Beware The Algorithm

Warning: not a blockchain post, but rather a post about social media and algorithms. But it is tangentially relevant to blockchain...

Introduction

We all know that social media platforms such as Facebook, Twitter, Instagram, and even LinkedIn have their own algorithm - the code that assesses your creative offerings and determines how widely they will be distributed across the network.

Social media platforms keep their algorithm secret, because publicizing it could result in certain users gaming the platform to spam the network.

Nevertheless, there are people out there spending significant portions of their time reverse-engineering the algorithm in order to make a living out of publishing articles and books and providing consultancy on how to get your message across to as many people as possible with as little effort.

In that sense, they're like advertising agencies. Don't use more than five hash tags in a LinkedIn post. Polls get more traction than articles. Want more advice? Sign up for a six-part course on how to maximize your social media impact!

The underlying implication is that the algorithm is a complicated blend of many input factors, a dash of AI, and some human moderation, and your input to all of these can be tweaked if only you know the right moves.

It's all a bit like trying to impress someone on a first date. Again, and again.

And underneath it all, the algorithms are indeed probably extremely complicated.

After all, they're designed and built by software engineers, most of whom like to make things that cover as many possible eventualities, especially if the use case isn't clearly defined. Just go and look at the design architecture for the Hyperledger Fabric blockchain if you don't believe me.

However, just because an algorithm is complicated, that doesn't mean a simple model can't be effective. Einstein's theory of relativity involves some fun (read: complicated) mathematics, but the much simpler Newtonian mechanics is good enough for most practical purposes.

Often most of the complexity just provides pretty froth on top of bulk of the wave, and it's the wave that we care about.

Sidebar: why am I interested in social media algorithms?

You may be wondering, "Why is that guy who walks through the woods talking about blockchain so obsessed with social media algorithms?"

The answer is that as a result of my blockchain studies, I'm also interested in automated incentivization. Many blockchains get people to spend time and money to maintain the security of the network through a reward system - the "block reward".

Social media platforms get people to spend time and money creating content (without which no one would use the platform), for free, through the lure of wide exposure and publicity.

See the parallels?

The big difference is that blockchain incentivization systems are open and transparent.

A simple life

About a month ago I started wondering if I could produce a simple formula that could reasonably capture the performance of individual LinkedIn posts, in order to predict whether they had peaked or not.

This was based on the observation that when you post a document, or text post, or video link on LinkedIn, the post goes through three stages - an initial phase (which is used to determine if your post should get some boost), the middle boost phase, and then the leveling-off phase.

Lets call them launch, thrust, and orbit.

I wanted to see if, after a few hours, I could roughly predict the final views tally for a given post, through as simple a formula as possible. Or, in other words, determine whether a post had reached its final disappointing orbit.

Without keeping you in suspense, here it is:

final views ≈ (reactions + (comments * 3.2)) * weighting * followers / 25,000

It is a linear equation, so far far simpler than some kind of machine learning algorithm with all sorts of other non-linear adjustments. You can even type it into a pocket calculator to get the answer (more on that in the penultimate paragraph of this article).

Digging in to the formula

It is helpful to look at each of the variables in the formula:

Final views

This one is simple: it's the answer we are looking for. If you were to get no more comments or reactions on your post, this is an estimate as to how many views your post will probably end up with.

I have found that about one in ten posts is an outlier - either getting unfairly killed (after all, every single post I make is pure gold), or boosted to about double the exposure of its fellow postings. In fact, I calculated that the standard deviation of my formula is about 20%, dropping to 10% if the outliers are removed.

Removing outliers - that's a brilliant trick to make your hypothesis work. I learned that in my ten years as a satellite navigation test manager. You should try it some time.

Reactions

I decided not to put in any weighting for different kinds of reactions. There is evidence that a curious ?? or a celebrate?? reaction has a different value a simple like ?? reaction, but putting that into the equation complicates things to much.

And the difference is relatively small. It's just froth.

Oh, and if you are running a poll, you can count the votes on your poll as reactions. Just add them to the reaction total. That is one of the reasons that polls do so well.

Comments

It is clear that comments are worth more than reactions. The question is, how much more?

