Skip to main content



Using Social Networks to Predict Deaths in Game of Thrones

As discussed in a couple of blog posts from previous years, there have been attempts to map out complexity of relations and interactions in the popular series Game of Thrones into social network graphs. One previously cited example used the book series’ writing to determine which characters should be connected in the graph with different weights associated with the edges between people based on how often their names were mentioned closely to each other. In preparation for this past season, Milan Janosov, a PhD candidate at Central European University, tried modeling the social network of the Game of Thrones universe to try and predict which characters would likely die in season 7.

The book series and the television show have a significant amount of differences, so for the purposes of his network, Janosov relied on character appearances in the television show. Characters are represented with nodes and they’re connected by edges based on their appearances in the same scenes. Each scene can be represented as a complete graph and when considered together, give the connections between characters different weights or strengths. Similar to other attempts at visualizing such a network, different colors represent members of the different houses which show strong levels of connectivity among nodes of the same house.

The goal was to use the information of the network to try and train the computer to predict outcomes. A character’s importance can be looked at generally by methods we have been discussing in lecture. One way is to determine how clustered connected characters are, which relates to triadic closure, or perhaps how connected a character is. The overall goal of machine learning to determine who is likely to die is based on training the algorithm with sets of the overall social network and by knowing who has already died. The algorithm can then make predictions based on similar characteristics among nodes of those who have already died.

Now that season 7 is over, we can look see how well the predicted outcomes from the machine learning match what actually happened. I’ll try my best to keep this vague for those who don’t want direct spoilers. While the character predicted most likely to die is probably dead (I don’t think we actually saw the death on screen), many of the other high-probability deaths didn’t happen. We didn’t even see some of the characters who had over an 80% predicted chance of death whereas a pretty important player with an approximately 50-50 chance of dying was killed. Perhaps what I found surprising were the last two characters in Table 2 from their generated predictions. Given the characters’ journeys up to now and having seen them in life-threatening situations this past season, I would expect their predicted chance of death to be much higher, but these predictions, noted in the text of the article as well, actually came true.

Obviously, no predictions will be exactly accurate, so the learned algorithm can’t always make the right predictions. But perhaps there is more to the Game of Thrones universe than what can be put into a simple social network. The article itself mentions how even more identifying information such as gender or social standing could increase the accuracy, but still acknowledges how tricky the lore is to model with people being reanimated or sort of undead.

Even then, I don’t know if a social network is sufficient for this sort of analysis. The edges of the graph were based on appearances in scenes together. This assumes that showing up in the same place in front of our eyes is just as important as other things such as how long you’ve known a person. Despite being physically separated for much of the series, the Stark kids, for example, still have strong family connections. What about those using ravens to communicate with each other despite being miles apart? As we started talking about positive or negative connections in lecture, this modeled network also doesn’t take into account the animosity some characters have towards others. Just because characters show up in scenes together doesn’t tell us much about the type of interactions they have.

At the end of the day, Game of Thrones is a network television show and not an actual society, so behaviors might be different from how real life networks might work. For the sake of dramatic storytelling and with plot armor, certain characters come and go. Even if you were to include positive or negative weights among the characters, they don’t always share their true intentions. Some characters have dragons by their side. Others are on a different continent and we don’t know if we’ll see them on screen again. There’s an army of white walkers and wights coming that could change the game (speaking of which, I was surprised to not see the Night King as a character of interest in this model). All in all, it’s interesting to see how much the game can be simplified to a social network while also acknowledging how difficult it can be. Perhaps adding complexity to the network would help in predictions, but at the end of the day, the endgame has already been decided by GRRM and D&D are in charge of getting the viewers there. We’ll have to wait and see how this clearly unbalanced network will resolve itself in the final season of Game of Thrones.

 

https://cns.ceu.edu/article/2017-07-08/network-science-predicts-who-dies-next-game-thrones

Comments

Leave a Reply

Blogging Calendar

August 2017
M T W T F S S
 123456
78910111213
14151617181920
21222324252627
28293031  

Archives