Oct 17

Nat Torkington

Nat Torkington

Homophily in Social Software

The Washington Post has a brief article called "Why Everyone You Know Thinks The Same As You". In short, you hang out with people who are like you, a phenomenon known as homophily. This happens online, and indeed the Internet can lower the costs of finding people like you. But homophily raises the question for social software designers of how much they should encourage homophily and how much they want to mix it up.

Consider sites like Findory, where machine learning techniques present news to you based on news you've said you liked. It's often been asked whether this filtering just encourages people to see the news that supports their prejudices and never see news that counters them. Indeed, people were saying this about Usenet killfiles in the 80s and 90s. As social software and recommendations engines become part of the fabric of Web 2.0, the issues of homophily become important.

Designers first need to decide whether homophily is a a feature or a bug. Life is easy when you're unchallenged: this is why people read the New York Times or watch Fox News or even just watch the 5pm news (the one with the deaths taken out) instead of the 7pm (the one that's all death). Do you accept that your audience wants to be around people like them and that your job is to make that as easy as possible? NYT and Fox News show that it can certainly be a path to financial success.

If you don't buy into homophily completely, what can you do? Recommendations increase your pool of interest in very short steps. To break homophily, recommend something for reasons other than "this meshes very tightly with your profile". This seems heretical at first: the whole logic behind recommendations is to guess at items the user will probably like. But it has to happen. For you to identify their complete region of interests, you necessarily have to show them things in and out of that region. If you prematurely narrow in, you'll end up only showing them stories about melting Antarctic ice shelves without connecting to the rest of environmental, travel, or scientific stories that they're really interested in. The best way to make those connections is to mix it up.

Doing this creates serendipity: pleasantly surprising the user. For example, don't show just the top 10 most similar items in your recommendations list, but show the eight most similar and two from the mid-range. Or call the "less relevant but also likely to be interesting" results out like you're advertising them: put a heading like "Take a walk on the wild side" or "Break out" on top and act like it's a feature you're offering, not a bug you're fixing. Breaking out of the tight circle of self-similar recommendations is a feature. I tried pandora.com and listened to all the bluegrass music that was like the music I like, until Pandora had no more that I wanted to listen to. It was briefly a bit like a poker machine--I spent another fifteen minutes trying to find new music before I finally realized there was no jackpot to be held and left. Pandora never said "look, I'm out of music to recommend to you--perhaps you'd like to head off on these related jags?" Don't make your software an exhaustible pool of narrow recommendations.

Another strategy that works is to take a leaf from Malcolm Gladwell's "Tipping Point" and find the connectors--people who join homophilic clusters. This is a feature a bit like "people who liked this story also liked" but it specifically eliminates your personal history and preferences--the point is to use the current object (person/photo/music genre/news story/...) as a gateway out of your shallow meme pool. Pandora could say "the genres most liked by people who like this genre are", for example.

With social software, I think there's a lot of room to exploit profiles. Think of social software as software that connects people through activities. The activities are necessarily around some common shared interest but they function as walls around those interests. Let people build out profiles to express the full range of their interests and is start conversations about other interests. Look at MySpace profiles for examples of how keen people are shove every facet of their life into a text field, preferably with blinking orange text and accompanying 50cent soundtrack.

The methods I gave earlier for advertising interests ("Did you know ... Joshua Schachter also competitively whittles Persian cats") let you learn more about the people you already know. That's an important difference from the prototypical serendipitous recommendation: "you and X should become friends!" As Liz Goodman pointed out, we're grown-ups and have lots of friends already. What else can the software do for us besides making it even harder to keep up with all the people we know? People are conversational animals, give them more things to talk about.

Another way to build in serendipity is to have pivotal navigation: tags, top ten lists, and Flickr's interestingness measure are all ways to break people out of whatever group they're in and take them to something new. Links are at the heart of this: we've all been lost in clicking our way through a drunkard's walk of the Internet at one point or another. Inspire that in people: build those links and the metadata behind them into your site from the get-go.

