Recommended! Wait… but why though?

Everything you always wanted to know about video recommendation services
In the era of classic media the world was simple. Everyone understood what pop culture is - it's whatever they put on big TV-channels, radio stations and newspapers. And all the rest is nothing more than niche products for the fans. As we moved into the age of social networks, an entirely new mechanism was born. Nowadays everyone is used to having their very own feed, news page and a bunch of recommended videos on the topics they follow. Putting aside philosophical question of how this brave new world influences society and culture, let's look on a more practical level at what are the logic and algorithms behind the recommended section.

Mutual benefits
First thing to understand: recommendation system can and should be mutually beneficial for content providers and the audience. Providers gather information on their preferences and acquires a tool for additional sales. Clients gets an easier navigation through the catalogue, especially when they're not sure themselves what is it exactly that they want. It's easy to forget about real personal connection with people while chasing the dry numbers of CTR and conversion, but for recommendations it's a crucial point! User must feel that this offer is not just another attempt to sell them something they don't need at all, but rather a part of a high-end service that helps them make their own choice. In fact, for platforms that offer content on subscription, there's no direct monetary benefit to recommendations at all! What they do greatly enhance though is clients loyalty and engagement by giving people incentive to spend more time on your website and keep visiting it more often.

Recommendation types
Simple selection method, as the name suggests, is the most primitive recommendation mechanic. The platform just puts together a generic offer based on titles popularity. Strictly speaking, the simplicity of implementation is the main advantage of such system. Which means it is only sufficient for platforms that build their strategy around promoting star titles (e.g., Hollywood blockbusters).
In the case of editorial method a small group of experts acts as experienced "sommeliers" and share their reviews with the audience. There are two ways to form such a group. One would be respected professional critics working for the company. The other is possible when your platform includes a social network element, allowing users to write their own reviews and "level up" their accounts by accumulating credibility between their peers. This mechanism on its own probably should be seen not as a sufficient alternative to other systems, but rather as an addition which boosts audiences involvement.

Finally, personalized method is an umbrella term for all of the more advanced mechanics of individual recommendation. The number of selection criteria is innumerable and the variety of algorithms for processing those – even more so. That being said, let's look at the most common examples.
Criteria
The most obvious parameter is, of course, genre preference. And the level of detail here is potentially infinite. Yes, you can stop at the most generic classification of comedy-drama-action-crime-period. But there's a reason why art theory defines genre as simply "a set of common tropes" – which in practice means you can build around any and every one of those. E.g., fans of sci-fi or heavy music absolutely love digging into sub-sub-substyles, explore the indie and underground scene. And if that's part of the audience you want to attract, it would be a good idea to distinguish Cyberpunk from Steampunk or Gothenburg's school of death-metal from Helsinki's.

Next popular criteria is cast and crew. Of course, you'd hardly find that many serious fans of individual cameramen or makeup artists. But even a well-known and respected producer can serve as a quality mark for part of the audience. Let alone actors and directors, who are some of the biggest celebrities of the modern era and bring just as many people to the screens as the movies themes and plots.

Country of origin means, once again, a set of typical themes and techniques that aren't necessarily codified as genre, but are easily recognized and even expected, especially between fans. Be it the singing and dancing of Bollywood, arthouse-ish Scandinavian touch or flirty and romantic French feel, the national tradition that movie creators come from inescapably influences their artistic vision. Similarly, you can look at the movie's premiere date, the decade or even specific year in cinema history.

Original language, as well as dubs and subtitles can appear to be a subsection of the previous point, but those have more technical implications. It's importance might be not only due to subjective taste, but also a deliberate aim to use content for language study or even simple inability to perceive a certain language with enough fluency (especially on the internet, where real world borders don't exist).
Also of note is the potential a lot of these criteria have for the forming of discussions and even fan communities around them. Especially on platforms that are all about adding not just functional but social elements. This might come as very beneficial, since given creative freedom (editing, reviewing, discussing) fan clubs on their own can generate content and excitement around their idols (especially actors).
The Math
There are two ways for collecting data, - direct and indirect, - depending on users activity. Direct means tracking their active choices: buying, voting, reviewing and commenting. Indirect includes all the potential signs of interest such as clickthroughs, reading descriptions or only watching the first few minutes of a video. With that, both types of data influence the algorithms, but the direct stats get priority or more statistical weight.
Working with "cold starts"
This term describes the beginning stage of data gathering, while the system still doesn't have enough info to make a good read. E.g., a newly created user account that we know nothing of yet in terms of preferences. There are two ways to go from here:

1. The platform first creates a generic recommendation based on simple selection (see above). As more information is collected, it keeps personalizing the assortment on the move.
2. The platform suggests that new users start their interaction by voting on a few titles (e.g., going through their favorite oldies). This becomes the base for a preliminary analysis. There is an unexpected mind trick though: turns out, people are prone to give answers as if presenting themselves in a certain light. So, many new users at this stage would include more respectable classics while ignoring more pop culture titles, even if the latter in reality get their attention much more often.

Probably the most mathematically advanced recommendation mechanic today is collaborative filtration off of calculating a reference group. Based on users votes (and maybe other ways of showing interest) the system creates a correlation field that tracks how similar their tastes are. For the most part this data stays hidden and is only used by the system itself. Although some platforms deliberately turn it into a social tool, showing users: "here are some like-minded people, you can talk to them". Whichever's the case, the system uses matching tastes to refine its predictions, since votes by people from the reference group are of course more relevant compared to generic rankings.
One more thing to look at is the contents "rewatchability". For example, for news this would grade very low; pretty much only the audience with professional interest would go back to those later. Something like a tutorial, on the other hand, often gets rewatched more than once. Similar logic works for the "news worth" in analytics. A fresh breakdown of a sporting or political event is on hot demand during the first days or even hours and loses relevance later, while a literary analysis of "King Lear" will be just as interesting to audience years after the initial upload. As far as specific answers to this issue, those are numerous – from adding corrective variables for news worthiness into the algorithm to creating several independent recommended sections ("Trends", "Watch again", "Liked" etc.) In any case, these factors must be addressed to avoid getting the recommended spammed with useless content and make sure the relevant one gets there in time.

As far as the way recommendations are presented, it is a task at the intersection of functionality and design. The most important point here is walking the thin line between constant visibility and subtlety. Like a good waiter who's "unseen yet always there when you need him", the recommended panel – whether it's a hard-set section on a screen, an overlay or a separate submenu – should tactfully remind you of its existence from time to time but never intrude upon your leisure.


Anatol
Chief Marketing Officer, Founder, seasoned tech manager and entrepreneur. Are you Interested in more information about our product and services? Please contact me on LinkedIn.