Table of Contents
We, as humans, have many different activities we take part in. This is true in the physical world as well as the digital world.
For example, you may use your computer to manage email, write a blog post, create software, watch cat videos, etc. There are many things you can do with a computer.
I wish recommendation engines would account for this.
That's an abstract statement, so let's take a concrete example. At the time of this writing Youtube has a recommendation feature which spits out a grid of recommended videos when you at the end of anything you watch.
What do we have here? The recommendations presented break down into roughly 3 categories:
- Programming-related (7/12)
- Music (3/12)
- Random (2/12)
This makes some sense—I indeed use Youtube to watch videos from those three categories, and of the three programming-related seems most likely to be related to a video titled "Statistical Graphics in ClojureScript."
So what's the problem? The problem is that, taken as a whole, this is a subpar recommendation. A wall of related programming videos would have been a superior recommendation, because that's the context I'm in right now.
As a viewer I'm motivated to find videos that meet my immediate needs, be it a need for information or the need for entertainment.
Youtube, as an advertising company, is motivated by viewer engagement. I.e. keeping its users on the site for longer so it can deliver more ad impressions.
I offer a prediction: I would spend more time on Youtube if it presented more relevant recommendations.
Youtube, the business, undoubtedly has many smart people working on optimizing its recommendations to drive user engagement. I recognize I may be fully off-base here, but I think they may have reached a local maximum.
Youtube has phenomenal content, both in depth and in breadth. I'm truly amazed sometimes at the quality of free content on youtube. However, its surprisingly hard to discover through Youtube itself. The best videos I've seen have all come from human recommendations, i.e. word of mouth.
So, what's the solution? Make recommendations contextual. How can the algorithm tell what context a user is in? It can't, or at least it would be difficult. Instead of automatically "detecting" (read: wildly guessing) just let me tell you what content I'm looking for.
However, I know I cannot count on Google to allow individual users any say in personalizing their personal recommendations. As such, maybe it can be automated. In fact, recommendation context doesn't seem like it would be hard to detect. Just surface content relevant to the current video, or even the sequence of videos the user has watched this session.
It's easy to imagine that the reason Youtube does not do this already is that eventually the viewer loses interest in the topic they are viewing. At that point Youtube hopes to capture them by surfacing something unrelated yet engaging.
I have little hope that we'll see contextual recommendations on ad-supported platforms anytime soon, which is too bad. Given the amount of information on platforms like Youtube, let alone the web in general, there's a real problem to be solved by recommendations.