Saturday Night Data Party: Hunting Down Hate

All right, it is Saturday night data party. It is uh Independence Day here in the United States of America. And wanting to continue from where we left off last week. We talked a little bit about um identifying networks of people. And wanna dig into this some more because it's something that I've been messing around with uh for the week, you know, pulling in a lot of data to analyze.

In the spirit of Independence Day, there are some things that frankly we would like to be free from. Uh one of them is uh active amounts of hate online and not like arguments, not uncivil debates, but actual racist sexist, pi uh pick your pick your poison, um people, and the potential that some of it is not even human. So we want to identify that. We want to uh be able to understand that. One of the best networks to identify this on is Twitter, uh, because it is an open system.

It's uh uh there's plenty of uh data that's available, uh and you can most importantly see the connections between different groups of people. How do you do this? How do we how do we extract this information? Well, first we have to to pick a topic that we want to know more about and identify behavior that's that isn't is not great. So we're gonna start with uh in here our uh plandemic stuff, which, if you remember from last week, uh plandemic is this bat shit crazy conspiracy theory that uh the coronavirus is some global world domination plot to I don't know.

Anyway, it's it's not even worth getting into. But lots of people tweet about it, lots of people share conspiracy theories about it and like how 5G cell towers are spreading the virus and you know, aliens and tinfoil hats or whatever. Needless to say, none of that has any credibility in the scientific community. What we want to do though is we want to dig into all of those tweets and get a sense of who those people are. So there's two there's two entities that are important when it comes to to understanding a network.

There are edges and nodes. Nodes are people, you, me, uh, anybody who's who's using Twitter, and this works by the way with things like Instagram too. Uh it's just it's a lot harder to get Instagram data, and there's no there's no good way to get it in accordance with the terms of service. Whereas if you use Twitter's API, you are in abeyance with as long as you respect you know limits and stuff, um, you are obeying the terms of service. Um there's nodes, and then there's things called edges, and edges are the paths that connect nodes.

So if I tweet at you, uh I've created an edge and uh a single degree of connection from me to you. If you then reply or you retweet me or do something else, um, that becomes sort of a two-way edge. And what we want to do is we want to take in all these tens of thousands of tweets. In fact, how many do we have here? We have we have gathered uh 59,000 tweets about this topic, and we're going to make ourselves a network graph.

So let's go ahead and run this. One of the things that uh I've had running for the last week, and it has to be done very slowly because again of Twitter's limits, is identifying using uh a neural network the probability that an account is highly automated that is that is a bot or um is you know does nothing but retweet etc uh because we want to identify that too we have heard a lot and there's some fun stuff you can go read let's just google for it about um usage of bots uh as part of these uh information campaigns by shall we say nefarious forces and we want to see get a sense of how bad that problem is and then also what do we do about it what kind what sorts of things can we do so I've got the tweets our next step and you'll recall this from last week is we need to draw the network out and there's a piece of software it's an open source piece of software called Gephy so let's go ahead and move this into the main window here go to gephi.org it's a fun piece of software to play with but it does require you to have sort of the the individual files and that's where that R code that we were just looking at comes into play because you need to have that code in order to be able to generate the types of inputs the gaffy needs. Gephy typically needs uh comma separated value files again of nodes and edges so let's go ahead and bring in our nodes first and these are again all the the people themselves and we're gonna make sure that our bot score is a floating point number here we have 42,000 unique edges and unique nodes let's go ahead and bring that into our spreadsheet here. And now we need to bring in our edges. And these are the connections.

Okay. We have 65,000 65,000 edges between 26,000 nodes. And if we look at the visualization of this, this is going to be in the beginning just a big big blob, right? So the first thing we need to do is we need to remove anybody who is just not. Hey Chip, good to see you.

Nefarious forces, yes. Nefarious forces uh lots of them. Uh we need to remove everybody who's just talking to themselves, right? Uh there's no interaction to or from them. And then we're going to do this by these degrees of connection.

If you're familiar with um Kevin Bacon, uh six degrees of Kevin Bacon, you know, so-and-so is connected to this, who's in this movie and so on and so forth. Same idea. We're going to try and and identify how interconnected the network is. So we want to get rid of everybody. You'll see there's two types of degrees.

There's in-degree, meaning a connection coming to you. So if somebody tweets at CS Pen, that would be an in-degree connection to me. If I tweet, say out to Chip Griffin, um, then that would be an out degree for me. People who have no degree in or out essentially are just talking to the air. And for the sake of this analysis, that's not going to be helpful.