Your challenge for this week: spot the social software features of a site you use that encourage homophily, and figure out two ways to break that homophily. Post your suggestions in the comments and on Friday I'll send a free book from our new releases list to what I think is the best idea (O'Reilly books only, sorry--my astounding freebie powers bounce off the kryptonite walls of Paraglyph, the Prags, and Syngress).

tags:   | comments: 12   | Sphere It

Previous  |  Next

0 TrackBacks

TrackBack URL for this entry: http://blogs.oreilly.com/cgi-bin/mt/mt-t.cgi/4985

Comments: 12

  Kyle Johnson [10.17.06 04:28 AM]

I work in student affairs IT at a university and we've been talking about just this thing. Facebook is obviously on our mind often, and I came up with a new feature for Facebook I call "The Stretch." Right now Facebook has lots of information about students in terms of what they like, their political bent, their ethnicity, etc. They also know what events students RSVP for. My idea was to develop some "likeness" score between nodes or groups of nodes on the network. The stretch would look at you, find a group of people a little different than you, and suggest events that are a little bit (but not very far) out of your comfort zone based on the people that are a little different than you. The idea is that it's easier to get someone to go to something a little different than radically different.

That's my idea.

  Ewan Gunn [10.17.06 04:33 AM]

Livejournal. Homophily sets in when you're adding friends (they're your friends already, or friends of friends, both of which mean you'll have at least something in common.)

How about instead of searching for users that have the most interests in common with you, you search for someone that has absolutely nothing in common with you what so ever, and have a browse through their journal. There will no doubt be links and articles that you'd never had looked at otherwise. Just have a peek at them.

Or (secondly) join one of the many many debating communities (hopefully you can find one that isn't all shouting screaming flame wars!) but make sure that it's one on a subject you have no knowledge of. Then you can read all views and opinions and come to a judgement yourself, and maybe participate afterwards.

I think the main thing with homophily and how to break it is to just jump. Pick one thing, completely at random (flick through a dictionary or something) and just read about it, follow links etc. Chances are you'll find a subject related to the random subject you'll like.

Of course the questions is if you then become interested in the new subject or make new friends, does that just extend your own homophilic network?

  Fred [10.17.06 05:17 AM]

Designers first need to decide whether homophily is a a feature or a bug.

I think this viewpoint is very telling - and it illustrates the core challenge to designers of social software. That is, do we attempt to design software to our users that reinforces their social patterns, or do we design software that breaks these patterns or introduces new complexity into their social world.

For example, a person might have one bar that they frequent. They like this bar because they know the bartender, their friends frequent the establishment, they like what is on tap - they are comfortable with the experience. Is it the role of the social software developer to reinforce this positive experience, or to force the user to experience new contexts? If you forced the user to leave their bar and try new ones, they might like these new bars, but this action would be breaking the established social pattern this individual is accustomed to.

Humans are homophilic largely because we are movement-limited. We spend the first 18 years of our lives around a small, protected group of people. Once we have mobility, we largely choose to remain around people like our original group - it is homophily, but it is homophily because it is what we are used to. That is, our social patterns have constructed a value for homophily that consistently operationalize.

Of course, on the internet, there is no limitation to mobility. A person can be anywhere - but this new context doesn't mean we leave our social values behind. CMC literature shows that in online contexts, we will grasp at the smallest notions of similarity to find group identity. This desire to be in a like group, a carryover from our real-world, homophilc existence, is extremely strong. Furthermore, we will go to great lengths to bootstrap our existing social models onto the online contexts.

Without question, we use the internet to find new things. However, just like we don't expect that comfortable bar to be gone one day, forcing us to use a new bar, our social software shouldn't force us to break our existing social models.

Therefore, it seems to be useful to not think of homophily as a problem to be solved - indeed, homophily has negative consequence such as reinforcement of stereotype - but social software can't break this. Instead, we must think of ways to coax users to experience new contexts, operating within their social experience. Homophily isn't a problem to be fixed, per se, but rather an important part of human nature that must be designed around in social software.

  Vera Bass [10.17.06 07:32 AM]

yes yes yes and...

Immediately before reading this, I read Gary Stein's post called Clique-Through: Deep Reach Within Small Audiences.

Drilling deep into a small core group for a market foundation is now a proven strategy, and how to broaden the scope of these strong pools is a current question.

