Taste Neighbors
I was walking home from work last night around 9pm needing to get something to eat and had a hankering for a fish taco and a beer. So I pulled up Yelp on my phone and looked for fish tacos in the west village. Mary’s Fish Camp came up first and I like Mary’s so I headed there. It turns out that Mary’s doesn’t serve fish tacos but I had a great shrimp taco (and an Anchor Steam) and so that story ends well.
Mary’s was at the top of the list on Yelp because it was well reviewed by 65 people, way ahead of all the other fish taco places in the west village including the places that actually do serve fish tacos.
The Gotham Gal always laughs at me when I take these Yelp reviews seriously. Same with the reviews on Menupages. She points out, rightly, that I have no idea who these reviewers are, what they value, where they like to eat, and if they know good food from bad.
And in fact I ignored a bunch of the reviews I saw on my blackberry walking through the village last night. Let’s take this one:
Cute and enduring? ehhh. Nice and fast staff? Yes. Grossly overpriced
even for west vil? YES! $28 for a lobster roll (just so-so) and $10 for
a plain ol normal sized hot fudge sundae?? these should not be the
prices of a tiny lil neighborhood joint. There are better places for
fish. Although Pearl’s aint all that great either.
This person is price sensitive and rightly so. $28 for a lobster roll is a lot of money. Same with $10 for an ice cream sundae. But if its the best lobster roll and the best ice cream sundae in the world some would be happy to pay that price. Different strokes for different folks.
Which leads me to the point of this post. We need "Taste Neighbors" to filter the user generated reviews that are piling up all over the Internet. It’s nice and all to have ratings for reviewers like Amazon pioneered and everyone copied. But just because most people think that someone’s a good reviewer doesn’t mean their taste matches mine.
This problem has largely been solved in music. Because its relatively simple to watch what music I listen to and what music millions of others listen to, there are many services now that use musical neighbors to drive recommendations. My personal favorite of these services is last.fm and this is a list of my musical neighbors.
I was going to say that I don’t know any of these people, but that’s not true. I know Dave Hyndman and Peter Hershberg. Very cool. So it’s possible that your neighbors in online profile land can be people you know, but most of the time they aren’t. They are people like this 20 year woman from NYC who I’ll probably never meet in person but who enjoys pretty much exactly the same music I do.
How do we apply this technology to user generated ratings and reviews? It’s not so easy to figure out who likes the same movies, restaurants, and books as I do. The consumption of these things is done offline (for the most part anyway) and its not as easily captured as music.
I think the first step is to perfect sentiment analysis so we can automatically determine the sentiment of any review. I like what Summize is doing in this area, today it’s largely for Amazon reviews but hopefully they can extend it to the rest of the web. Once we know the sentiment of every review we can start to match reviewer and the sentiment of their reviews to build online profiles of people who fill out user generated ratings and reviews. We can use cookies to identify who they are but it would be better if we had some kind of universal registration system, maybe built on Open ID, to better identify and track who is leaving what reviews.
Once we have profiles of reviewers/raters, we can start to match them up with others, to create taste neighbors. But what about the people who don’t review and rate? It’s generally true that only about 10% of all consumers of user generated content actually contribute to user generated content. So we need some way to tell you who your taste neighbors are if you don’t rate and review. Thumbs up/thumbs down style rating of reviews would help. If the 90% that don’t rate and review would simply vote on the reviews we could determine taste neighbors for them that way.
But I am more optimistic about watching what they actually do. As my former partner Bliss used to say, "watch what they do, not what they say". With online finance services like Wesabe (one of our portfolio companies), you can easily build a database of every restaurant you eat at, every movie you go out and see, every book you buy from Amazon, etc.
If I contributed that data (with the financial data stripped out) via the Wesabe API to a "Taste Neighbors" service, it would be simple to profile me based on my actual behavior. You’d know, for example, that for sushi I prefer Blue Ribbon and Yasuda, for burgers its Shake Shack and the Burger Joint, for Italian, Babbo, Lupa, and Esca.