I fiddled this one by adding up the total number of reactions I got to about 40 posts, and the number of comments, and then divided the latter by the former to come up with a weighting factor of 3.2. So a comment is worth about 3.2 times what a reaction is.

I probably should have determined if your own comments have the same influence on the final views that a follower or a a random walk-on comment has, but I'm lazy, so I didn't.

I told you this was highly scientific.

Weightings

It is well-known that some post formats do better than others. Polls get masses of traction, and attached documents do very well too. Text and image posts do okay, followed by videos and external links. And articles (such as this one) usually get extremely little exposure.

LinkedIn currently loves banal badly-worded meaningless polls, and hates well-thought-out considered articles. Why? I have no idea. Perhaps they couldn't resist the mindless race to the bottom of the social media sewer.

I no longer use LinkedIn's native video format, because one of the things I like to use LinkedIn for is to herd my following towards my Youtube channel and my podcasts. It's part of my "don't keep all your social media eggs in one platform basket" strategy. After all, it is not inconceivable that I could wake up one morning to find my account deleted, and that would mean years of shameless self-publicity work down the drain.

And I intend to keep following that strategy until a blockchain-based self-sovereign social media platform actually gets some real traction.

But back to the post weighting system: I just cross-compared the peak views for a number of average posts using different posting formats, and came up with the following comparison table, giving a multiplier factor for each post format:

This means that for every view an article gets, a text only post will probably get eleven times more, everything else being equivalent.

You can see how abysmal an article performs compared to a document or poll. And remember that polls probably count votes as reactions, boosting them even further.

Followers

It is a sad fact of life that the rich get richer, and the poor get poorer (relatively speaking).

This is as true of social media as is it of real life, and the best way to succeed in social media is to start using it before everyone else.

Unfortunately, for most of us the only way to follow this advice would be to invent a time machine.

And so the formula needs to take into account the number of followers a given content creator has. I was kindly provided with some data points by Anthony Day, who annoyingly has far more followers than me.

But he conveniently posted the same post as me at the same time (the one with the NFT primer), and I also used data from the increase in traction that came along with an increase in my follower count over the month that I gathered data for this experiment.

Which explains the followers / 25,000 part. There is a caveat here - the formula hasn't been tested at the extremes, so I expect it won't work for people who have ten followers, or ten million.

Compensating for that would require some kind of hyperbolic adjustment multiplier (sounds difficult, doesn't it?), and that would just make things too complicated.

But what about newsletter articles?

What about them? Oh, you mean the two entries in the weighting table?

Well, the first newsletter I posted bombed. It was about twice as good as a normal article, and two times almost nothing is still nothing. So when LinkedIn asked me for feedback on the newsletter feature, I gave it to them. Rather vitriolically, in hindsight.

Now I'm not saying that LinkedIn acted on my feedback, but if enough people had the same reaction, then they may have pushed the "reach for newsletter articles" up a bit.

Or perhaps the first or the second newsletter article were outliers. As this is only my third newsletter article, there really is not enough data to reach even a vague conclusion.

How to use it

Okay, so I've published this equation. So how you can crank the handle and use Microsoft Excel or LibreOffice Calc (or even just a calculator) to see if your post has peaked, and if not, what kind of exposure you will get?

Post something, and engage with all the comments that you get. And then after a few hours, look up the number of reactions, add the number of comments multiplied by 3.2, and then multiply that by the weighting value for your post type (as looked up in the table above) and your follower count. And then divide by a hundred thousand.

If the number of views you calculate is above the actual view count, you can expect your actual views to rise, and if it is below or equal to the actual view count, your post has peaked. Probably.

And then rinse and repeat an hour or two later, just to check.

Conclusion

I have to add that my high-school physics teacher would be horrified by the slap-dash back-of-an-envelope empirical approach I took to coming up with this equation. I should probably stick the MIT open source software license disclaimer in here at this point. You know, the one that says:

THE FORMULA IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED...

You have been warned. Hopefully, at the very least you found this article entertaining.

The next one will be purely about blockchain, I promise.

Update: It looks like the post has settled down, and the weighting for newsletters appears to be 30. The formula predicts 2761 views, and the actual views are 3020, which is about 10% out.