents a number of spin-off works featuring one or a couple of the original cast from the central work of the franchise. Considering the cited features of Japanese media mixes, it seems to make sense that this type of franchise expansion could be more prevalent in non-Japanese cases. This technique of building spin-offs around individual (or only a few) characters does not conform to the prototypical movement of expansion found in Japanese media mixes, namely via adaptations or divergent retellings of the original story.
Second, we have two potential Japanese trends, not necessarily tied to the media mix form, and which would also need to be further substantiated by future research. These two trends are the presence of celebratory and parody crossovers. Examples found of such crossovers include: the Gundam vs Hello Kitty anniversary crossover project, the Infini-T Force series, and the celebratory parody crossover project Minna Atsumare! Falcom Gakuen. Unlike in the case of the “knotty star” these celebratory and parody crossovers had no particular network configuration associated with them in our sample.
In summary, although examining networks of co-appearing characters might not be the best approach for trying to identify/verify the characteristic traits of Japanese media mixes, it definitely offers an avenue for the exploration of various configurations found in franchising and crossover practices. And although the exploration of these networks can only be partially automated, the creation and separation of various network types can help delineate the franchises that need to be examined in detail manually by the researchers.