For social search, similarity could trump friendship

There's a difference between people you know and the people you're like.

Just because I friend or follow someone doesn’t mean we’re alike. We might be colleagues or classmates. Maybe we’re both fans of a baseball team or a certain twisty TV show. That’s the extent of our overlap.

Altimeter Group founder Charlene Li (@charleneli) touched on the difference between people you know and people you’re like during our interview at Web 2.0 Summit. Here’s our exchange:

Will social search and algorithmic search eventually combine?

Charlene Li: I think social search has been bandied around for a long time as an additional signal to pure algorithmic search. Social was just another input into the algorithm in the same way that Page Rank or word frequency might be …

You have two types of social graphs in there: people you know and also people like you. That’s the part around personalized search that’s also in semantic search, because language is unique to each person.

During our chat, Li said the partnership between Facebook and Bing is a step toward a unified social-algorithmic engine because Bing is extracting patterns from Facebook data. It’s not blindly tossing queries into your Facebook network.

After speaking with Li, it occurred to me that the current state of social search is reminiscent of those pre-Google days when results seemed arbitrary. I’ve encountered that old “ask and pray” feeling myself whenever I pose questions to my networks. Sometimes I receive useful information — stuff I’d never get from a strict search engine — but there’s no rhyme or reason to that process. Occasionally, there’s no response at all.

Social search built around similarity — the “like” rather than the “know” — could improve its reliability. To increase the chances and relevance of similarity, social engines need to also expand the boundaries of “social proximity” to include friends of friends and others adjacent to your social graph. This cocktail of similarity and expansion could yield the Page-Rankian shift that transforms social search from an occasional option to a reliable resource.

If you’re interested, the full interview with Li is included in the following video (the part about social search starts at 1:45):

Note: Portions of this post were edited after initial publication. Changes were made for clarity; it was hard to determine which comments were associated with Charlene Li and which were made by me. My apologies for any confusion. — Mac.


tags: , ,
  • Alex Tolley

    “For social search, similarity trumps friendship”

    Charlene Li does not say this. Are you saying it, and if so what is the evidence for this statement?

  • Mac Slocum

    To be clear, I’m making that point. Your comment made me see the original headline and parts of the post were misleading, so I made changes that hopefully clarify distinctions between who is saying what.

    Expanding on the importance of similarity a bit: That aspect occurred to me after Charlene made her very adept point about “people you know and also people like you.” That was an eye-opener for me.

    I think evidence will come from relationships like Facebook and Bing and others that form. We’re in a very early stage with social search, which is the point I was trying to make with the comparison to “pre-Google days.”

    I think the issues, however, are evident in the networks we’ve all assembled. The people I follow on Twitter or friend on Facebook or link to on LinkedIn do not necessarily share traits relevant to my specific queries. To me, that feels like a shot in the dark.

  • Chris Anderson at Wired said something similar back in 2005, that friends aren’t the best people to use for recommendations, because “someone you don’t know has found the coolest stuff”. See


  • Mac Slocum

    @Greg: Thanks for those links.

    I was struck by this bit from Anderson’s post:

    > The filters that work best for me typically earned my trust by liking some of the same things I did, then turning me on to new stuff that I liked even more. I really don’t need to know anything more about these people, other than that they’ve got more time than me and are willing to listen to a lot of junk in search of undiscovered gems. And they, in turn, don’t care much about me. When it comes to recommendations, friendship is overrrated.

    It makes me wonder if the vaunted personal social graph — which is so useful and interesting for maintaining ambient relationships via curation — perhaps does *not* have a future on the search side. At least not as it’s being currently portrayed.

    Rather, for a search tool to really call out similarities and allow for recommendations from people you “don’t need to know anything more about,” you’d need an engine that taps the entire social graph. My small graph snippet, which has been customized for a distinct use case, isn’t enough. Not by a long shot.

    Put another way: A social search engine needs all of Facebook’s “Likes,” not just the ones from my friends.

  • Keith Williamson

    I am glad to finally see an article that is trying to objectively look at the importance of social search or the social graph itself. The social graph has been highly praised by many to be the next big thing in the search/recommendation space. Your article correctly points out that social networks might not be the most reliable when it comes to search results since they rely heavily on “self-similarity”.