So let's go ahead and just take those people out, delete them, and our network graph should be a little bit easier to see. Not a whole lot, but a little bit better. Let's rerun our connection. Now we can see we're basically up to everyone in this network being almost two degrees of connection to each other. Uh we're next going to run our importance measure.

So this measure is called eigenvector centrality measurement. And what this does is this is like if you're familiar with Google's page rank, eigenvector centrality analysis is the same general idea where if I link to your blog and you link to a bunch of other people and more people and a lot of people link to your blog, uh then the number of connections coming in is judge basically says your blog must be important because a lot of people are linking to it. Likewise, if you link out to a bunch of people, um chances are you're saying okay, I I'm sort of giving away my credibility. So eigenvector centrality is a good measure of influence in the sense of saying who does everyone else talk about. Now, for our coloration, let's rank, and we're going to rank on the probability that somebody is a that an account is a bot.

I think that's a really fun uh easy way to do that. Let's also make sure we've got our labels set here. Uh labels are uh just the names of the accounts. And let's make sure our labels are scaled to the size of the node, turn them on, and for coloration, let's apply green, yellow, red. Red being more likely to be a bot.

That's a pretty dense graph. Um let's then apply for sizing. Let's start with eigenvector centrality. Well, so this is plandemic. This is people who believe that the the coronavirus is a is a secret conspiracy theory.

And there's one node that stands out real large, and that's for the purpose of this analysis not gonna be super helpful. So let's look at the number of nodes where there's just a lot of stuff outgoing. Just stuff constantly coming from those nodes, the out degrees. Okay, we got a little bit a little bit easier to see in here, and we're gonna want to simplify this. Simplify this um and visualize it as best as we can.

So let's go ahead and set our gravity to four. And this may or may not crash the live stream because this is a very, very computationally intensive thing to do. We're going to essentially draw out this graph and get a sense of what the network looks like. You can see there's an awful lot in here that you got of yellow, orange, and red dots that indicate there's a fair number of accounts here that may be automated. And those automated accounts um could be sharing information that frankly isn't all that helpful, right?

If you're trying to promote uh useful scientific information about uh fighting a pandemic, uh having automated accounts sort of shoving this plandemic narrative might not be the best thing in the world. Okay. It's a big graph. Let's take out, to simplify this, since we're going to be focusing on the out degrees, let's also take out anybody who is not sending any information out. So we'll sort by here and just take all the out degrees out.

That will remove a few thousand more. Okay, that's a bit more manageable. Down to 17,000. But it did change the network density and made it much less dense. Okay.

We can already see that this network is starting to come together. There's a little bit of order to it. Um there are definitely some super nodes, uh, really, really big nodes. So let's do this now. We can render the pretty picture another time.

What we all want though is we want to take out what this table is of the of the highest degrees. And we're going to turn this into just a regular old fashioned spreadsheet. Let's go ahead and open up our Microsoft Excel. Alright, and again, we're using out degree as our measure. Let's sort that by largest to smallest, the the biggest, loudest nodes.

Let's take out a few, take out the top 50, right? That's a a good starting point. And put them in a file here. I'm going to turn this into just an array. Now, what we want to do is we've established that these are the four sort of 50 loudest accounts in this whole pandemic conspiracy theory.

You cannot and should not, should not be able to get rid of an account just because it is sharing information that you disagree with. But you absolutely can get rid of an account if they're spreading hate speech. That's pretty straightforward. So our next step is to see to validate are any of these accounts actually doing that. So let's put them into a new piece of code here.

Now here's the catch. Hate speech is still a little vague. But for the purposes of this exercise, what I've done is I've put together a list of essentially basic uh hate keywords, things that you shouldn't really should never call people. Now the question you know all this? Um there are some really great resources online.

The uh the ADL, for example, has a fabulous database of uh uh hate speech, hate symbols, um slang, slogans, stuff like that, uh, and a lot of it's extractable, a lot of it's downloadable, you so you can download that data and uh put it into a data set. You can also get uh information. Uh there's uh another fantastic one called the Racial Slur Database. Um they're also uh sexist and gender slur databases. When you start doing stuff like this, I will lend you this note of caution.

There is a quote, and I want to say it's by Nietzsche, that when you stare into the abyss, the abyss stares into you. If doing this sort of work, I'm gonna get gonna get this started while we talk about this. If doing this sort of work is of interest to you to identify and root out uh people using hate speech, you're gonna be exposed to a lot of it. And you're gonna be exposed to some really, really nasty stuff. The stuff that uh makes you seriously question whether humanity um is worth saving.

