LinkedIn’s feed is boring and could be improved by toning down promos and likes, and introducing an easy model (User-System-Content) for evaluating recommender systems in general. Like 40% of LinkedIn users, I’m on it every day. LinkedIn has a monopoly on public CVs which are useful for lots of reasons. And many of the articles are interesting. But the rest — epitomised by the activity feed — is dull. No wonder LinkedIn is one of the least sticky social media sites.
This dull unstickiness intrigued me — what’s stopping a company of LinkedIn’s resources loading my feed with compelling reading? I decided to try and work it out.
I analysed 100 items in my own feed and classified each one: why is it in my feed? (e.g. one of my contacts liked something); what is it about?; how interesting did I find it? (scored out of 10).
This article describes what I discovered and why the feed is no better than it is, and a way of thinking about recommender systems which might help LinkedIn improve theirs.
Of the 100 items I looked at:
- 8 were interesting enough to act on e.g. open a link, comment or share. 92% not interesting doesn’t feel like a success. So the stats bore out my hunch: LinkedIn is boring
- 49 were promotional. By promotional I include personal promotion, what someone did that day, awards (so many of these!), job adverts and LinkedIn’s own promos
- 56 of the items were there because someone liked them
Every couple of years we generate as much data as the planet has ever produced. More people, more news, more info, more knowledge, more conversation, more entertainment, more everything. So cutting through to what we need is now imperative to succeed and not just feel overwhelmed. Exactly the kind of problem to which algorithms, machine learning and AI are suited. But these technologies require crystal clear thinking at the outset to succeed. I’m going to have a go here.
User + Content + Recommender = Feed
A simple way of breaking down the problem is to consider user, content and recommender system separately. All need to be very good individually and work together to personalize the feed well. Addressing these in turn…
LinkedIn knows us. They must have the best career and work related data in the world. And now they can add Microsoft to that data set! But do they use that data well? It’s hard to know how exactly how they’re using it of course. But for example, I get told about trends in the ‘internet industry’. Yet LinkedIn knows much more precisely what I do. So why not recommend, in my case, trends in the education, learning or edtech industries?
They also know which articles I’ve looked at and liked but I’ve never seen a relevant-content-driven recommendation in my feed. They also aren’t trying to get data which is fit for the purpose of filling out my feed. They don’t explicitly ask me what kind of material I’d like to see. Of course, I don’t want them to bother me with this often. But given the amount of time I (and many of the world’s 230m knowledge workers) spend on the platform, a few choice questions would be welcome; the right data is more important than the right deep neural architecture. I think LinkedIn is punching well below its weight here.
The recommender system
Machine learning is a big part of this. (But note that it need not be — LinkedIn could just set up some very well chosen, non-adaptive rules for which items go to which users and that could, conceivably, do the job) . With all their data, the activity feed should get better with time — that’s the machine, learning.
Three observations. First, for me, out of these 100 items, I found just 8 of these sufficiently interesting to act on them. Given the content at LinkedIn’s fingertips, this seems unsatisfactory. Second, if I browse the site (e.g. someone else’s activity feed) the material seems about as relevant as that which makes it onto my feed. This is finger in the air stuff of course but the content in my feed certainly doesn’t feel very special. Thirdly, LinkedIn doesn’t make it very easy to explicitly mark irrelevant content as such. You can do it…
…but it’s two clicks away and that means it’s hidden. I’ve never done it.
LinkedIn has a lot: profiles, articles, posts, status, updates, SlideShare, LinkedIn Learning and no doubt plenty more. Some of it is brilliant and some of that will be highly relevant to me. But half (49%) of it was promotional. Promotional material is inherently dull, almost whatever’s being promoted. That’s why we skip ads on YouTube the millisecond we can! No wonder promo material scored 2.5 out of 10 for interest (vs 5.2 for the rest). Liked material is also pretty dry. It scored 3.4 / 10. Not surprising — it’s too easy to like stuff on LinkedIn and inadvertently contaminate your contact’s feeds. LinkedIn have beefed up their team of human curators to surface more of the good stuff. I’m sure that’s helped on the article side. Imposing some caps for promotional and liked content would improve the feed too in my opinion.
My LinkedIn feed sucks 92% of the time. But if LinkedIn made better use of their user data and filtered out obviously uninteresting content it would suck a lot less.
You can apply this User-System-Content analysis to Netflix, Airbnb, Tinder, Timeout, Amazon, Spotify, TripAdvisor, etc. Any recommendation platform needs to:
- Know its users
- Know its content
- Understand the underlying dynamics between them, ie the human desire / need that’s being provided for
- Build the technology around that.
We’re still a long way from general AI and further still from superintelligence. So for now and the next few years, AI needs every advantage it can get to be useful to us in real world situations. We need to provide it with those advantages and that comes from our understanding of the structure, applications, relevant examples in adjacent industries, the idiosyncrasies of the specific problem we’re trying to solve. As Bradford Cross says at this IBM event, ‘Start with the need first, don’t start with the tech first.’ It’s real business understanding that drives the technological architecture, not the other way round.
One more thing on those underlying dynamics — LinkedIn’s objectives actually seem to be pretty well aligned with their users. Most (two-thirds) of their revenue comes from recruiters and a further 25% comes from premium subscribers like me (most of the rest is from advertising). Recruiters want LinkedIn users to be on the site, keep an updated profile, respond to LinkedIn-branded communications. And Premium subscribers want a good, interesting user experience. So it really should be a priority to LinkedIn to improve the flavour of my feed.
All recommenders must and will improve their capability over time. When LinkedIn does, I and others will be able to enjoy it much more than 8% of the time.
If you enjoyed this post please do feel free to click on that little heart below and check out our blog.
Are you looking for a new training solution for your company? In January we launched a new learning recommendation engine to inject some personalization in the world of L&D. Find out more about Filtered’s learning recommendation engine here.