The Importance of “People like Me” features

Ha, just saw this post after emailing you. So now I feel compelled to comment.

First of all, person-to-person similarities does not scale, which is why Amazon does not use it for recommendations (we published that fact several years ago, back when we published academic papers). This is because of the slow ramp up before getting recommendations - it takes a lot more ratings to generate person sims than if you just did item-to-item similarities. With item-based sims, I could buy 1 product and have zero ratings and still be able to see recommendations, since the 1 item I purchased already had similarities precomputed.

As I said in email to you, there are a lot of people who react negatively to “people like me” because many people have this fantasy that “nobody is like me. I’m unique”. I think most of the issue could be solved with improved messaging - “Recommendations from customers with similar purchases” is much more innocuous than “Recommendations from customers just like you”. Another good idea (credit goes to Josh for this) is choosing the similarities from a set of already trusted people - your friends or designated experts. Of course, everyone >> friends + experts, so there may be a coverage issue where there really aren’t good similarity matches amongst this trusted group, but I guess you’d have to try it and see what happens.

TrustedOpinions.com is a site that was working on something like this, as was crowdstorm.com. I’m not overwhelmed by either one, but I’m glad to see people investing in this area.

I hope this isn’t too shameless a plug, but the ‘people like me’ idea hits home. We’ve built a disease management site around it: PatientsLikeMe

Patients have tools to enter extensive data about their disease progression, symptoms, treatments, etc. This information is reflected back in a series of charts and other visual representations that become their profile within the community. The profile becomes an identity of sorts and is the basis for finding ‘patients like me’. I should add that ‘like me’ can extend along a variety of axes - age, disease progression, symptoms experienced, treatments taken, etc.

One identity feature we like is the concept of a patient ‘nugget’ which is a small visual summary of a user’s disease state - identity in a nutshell. It appears in the profile and next to users’ forum posts, giving others in the community a quick way to gauge what kind of patient they’re talking to (early or late stage in the disease, mild or severe symptoms, etc). It’s proven to be a useful snapshot of identity within our communities and both a starting point and filter for conversations. We’ve even seen cases where patients without much of their profile filled out post to forums and get pushback from others to fill in more so they know who they’re talking to, whether that person actually has the disease, etc. Those with extensive profiles tend to be taken much more seriously and, in addition to having more to offer, have more ways to find similar users.

An example of a profile can be found here:

http://www.patientslikeme.com/patients/view/993

The nugget is the image next to the photo. You can mouse over to see what each visual element represents.

Interesting topic Josh, and equally interesting comment by David. When it comes to “people like me” or “people who bought X also bought Y” features a couple issues need to be considered. I always take reviews and ratings on product sites, like Amazon, with a hefty grain of salt. I don’t know who those people are or what makes them tick - do they have a stake in the product, or a competitor’s product? Would they use the product the same way I would use it. My naturally skeptical nature prevails.

On a site like Last.fm (one of my favorites), the “people like me” or “neighbors” feature is a completely different experience. Last.fm users’ musical tastes are being implicitly tracked behind the scenes. I know that I have a good chance of discovering music that I may enjoy because the chances for bias or gaming the system are extremely low.

I guess it comes down to this; I am more skeptical about explicit (you know what I mean!) reviews by users than implicit actions.

@Scott - to be clear, “Customers you bought X also bought” as well as Personalized Recommendations is completely unbiased. We mine the data in a way that is agnostic as to who bought what. You can trust that what we show you is what the data says to show you - there is no editorial or manual process.

@David - my apologies. I actually did know that Amazon uses item-based collaborative filtering, but I bungled the post. I wrote it completely out of order, adding the bit about Netflix and person-based filtering after the fact. I’ve updated it to be more accurate.

Now you have me wondering…it seems like a difference between Amazon recommendations and Netflix recommendations is that Amazon considers purchase decision to be a positive action, whereas Netflix does not consider renting a positive action.

Netflix, as far as I understand it, explicitly looks at the rating when making recommendations. Does this have the effect of necessitating person-based collaborative filtering, or could you say “people who rated this highly also rated this highly”? Perhaps that’s what they doing anyway?

@Rich: You’re absolutely right! Patientslikeme is probably the best example here. I’ve long used it as an example, and should have used it in this post! So plug away!

You bring up a great point, as well, about trust. You’ve built an app in an environment where trust is crucial, and for patients with diseases they don’t always understand (or are completely new to) it makes all the sense in the world that they’ll try to find someone who has been through the same battles. I like the idea of showing similarity through graphs of treatment data…almost adding to the trust factor.

Thanks for writing, Rich. I’m sure we could all learn a lot from your experience at patientslikeme.

@Scott - Yes, the distinction between implicit and explicit is interesting, especially that you trust one more than the other.

I wonder if it matters who people rate items for. If they’re rating it for themselves, the ratings will probably be more accurate than if they’re rating for others…since the social performance might affect how they act. (at least for some people)

I think many people are desperate to make sense of the sometimes overwhelming number of options. It might not be the greatness of the opinions of other individuals, as the spin of corporations that makes the alternative so attractive.

