Election 2016 — Debate One
Structure of conversation on Twitter
At Right Relevance we provide information about influence on social media and, particularly, Twitter. We have a free service where you can discover information about topical influencers on thousands of topics. We provide an API framework to provide access to our data on influencers which we call “Relevance as a Service”. In addition, we undertake deeper consultancy projects to provide detailed analysis of topics.
We analysed 1.1m tweets from 380k users on the general topic of the Presidential Election during the 48 hours covering the first candidates debate.
Here is the overall conversation on Twitter in that period shown as a conversation map.
This shows (as you would expect) a strongly partisan split between the Trump and Clinton sides. On first inspection the map alone does not show much more information beyond that.
Compare this map with a map of the climate change conversation run with a similar analysis configuration.
You can see that in the climate change map there is a lot of rich information in the map about the kinds of different groups and topics contained within the conversation.
Another example is the Brexit conversation in the UK, where there was (like the US presidential election) a two sided public debate about the referendum to leave the EU. The choice in the case of the Brexit referendum was a choice between the status quo (remaining in the EU) and taking action to leave. Whilst, this is not quite the same as a two horse race such as a presidential election there were still two clearly defined sides to the conversation.
The conversation map for Brexit, like the climate change one, also revealed more about the conversation than the initial Presidential Debate map does although this was less pronounced than the climate change example.
As we demonstrated in our last analysis of a single day of September 11th our Insights tools can reveal more information available from the Analysis by examining the groups of users detected.
However, in this post, rather than examine the details of the stories within the conversation, we will examine the structure of the overall conversation and what that can tell us about the debate as a whole.
As discussed above the initial map does not reveal a lot about the conversation at first inspection, however, there are a few things we can observe.
- The number of prominent users (indicated by text size) is small compared to other maps like the climate change & brexit ones above.
- There is a small but noticable third community containing Joe Biden, President Obama & Bernie Sanders.
- Jerry Springer — we will come back to him.
The reason for point 1. is that the size of each user is dictated by the users Pagerank value calculated for this network. This is usually the best way to size the users as can be seen in the climate change and brexit maps above. In this case, however, the two candidates (plus Joe Biden) are so dominant that the the pagerank value is skewed and makes those with lower pagerank hard to see.
To solve this we can use a measure that we calculate based on several measurements combined, called Top Overall Rank .
This has the effect of smoothing out the wide range of values in other measures as can be seen in this table.
If we use these values to size the users names on the map it is possible to see more of the important influencers.
There is still a very large amount of ‘noise’ in the conversation. These are the small nodes which are responsible for a large volume of the connections and can be seen in the outer areas of the map. The direction of the lines is important. If you imagine a Catherine wheel spinning, if it is spinning clockwise the lines going into the user at the centre are mentions or retweets OF that user, if anti-clockwise then it is mentions or retweets BY that user. In other words any user which looks like a catherine wheel spinning anti-clockwise is just retweeting or mentioning lots of other users/tweets and not being retweeted/mentioned a lot by others.
There are several ways of filtering out this noise and looking at just the most important users.
In this case we can use the Influence ranking provided by the Right Relevance service.
The Right Relevance service provides information on the Influence of users on Twitter on a range of over 50 thousand topics.
Here is the list of topics in which Hillary Clinton is influential.
This is a list of all influencers in the topic of the Democratic Party.
By way of balance here is the same information for Donald Trump.
The Right Relevance service is not limited to the most well known figures and we track millions of ‘ordinary influencers’ in over 50 thousand topics from F1 to Opera.
This information is useful in the context of making sense of the noisy conversation about the Election debate.
Each influencer in the Right Relevance system has a score (0–100) for each topic in which they have influence. This information is calculated by machine learning from each users social activity.
Using this information we can filter the map to show just those users who have an influence of above 70/100 in any Topic.
This is what that map looks like when it is cleared of the noise which is visible around the edges of the map but contains thousands of users who are mainly retweeting others but not gaining many retweets themselves.
Note that Clinton and Trump have also been removed as they also generate a lot of connections which add no information for the reason that every tweet in the analysis is about the contest between them so showing those links adds no value.