...my specific focus for some time has been on linking core groups through the places where they overlap the most.


  Keith [10.17.06 08:05 AM]

Anti-homophily and homogeny features will be most successful when they focus on unresolved user needs. The argument that homophily breeds narrow-mindedness is somewhat irrelevant if your cool meet-someone-new feature dies in obscurity. That being said, the time is nigh for web 2.0 user contributions to go beyond content and into service and support. Take a service like Netflix that could really benefit from some mold-breaking recommendation capability. As a customer, I could flag my profile as one that is looking for some personalized movie recommendations from other users. By rating how good the specific user recommendations I get are (e.g. "I've already seen it" vs. "ordered it and it blew me away�"), recommenders build a matchmaker quality rating, show up on leaderboards, etc. This reputation-style rating in turn is used to rank recommendations more highly if they come from certain guides, and one could imagine this going deeper in a number of ways (who's best at finding sleepers?). This concept also meshes well with the blurring of site user vs. site administrator phenomenon that seems to be unfolding as participatory services mature - take Wikipedia (bureaucrats) and Slashdot (admins) as two examples.

  Shawn Smith [10.17.06 12:51 PM]

Books, music and reading material are one thing, but this issue is especially profound in the context of online matchmaking services.

I took a few small steps into the world of online dating a couple of years ago, but I met my current girlfriend in a very old-school, offline way. On the surface, we have very little in common. We come from different ethnicities and very different backgrounds. She's a techie. I'm artsy fartsy. Our hobbies and interests barely overlap. Where it counts, however, we are perfectly aligned.

It's unlikely that an online matchmaking service would have steered me to someone so apparently different.

  Beau [10.17.06 01:02 PM]

Kudos for naming this long standing need in software. When I think of homo-/hetero-phily, I always go to Kai's Power Tools and its mutate function. It's one of the earliest software implementations that I can think of, and it gave the user the power to choose how similar or different they wanted their results.

  Calo Bob [10.18.06 11:19 AM]

Humans are homophilic largely because we are movement-limited. We spend the first 18 years of our lives around a small, protected group of people. Once we have mobility, we largely choose to remain around people like our original group - it is homophily, but it is homophily because it is what we are used to. That is, our social patterns have constructed a value for homophily that consistently operationalize.

  Greg Linden [10.19.06 09:15 AM]

Great post, Nat. You're worried about pigeonholing people, never showing them anything new and different.

From its early development, Findory was designed to avoid this kind of pigeonholing. There is good evidence that those efforts are successful. For example, of the few complaints Findory receives, the most common is that we don't pigeonhole people enough, especially on opinion articles where some people become irate when presented with opposing viewpoints.

More generally, the point of personalization is to aid discovery, to help people find stuff they probably wouldn't find on their own. As you suggested, in well designed systems, pigeonholing is avoid by not limiting to narrowly defined interests and by supplementing the selections with experiments to learn about new interests.

The key is to make sure the personalization reaches beyond the obvious and into the surprising. If you do that, personalization reveals the full breadth of the options available and enhances serendipity.

  2cWorth [10.27.06 04:35 AM]


Just posted a longish comment at the Profy blog (url above) - with the spate of new(er) social networking sites that come up every so often, as well as the advent of those who game each system for commercial gain, I doubt if any single monobloc is going to emerge.

Not to mention the other 7 billion who aren't online, and therefore not part of this.

  Justine [11.16.06 06:11 PM]

I'm late to this (too late for a copy of Make magazine? ;-) but here's an example of anti-homophily: LibraryThing's 'UnSuggester' - books that are least likely to share a library with the book you suggest.

  Julie Dole [07.25.07 09:58 AM]

I think the absolute easiest way to bust up the homophily on a social site is to simply add a button that links to the Arts and Letters Daily site (www.aldaily.com.) It is intellectual serendipity defined, and having it as a homepage has painlessly introduced me to useful and entertaining stuff for years.

In fact, it led me to the WP article you cited, and then to your blog.


Post A Comment:

 (please be patient, comments may take awhile to post)

Type the characters you see in the picture above.