People who share those restaurant preferences with me are going to generate much more reliable reviews and ratings for restaurants than the person who panned Mary’s Fish Camp. When I look at her profile in Yelp, I see that she also panned Stantion Social, a place I like very much. So we are not Taste Neighbors. And that’s fine. But we need a system to figure that out in advance. Otherwise the Gotham Gal is going to keep laughing at me for reading these reviews. And that’s gotta stop.
Comments (Archived):
great post fred.it’s a big problem and oppty.I always hoped that LinkedIn or someone would figure this out for us for some sort of taste matching.I had a post on this subject too.http://sabet.typepad.com/bi…
I was at Stanton Social on Saturday night and we had the exact same discussion about Yelp limiltations. The conversation was with Stellah, queen yelper from san francisco, who was pointing out the same issues.small world indeed.
Fred – Amen! We also believe that user written reviews need rich profiles attached to them. You should take a look at how Viewpoints.com is coming along. By allowing people to tag themselves each time they write a review we help them build a rich profile like you are describing. Our “I Am” tag is like a tag for a picture, video or news article. It is an effective way to achieve two important goals. One is to allower the reviewer the flexibility to self describe themselves with the exact words that fit them. The other is to help the reviewer filter and navigate to reviews based on these tags. If you were to write a review you might tag yourself as “looking for a fish taco” or “not price sensitive”. Three months into our beta we have 80,000+ users. Next month we are launching the 1.0 release with some very important search and navigation features.One lesson we have learned is that people’s profiles differ depending on what “role” they are playing in their life. At dinner with the Gotham Gal, you might be a romantic, slow eating, non price sensitive guy, but when you go out for lunch, it you might be inpatient and looking for a healthy alternative. This “situational dependence” is one reason why aggregating anonymous reviews or aggregating behavior from transacational services may not lead to the best result. If you looked at my credit card you might see what I am doing, but you don’t know if I was doing it as a father, husband, or business traveler.Keep those posts coming. Always thought provoking. And thanks for the Kings of Leon recco.
Quick Plug http://www.tastespace.com
It would be more interesting if I had a better “address book” – one where I could include userID’s from the various social networking sites. In my address book I could weight your opinion, then when visiting a site a widget on the side could list any associations with their weight….Just a little something I have been toying with.
Take a look at http://www.cellartracker.com. Excellent platform for putting this type of system in place, because it allows you to see what other users also own the wines you use, which immediately defines a community.
I agree with Gotham Gal, that any given review on a site like Yelp or MenuPages has a chance to be crappy. That’s because a one-person sample is flawed (ie “law of small numbers”). There are a lot of competing restaurants posting biased reviews, and a lot of people with a chip on their shoulder silly reasons. However, I find that the aggregate of the reviews is a pretty accurate representation of my own opinions.Furthermore, Zagat’s is panel-based, so if you don’t trust the aggregate of Yelp or MenuPages, then I wouldn’t trust Zagat’s either. Yet, most people see Zagat’s as the gold standard.
MOG (http://www.mog.com)has added this feature to editorial postings. you can push the magic button in any section of MOG (music news, mp3 uploads, watch, mogtv, etc) and it filters the user generated content based on its profile of mogomatic data.
Generally “tailored ratings” are a big opportunity that has never received much attention. Amazon has “those who liked x also liked y”, Netflix has it, but Digg, etc.? No. You made an important point that the user base has to be a certain size, but I don’t think it has to be tremendously large. Remember that things get statistically significant at 30 samples.We’ve put so much great information out there now, but as it reaches a certain size threshhold its value slowly deteriorates. We have to come up with better ways to winnow away the chaff appropriately, and tailored ratings is probably the best tool out there.
Thanks for the tip on mobile Yelp, Fred — I’ve added it to my mobile bookmarks.
I really like how Netflix does this with their user reviews. Next to each reviewer, it says “76% similar to you,” or whatever the percentage is.