Whilst the filters mentioned above remove users from the visualisation their influence is NOT removed. So the size of the users visible in the map is a reflection of their influence and importance within the whole population of over 1 million users, not just the other users in the filtered map. The reason for filtering out the ‘small’ and ‘unimportant’ users is that it is impossible to fit so much information into a single visualisation at a detailed level.
It is the overall effect of the activity of millions of regular users that we are measuring, as that is what counts in an election.
Topics in the Network
Here is the list of the top topics associated with users in the debate analysis which is used to filter out the users with topical influence. To be clear, we are not making any judgement about which topics are ‘worthy’ or important. The filter based on topical influence keeps users with an influence score over the threshold in ANY of the topics.
Influential Users — Differences Between the Sides.
If we look at both these views it is noticable that the clinton side seems to contain a higher proportion of influencers than the Trump side. Whereas the Trump side appears to contain more noise.
If we separate the map into two parts the additional noise in the Trump side is more noticable.
To quantify that effect, the table below shows the ratio of Influential users to All Users at the Influence score of 70/100.
There is a marked difference between the two sets of supporters with the influence score at 70/100. With Clinton supporters making up 37% of the overall Clinton supporters in the network and Trump influencers only 24%.
These differences appear to be quite significant. However, if we look at the make up of these groups the effect appears to be significantly understated by the figures above. We can illustrate this with a filtered view of the maps.
Trump Core Supporters
Firstly, looking at the core influencers within the Trump supporters. This is just the main influencers in the Trump side of the map re-organised to show them closer together.
There are two things to notice:
- There are clearly a lot of users who are known not to be Trump supporters e.g. Bill Clinton.
- The remaining users are mostly very partisan supporters of Trump.
The reason for 1. is that in a large network with hundreds of thousands of users some users attract a lot of communication from the opposing sides — enough to pull them into the other side on the visualisation.
Here is a list of the top tweets about Bill Clinton which illustrate that effect. These tweets are from users who are clearly supporting Trump and mentioning Bill Clinton to have a negative effect on the Hillary Clinton campaign. Sometimes this activity is so prevalent that the target becomes more associated with the opposite side.
The grouping sizing and layout of the users in the maps is done by machine learning algorithms which can analyse millions of interactions between users without any human input. The objective is not to definitively state which side any given user supports but, rather, to illustrate what can be learned about the overall conversation from the public conversation on Twitter.
Clinton Core Supporters
Here is the corresponding group of users in the Clinton supporters side filtered and organised in the same way as the Trump supporters above.
By contrast there are also two points to notice:
- There are much fewer (and less strong) connections between the users.
- There are many more, broadly, independent or balanced users in this group and far fewer highly partisan users.
This information can be summarised in one image as follows:
This is very reminiscent of the situation in the UK Brexit debate as illustrated in this map.
Note on Jerry Springer (and ‘Bruce’ Springsteen)
Jerry Springer & ‘Bruce’ Springsteen
Two prominent users in the non influencers are worthy of a quick inspection.
Firstly, Springsteen. It would seem surprising that Bruce Springsteen was not highly influential in some of the topics contained in the list.
As can be seen here Bruce Springsteen is indeed highly influential in a few topics.
However, the ‘Bruce Springsteen’ who shows up in the map is a fake account and therefore only has influence in Springsteen related topics which are not relevant in this context.
Jerry Spring on the other hand, is the real pearson.
The latest Tweets from Jerry Springer (@jerryspringer). Talk show host, ringmaster of civilization's end. Host of…twitter.com
However, Jerry is not a particularly big user of Twitter and only has 63k followers. He is not a highly ranked topical influencer in Right Relevance.
The reason for his influence in this network is that Jerry was responsible for a very widely retweeted tweet during the debate.
CoolLoon & Variety
You may have spotted these two users in the tables above and be wondering why they are so influential in the network.
Here is the reason.
Donald Trump tweeted mentioning both these accounts and stating that he was ahead in the Variety poll.
This tweet was subsequently deleted.
Presumably the tweet was deleted when it turned out that the Variety poll had Hillary Clinton as the winner.
The purpose of the post was to illustrate the methodology for identifying the most influential people using the public conversation on Twitter using visualisation techniques to gain an understanding of the overall picture.
In particular we have shown how information from the Right Relevance service can help identify the most important and influential people within a public debate.