Uh and if you do too much of it, you will get to a point where you're like, yeah, you know what? Let's just let the human race exterminate itself. So any kind of work like this uh is work that you have to know that you can handle doing, and you have to have a good support system personally to do it. That means you know, maybe a therapist, a group of friends, uh close friends and things. Um anybody who can help you bear that burden and can keep an eye on you to make sure that you're not losing yourself in it, too.

Um one of the things that you know, in talking to law enforcement officers and and working with them you know in a martial arts capacity, um, they have to be real careful, real careful, not to become the thing they're fighting, right? Uh Chip says, I once had a briefing from NCMEC and the impact of the work they researchers and investigators do to their own mental health. Yeah. The when you stare into the abyss, the abyss stares into you. And ideas and words and things are just as infectious as viruses.

So it's something that we have to be very, very careful of, and we have to we have to have a good a solid support network. So a lot of this stuff I'll do like once a week, um, and then go hang out with friends and things you know online because pandemic, um, to make sure that I keep my own bearings and that I don't get lost in this. When this is done, what you will get is essentially that large list of users, right? And we're gonna look at it's essentially the software's gonna pull their last hundred or so tweets and just do simple bag words matching, like in the context of you of the plandemic conspiracy theory, right? There's absolutely no reason that you should ever be using the N-word, right?

There are cases where, and certainly um like black people can use the N-word. That's is their word, right? 100% theirs, they've retaken it, and they should be using it. Nobody else. Um and so, in the context of this this conspiracy theory and the audience that it's in, you can feel pretty confident that uh somebody using that particular word in this context is not using it in a positive way.

So, what we'll get is let's go into our code folder here. We get a list. We get a list of these accounts and scored by the number of different words that they've used. And so now that the next step is, and this is where it can get a little dodgy, you then have to take these and manually go check them. And when you manually check them on Twitter, what you can do is immediately just hit the report button and report it for hate speech, report it for this or that or whatever.

Again, it's not really good for your mental health in the long term, but it is something that allows you to look at these highly network people. And I work on the premise that in these networks, we know this is true for network theory, right? There are nodes that just are little and spouting hate and whatever. But there are some of these hubs of the network. If those hubs, which is the ones that we've identified, are also sharing hate speech, and we can remove them from the network by getting their accounts canceled for using blatantly hateful speech, what will happen to the network?

Will you be able to disassemble the idea or slow its spread by taking out those hubs? Now again, I want to emphasize we're not censoring somebody's ability to spread batshit crazy conspiracy theories, even though it is publicly harmful. We are absolutely censoring them for using hate speech. And so this is the experiment to see if we do this, how it goes, and what the impact of uh is on the network overall. So ideally in 30 or 60 days or whatever, uh, after Twitter goes and reviews the reports and validates that yes, in fact, this person is using the N-word 440 times an hour and is you know foaming at their mouth, and that account gets canceled.

What happens? What is what happens to the network? What happens to the idea that's being spread? That's what we want to find out. Because if we can help to slow willfully malicious ideas, not points of civility, not points of debate, not points of political philosophy, but just outright stupidity that contradicts basic scientific evidence and data, which is what I care about.

Can we help? Can we make a difference? So I would encourage you, as part of your own work, to look at ideas like this, concepts like network graphing and the use of network graphing technologies, uh, concepts like using neural networks to to score sentiment. Even even as simple as bag of words stuff to stop willful, harmful misinformation. Give that some thought.

Thanks to everyone who stopped by to watch and uh and leave commentary. I appreciate it uh you being here. If you have questions about this and follow-ups and you want to to chat about it, go over to uh our Slack instance, trustinsights.ai slash analytics for marketers. Happy to hear your perspectives and uh get your feedback on the code, the graphing technology, the ways that you can use this technology to in positive ways too, not just to remove uh intentional hate speech from the internet one little bit at a time, but maybe to also identify how can you amplify good ideas? How could you turn this around and say, okay, well, this is a hub.

Can we introduce a good idea like wear a mask to the hubs and get the hubs to boost that signal through their networks? Networks, this is the essence of social media, is the network itself. But most people who are in social media marketing never actually dabble in the networking technology part to maximize the its impact. So uh give that some thought and uh thanks for tuning in. Catch you next time.

Want help solving your company's data analytics and digital marketing problems? Visit TrustInsights.ai today and let us know how we can help you.