We are all cohorts

It is said that every user sees something different when recommendations are displayed. Well, it’s true, but not in all cases, and not always… Do Spotify, Netflix, your favourite eCommerce, in fact, handpick every piece especially for you? Or do they just pretend?

Short historical recap

I heard somewhere that recommendation algorithms are said to be the first ones used in marketing. Is it true? I think search engines were the first that could be used; for sure, they were developed before the search engines we know today, yet eCommerce wasn’t then mature enough to embrace it, and they were rather simple tag-based tools. We date their creation to 1990/93 (it depends on how we understand the search engine).

Origins of recommendation systems are dated similarly – early 90’s. The concept was developed by Dave Goldberg and his research team at Xerox PARC, a research facility based in Palo Alto, California. The team also came up with the term ‘Collaborative Filtering’, one of the most popular recommending systems, we can apply the term Machine Learning to.

Disclaimer: Another popular recommendation algorithm is called market basket analysis. Besides, it is a very efficient tool; Affinity Analysis is more complex, so it’s fair to say that it is closer to Machine Learning. You can learn more about both of them in this article.

So which one was the first one?

Hard to tell nowadays. Another thing is that if we take a look at them closer, it turns out that search is, at some point, nothing other than a just recommendation engine.

Never mind that now, let’s focus on recommendation frames, known as the dynamic elements displaying the desired content depending on our implicit and explicit preferences.

Segment-of-one

You may have heard somewhere the quote, “our website has as many versions as customers.” It is attributed to Reed Hastings, one of the founders of Netflix, but the thing about famous quotes is that you don’t know who actually said them or if they actually were said. To be honest, I’ve never found evidence that he said something like this, but never mind that – it’s catchy… and pretty accurate.

The quote describes the idea of personalization using recommendation methods. With stationary, the exposure is always the same. When you visit when dynamic recommendations are used, each customer actually sees something different, so technically, the page is different for each visitor. Simple, right?

At edrone, we also use this trick. The functionality is called Marketing Machine, and the offer that’s generated in the frame is unique for each visitor. But does other types of “personalization methods” really display original content to every user?

Cohorts of Interests

It’s worth introducing the concept of cohorts here. What is a cohort? If you have associations with Asterix and Obelix – the Romans who reported themselves by the number of manipule, centuria, legions and cohorts – then you are right. A cohort is simply a group.

The term in the context of digital marketing appears more broadly in 2 places.

• The first of these is Cohort in Google Analytics. This is “a group of users who share a common characteristic, identified by the Analytics dimension”. However, it is not an exciting and helpful feature; we will not spend more time on it today.
• A second example, also related to Google, is FLoC, short for Federated Learning of Cohorts. This usage is closer to us, as it refers already directly to recommendations.

FLoC

Let’s start digging into this topic using Google’s example. According to papers, “an interest cohort is an interest group assigned to a user.” The cohorts should be rather large and as consistent as possible.

FLoC is meant to replace dynamic ad targeting since soon (start of 2022), cookies won’t be available for use on the third party relation.

The reason for using FLoC is strictly related to privacy concerns, and it’s not the time or place for diving in this (truly interesting) aspect of it. You can learn more from the following article.

https://blog.edrone.me/en/the-evolution-of-browser-privacy/

• The cohorts should be rather large – several thousand users.
• It should be as coherent as possible – containing very similar users.
• The same sets of ads are displayed to groups of users that share their interests.

It looks like we have the first imposter here. But let’s spare poor FLoC. We can call it a Segment-of-one tool – for sure, but nobody said that it’s genuinely “personalizing content”.

There is nothing wrong with it because the techniques of Artificial Intelligence – the hero of this text – are mainly based on illusion and sometimes cheap tricks. If something is stupid but works, it means that it is not stupid.

One-to-one

According to Cambridge dictionary:

Something that is in a one-to-one relationship with another thing strongly influences the way that the other thing changes.

Does it look similar to a famous quote? I think so. Indeed we cannot tell which algorithms are used by the most prominent companies, so at least take a look at the biggest ones.

Netflix is truly one big recommendation frame. What makes our analysis simpler, that we can easily compare it to eCommerce.

• Popular in your region – a piece of cake. Bestsellers filtered by region you are watching.
• News – Netflix’s offer does not really grow that fast, so there is a chance that it’s simply everything new and available in my language. In addition, these are the movies from across the world, so yeah, I’m pretty sure that it’s just new in the offer.
• Chosen for you. Yup. Pure Collaborative filtering (maybe enhanced with Basket Analisis).
• Proposition after watching some movie – Product-based recommendation. Pure Market Basket Analysis.

So on, and so on.

Netflix also has some exciting feature that ultimately classifies it as one big recommendation frame. Did you know that every film’s poster is also tailored for you? For example, you’re watching every movie with Henry Cavill. I bet dollars against doughnuts that you’ll see Geralt of Rivia on the poster with the recommendation if you haven’t seen Witcher yet.

Spotify

This one is rather hard to classify, being honest. I lived my life so far with the certainty that the music streaming platform tailors everything for me. For sure, certain mixtapes proposed to me are based on my playlist history, but it was until I was the playlist master at a party time ago.

Remember Netflix? Even though I share many movie interests with my friends, we constantly recommend another title to each other. Our offer is unique.

With Spotify, it’s something very different. Along with me, my friend’s sister composed a “playlist” for a party. Although we listen to similar music, we do not listen to the exact same stuff, and I assess it to a 60% match. When we compared the songs recommended for us by Spotify, what happened was that we shared the exact recommendations for our playlists. My playlist wasn’t original, as somebody “sees the same version of this app.”

Spotify puts a user into several different cohorts. This means that in a particular (often only known to Spotify) aspect, you fit into a group of other users. It creates the illusion of total personalization.

After all, we are cohorts?

At Spotify there are several times more songs than Netflix has films. It’s hard to define songs’ features as we can do with films. Thus we can treat such cohorts as the mixtapes chosen for us. What’s more, music is always about some community. Are we cohorts on Spotify? Definitely, does this recommendation strategy work for Spotify? Sure. You always have to pick the proper application of AI and tailor it to its task.

Let me write it down once again.

There is nothing wrong with such tricks and gambits. AI always mimics the real world. Does it work? For sure; And if something is stupid but works, it means that it’s not stupid.

Check out my Spotify-recommended playlist. Is anyone who is reading this article my cohort-mate?