@Josh Again, you make good points. Two, in fact. First, the fact that Amazon wants you to purchase but Netflix does NOT want you to rent (especially high demand items) is a critical and oft-overlooked difference between our systems. The ideal Netflix customer is the customer that only rarely rents a movie, and that movie is never a high demand item. Amazon wants you to buy whatever product you want. In terms of ratings, you could think of ratings similar to purchases; the weight you assign the rated product may or may not differ, but it’s just another item to be considered. So, it doesn’t require person-to-person similarities.

Your second point (directed @Scott) is also very interesting. When you rate an item, are you rating the item (like a customer review)? Are you rating the quality of the recommendation algorithm? Is my 4 star rating the same as your 4 star rating? Maybe the ratings should be normalized across customers?

Here’s an interesting twist on this subject. A Facebook Beacon experiment that produced some pretty invasive results for some employees at Yahoo and Microsoft who unknowingly had their profile pics used in an ads with the taglines “Leaving Yahoo?”, and “Leaving Microsoft?”, because they had become “fans” of this a company’s Facebook profile.

While the itent of the program was to provide a “genuine” peer endorsement experience for a product, it’s pretty clear that it can be misused as someone indicated that the ads looked less like “ads” and more like defametory propaganda, for lack of a better term.

Here’s a link: http://valleywag.com/354279/vc-freaks-out-yahoos-with-shocking-facebook-ad

No specific examples, but I would love a blog comment plugin that says, “You have a similar comment to (name) and (name), would you like to respond to them here?”

That would be awesome. Keyword recgonition I guess…

Years ago Amazon did have exactly a “People Like You” feature, and it was fantastic. It might well have been the first mainstream application of this sort of thing. You could page through a list of people with similar purchases and see items that they’d purhased that you hadn’t. I don’t recall, now, if ratings factored in or not.

Honestly, I still miss the feature today. It always provided far more interesting choices than their regular product recommendations. Amazon’s main problem there is a lack of real understanding about their products. Recommending a white KitchenAid mixer because I’ve said I own a blue one, or a Nikon D80 with lens because I’ve said I own a Nikon D80 body-only is a waste of everyone’s time.

Clearspace, from Jive Software has several features like this including a list of “Similar Users” when looking at someone’s profile. They also have a “More by UserX” feature when looking at a specific page.

http://www.jivesoftware.com/products/clearspace/

I’m surprised, and a little skeptical, to hear that Amazon’s Personalized Recommendations feature doesn’t use person-based collaborative filtering. Obviously “people who bought this also bought” is item-based, but the Personalized Recommendations seem to leverage your *ratings* of products, too. I can go through Amazon and rate a hundred products to improve the recommendations, regardless of what I’ve shopped for or purchased in the past.

The current Personalized Recommendations page seems to say “We recommend X because you bought Y” for every item, but it may well be leveraging aggregate person-based data as well. There’s no reason why both systems couldn’t be combined into a single engine. I recall it did so in the past (I ever read a white paper on it), but maybe they’ve decided to lean more towards item-based filtering — or at least they are showing that aspect to their customers, since it’s a little easier to understand.

All this being said, I agree that actually *showing* people who are allegedly “like you” is a lot different than simply using that data behind the scenes to make suggestions.

Part of the “black art” of person-based collaborative filteting is that it is NOT about finding another single person very much like you, but that it aggregates thousands of people loosely similar to you. This is hard for many people to understand, which is why they often simply resort to displaying “you are like these people” instead of displaying all of the complex math used to create your profile and recommendations.

For example, a good collab filter engine might look at the fact that I like The Ramones and Devo and, instead of finding another person or group of people who like both of those two bands, it might look further, to 2nd and 3rd or 4th degrees of “like”: Ramones fans also like Talking Heads. Devo fans like Talking Heads. Then it sees that Talking Heads fans like Shriekback. So it then recommends Shriekback to me. If the system only looked at people *just* like me, it might never find that additional recommendation. Multiply this sytem by hundreds of things I like, cross-referencing with millions of other people’s lists of things they like, and you can see how “explaining” this to users is impossible. So they resort to simplistic things like “People who bought X also bought Y” and other transparent, easy-to-grok systems. But that doesn’t mean that arcane true person-based collab filters can’t also be used on the back end to smartify these engines.

Let me make this more complex just to show where a smart engine can go. I’ll add a third band to my list:

I like the Ramones, Devo, and New Order.

So:
New Order fans like Shriekback.
Ramones fans like Talking Heads.
Devo fans also like Talking Heads.
Talking Heads fans like Shriekback.

That’s one first order point for Shriekback, plus two second-order points for Shriekback.

So Shriekback can be recommended to me.

At PowerReviews we license our review technology to retail sites like Staples.com, RadioShack.com, Toysrus.com, etc. etc.

Our review template asks people to select what type of user they are of the product, so for Digital Cameras for example, they can select whether they are a Casual User, Professional, Hobbyist/ Enthusiast, and so on (or add their own). We then aggregate this data and allow our retailers to add a ’social navigation’ component to their site. We also aggregating this data and enabling users to narrow results on reviews based on their user profile on our own product review site Buzzillions.com

By the way Joshua, I discovered your site after hearing you speak at SNAP earlier this week, and I’m really glad I did. Great job and extremely helpful - I’m looking forward to your book.