After much clamoring by its users, Twitter has to decided to make it easier to switch to a chronological view of your tweets. Originally launched this way, the change comes just a few years after Twitter created an algorithm that identifies and provides the user’s ‘most relevant’ tweets first. The change created quite an uproar when it first happened and has been the bane of many users since. Many feel that the chronology is what gave Twitter its benefit and have been demanding this reversion. This Verge.com article from 2016 quotes a user talking about the analytic-based feed:
“It tears conversations apart, and it’s really confusing when some people have been live-tweeting an event and those things get scattered all across my timeline. It makes it extremely hard to follow events, and destroys one of the core values of Twitter, in my opinion.”
It’s difficult to argue with this. I’ve definitely felt the challenge of following a breaking news story in real time since the change has been made.
But Twitter created the relevance algorithm for a reason. They have a major problem that they need to address. Basically, if you follow a decent number of users, how can you expect to find anything relevant among all the content that’s posted? It’s a signal to noise ratio problem. There is some amount of signal buried in a lot of noise. I have been bothered by this problem for some time, even though I only follow a bit more than 300 people (at the time of writing this). It has made me very judicious about who I follow.
This leads to a data triage challenge and it’s not unique to Twitter. Almost every popular technology company has this issue. Amazon has millions of products for sale. How am I going to browse and find something to purchase? Netflix has thousands of TV programs and movies available. How can I find something interesting to watch? Facebook and Instagram feeds are similarly clogged with posts. The more data that is collected for a system, the more difficult it is to find the relevant information in that data.
When companies have products that provide users with large amounts of data, they use standard approaches to help users manage that data, including filters, search, pagination, and Twitter’s choice, algorithm-defined organization. *Twitter also has some forms of filter and search, but the focus here will be on the algorithm.
With an algorithm, the company finds some way to calculate what any particular user might find interesting. Companies use these algorithms because they do show some benefit. They help drive conversion or engagement. But they are not complete solutions, for a few reasons.
First, these analytics help remove some noise (actually organize it) but don’t necessarily make the problem any better. There is no guarantee that signal is not also thrown away. There is also no guarantee that enough noise will be thrown away that a viable signal can be found. And most importantly, there is no way for users to know if the analytic has made the problem better or worse for them. They don’t provide users with any indication about why the tweets are defined as relevant and others as less relevant.
Consider Netflix. After watching Netflix pretty regularly for the past couple of years, it’s clear that their analytics have no idea why I like the shows that I do (I’m pretty sure I don’t either, so good luck to them). Together, my wife and I have used Netflix to watch countless programs across a variety of categories (Dramas, comedy specials, documentaries). I could probably count on one hand the number of successful recommendations they’ve made. I rarely see a recommendation that looks worthy of my time, especially one that I’ve not already considered without their help. These recommendations have not significantly changed my experience with Netflix (your mileage may vary of course) and I’ve largely stopped browsing for new content.
Similarly, I basically spend no time at Amazon unless I know exactly what I want, because browsing is so difficult. Their recommendation engine is fairly awful, and offers alternatives to things I already purchased. When there are enormous amounts of data, the traditional approaches don’t do a great job of strengthening the signal to noise ratio.
The second problem with the algorithm is that any success is a form of self-fulfilling prophecy. Let’s consider Twitter’s timeline. They use their algorithm to prioritize tweets for you. These tweets get the most exposure, which mean they get more attention. This increases the likelihood that people will engage with the shown tweet and greatly reduces the likelihood that something not shown will be engaged with. This makes it more likely that the algorithm is reinforced. This creates a cycle that’s difficult to break.
While there are ways to mitigate the problem (broadening checks), they still never overcome the self-fulfilling prophecy. Amazon’s recommendations have the same issues. Amazon suggests other things that you might want with your purchase. If you purchase them, then the algorithm was right. If you don’t make a purchase, it’s unfair to say that the algorithm was wrong. Maybe it was a good match and you already had it, or maybe it was a good match and you can’t afford it. The only information that provides good feedback for the algorithm is the purchase. This exacerbates the self-fulfilling nature of the algorithm. It makes it look like the algorithm is working, but it is really impossible to tell whether it is or not.
Now look at the feedback Twitter gets. They can clearly get engagement metrics: how many tweets people retweet, quote, or like per actual tweet viewed. They can also see how many tweeted links are clicked. They cannot (I assume) measure how many tweets make me laugh or pause and think, but do not make me interact with their tool.
A related issues is that while Twitter’s relevance algorithm may be a net benefit to users, it is impossible for users to grasp that value. It’s calculations are a black box. Users have no idea why the algorithm is giving them the tweets they see. And they have no knowledge about what they aren’t seeing. This amplifies the problems the users have with the analytic and make them clamor for something that they at least understand — chronological order.
Given the signal to noise ratio that exists in the Twittersphere, a more nuanced solution is necessary. Twitter should not be forcing one particular view on my timeline.
The Outline of a Solution
I would love to provide an answer here about what those perspectives should be, but Twitter isn’t paying me to solve this problem. I know that the answer isn’t easy, but I have helped to solve similarly difficult problems for customers in the past.
A good system will allow its users to explore the space, and decide for themselves what is relevant. It will present all the data (the Tweets) in the frames of reference that helps users decide what is potentially valuable. Time may be relevant to tell a story, but Twitter is also filled with non-sequiters. Trending topics are important, except for when I don’t care about them.
A good system will allow its users to explore the space, and decide for themselves what is relevant.
At a broad level, the goal should be to utilize everyone’s personal context to guide discovery in their Twitterverse, wherever that path may lead. This does not mean user designed, but a space that shows relevant data to the user to let them choose their own path. Show users the boundaries of the space and give them free reign to operate within that space and tools to help them explore.
What the Twitter product team should be asking is, in what ways will people best manage the tweets that they receive? There are likely to be multiple answers. There are some people whose tweets I would probably always want to see — my brothers for example. There are others for which the hit ratio for me is much lower. Twitter should help me organize these, so that I can see when my brothers tweet (infrequently), but still see what interesting things others I follow are tweeting about. And no, Lists isn’t this. I can’t read every single tweet, so Twitter needs to provide a high-level overview of the Tweet space using multiple perspectives.
Part of the answer would be for Twitter to distill what I missed since I last checked in. The first thing I see when I open Twitter might not be a list of tweets, but an overview of my Twitterverse. This isn’t just a summary (counts of tweets, trending topics). It goes beyond that to place the tweets in multiple frames of reference that help me choose which tweets to view. This overview should let me choose where, when, and how to enter. It allows me to switch perspectives and jump around as I see fit. I might want to prioritize my follows so that I can more easily identify when people I truly care about are tweeting. Maybe I can see when people I follow tweet about topics I care about. Importantly, none of this can be captured in a single view or representation and certainly not in a list order.
It’s not enough to find one way to organize the tweets, or find the best way to filter the tweets. There is no single solution that will be great for everyone — and finding a middling solution will only make everyone frustrated with their experience. I think this is what Twitter did with its relevance algorithm.