    But let’s assume that we have an ideal social network where a user only has friends who are similar to him/her. Then what! Knowing about friends who are similar to the user is only useful for directing recommendations from the friend to the user. The fundamental problem with using “self-similarity” in social graphs for actually creating recommendations requires you to at the very least deduce information about the user itself, which is a very difficult problem. Knowing similar friends doesn’t make it any easier.

  • There is a lot of difference between the people we know and the people or groups we like. We may know people back in grade school but is totally different from me. I may also like a person now but later on may dislike him after finding out more about him but kept our friendship. It is hard to base one’s personality on his social activities. I don’t a computer algorithm can do this, maybe human analogy can.

  • @ mac, keith and greg

    At last, an article that has got to the heart of the problem – it isn’t about people you know but about people like you. Of course, there may be people you know that are like you in your social network but this tenuous link is a poor basis for social search and recommendations. People are social at various levels – from very close to not so close – but yet most of us want to retain our individual identity.

    As Mac says, ” … you’d need an engine that taps the entire social graph. My small graph snippet, which has been customized for a distinct use case, isn’t enough. Not by a long shot.”

  • A point that hasn’t been brought up: similarity between people is very domain-specific. Subjects overlap somewhat, but a search algorithm that blindly recommends everything I like to another party simply because we both like Linux is going to be poor.

    Similarly, an algorithm that only forwards my recommendations to people who like Linux AND Ruby AND Democrats AND Solar Power AND Jane Austen AND Low-Budget Zombie Flicks AND Martial Arts AND … you get the idea. The population will be too narrow.

    Arguably, this may mean there are limits to the added benefits that social network data can provide. If everyone who is interested in Harry Potter is providing grist for my Harry Potter search, then it’s not clear to me what the social network is adding.

    I may be missing something.

  • Similarity may trump friendship, but at least for me, serendipity trumps similarity.

    I do not want to be exposed solely – or even primarily – to things people like me like. I want to be exposed to a broader range of alternative perspectives.

    Social search is an intriguing idea, and I do believe that using information about what people like me like may help inform search, but an excessive focus on similarity may reinforce the negative effects of echo chambers, silos and increasing polarization.

  • Julien

    Isn’t there a confusion between recommendation and searching?

    I trust expert strangers or profiles similar to mine for finding a needle in the haystack.

    I trust friends for recommendations.
    (this is true in the real world, I’m not sure it’s really true online)

    I feel that the two notions are becoming entangled. Maybe before defining the solution the problem needs to be defined. Are we talking search or recommendation?

    Of course, there’s also those that I know because I have followed their recommendations often because they are knowledgeable… which is the inverse logic. :-)

  • @ julian
    I prefer to talk about “personalization” (which encompasses recommendation and serendipity engines). I think that social search should also fall into the personalization category otherwise it becomes confusing wrt web search which the majority relate with Google, Bing and Yahoo!.

  • It all comes down to making the connection — consumer queries and business responses. It seems that most Web products are currently concerned with retail [purchases]. Picking something off a shelf and/or putting it in a shopping cart (in a brick and mortar store, or on the Internet,) are not heavy lifting, so to speak. Business responses are true “results.” Forward search, semantic or not, is still research, which will always be the case, and many times be required prior to actually querying to make a purchase. “Social proximity” is not an issue. In the connected age, we’re near most everything. “Friending” someone doesn’t build a true social graph. The deeper we dig into “likes” and “similarity,” the closer we get to inverse searches where the query comes before the information we seek – as opposed to forward searching, where the information is already out there, and it’s our job to sort through the starting places to query and inevitably arrive at a point at which we need to query. While ratings may aid us in making decisions, they are still part of the research process that comes before the query. Change of taste, circumstances or other factors might require “unlike” buttons, which would further complicate things. Research is one thing, but ecommerce is a whole different issue. Inverse searches can be used for research and ecommerce. Asking people for recommendations and Q&A can be public to a point, but business transactions are private.