Also, there was an interesting article on the Netflix community blog about why they don’t open up more of their data (it is illegal, who knew?) http://blog.netflix.com/200…
I’ve long wanted to mash up my Netflix movie ratings with RottenTomatoes, to filter the reviewer list down to a subset of those with whom I tend to agree. Then RT can show all reveiwers, Cream of the Crop, and MyReviewers. Seems obvious to me. RT is a good site but throw in community reviews, ratings, and critic matching and it’s a huge hit. Then make it all work as a Facebook app.
Snooth.com does taste neighbor matching for wine.
Fred, I have tried a couple of times to email you a summary for a business model, digiSelf, that profits from exactly these concepts you discuss but I can’t seem to break through your Inbox back log. If I sent it again would you give me your thoughts on it ?
Just like Gotham Gal and Mr. Parker noted below, reviews are subjective and what someone says may not be relevant to you. Money may not be an object. You may dislike a certain type of fish that someone else loves. What may be loud may be moderate to someone else. Lots of assumptions and variables. If we can get around that (thru technology, we can), we’ll start building an intelligent system. However, in order for this to happen, we must establish the browsing users identity.
Isn’t what you describe essentially… Zagat?
last.fm succeeds because the scrobbling is passive. All the user needs to do is agree to it, and let it run. With restaurants, there are greater outside motivators that affect decisions and no passive submission process. With that said, I have always appreciated review relevance and scoring – ie what Metacritic does with their metascore. Also, being able to weigh or exclude reviewers from a search or result set is key to allowing refined results. One thing I’ve taken to doing when looking for the bones of something is to do a blog search on Google. I find that if a user posts about something (whether it being code, a restaurant, or new gaget) it has more relevance to me and generally provides more insight (or links) into the information I am truly after.One comment about your new comments system. As a UI person, I dislike the field labels being in the field only. This approach works well for a single field item like search, but not so for multiples. I found myself tabbing out and back to see which field was what when leaving my name, email, and website.
There might be something to be learned from the world of content discovery here (ie matching two pieces of similar content).There are two primary ways to match content with content. Method one is a collaborative filtering which compares subsets of data – ratings, clicks, purchases, listening history, etc. The strength of this approach is that it’s quantitative – if you have the data on each individual, you can crunch those numbers and calculate out matches. People who liked/bought/listened to A, also liked/bought/listened to B. The problem with this approach for consumer rating, as Fred points out, as you typically DON’T have this data. Most people don’t rate enough stuff online, and in the same place, to draw meaningful correlations. For example, Fred generates a lot of data online, but I’d be willing to guess most of it isn’t in a neat 1 to 5 rating scale.The second approach used by content discovery services is to actually spider / process / analyze content, and in an automated fashion, infer its meaning. This is the approach used by services like Google (to match queries to content), Media River (to match content to articles), and Inform. It’s extremely difficult to solve this problem, both from a technology perspective, and from an access to data perspective. In the world of consumer rating, the equivalent, as Fred entions, would be crawling and processing a master set of user profiles, inferring tastes from those profiles, and matching profiles which each other. There are numerous challenges with this approach as well. Profiles vary wildly in their completeness, many are held in walled gardens and not accessible by third parties, and their structure is often very different. Profile aggregators could fulfill a valuable role here if any were to gain traction.A third approach, as used by several social networking sites including my own employer, is to enable the filtering of ratings and recommendations by network of friends (just like sites Facebook filters newsfeed activity by your network). While this is a nice feature, cross property filtering is a challenge, and it’s unclear if a recommendation from a friend is any better than a recommendation from a non-friend, all other things being equal. I have friends, who if they recommended a band to me, it would be the last band in the world that I’d listen to.My take is that the “answer” lies somewhere between approaches one and two – crunching hard, explicit data, and parsing profile and ugc from individuals.
I love fish Tacos from taco Milagro in Houston.
First of all, I’d like to let you know your blog is being avidly read by many – including me – from 5,000 miles away – in Korea! As for your article, I’d like to come from a slightly different angle: Do we really want to find our Taste Neighbors? And if so, what’s the purpose? I mean, it would certainly be interesting to find someone whom I haven’t met actually possesses a very similar taste to me. But then… what’s the next action? If you are a (straight) guy, and you found another guy who listens to similar music, watches similar movies, and goes to similar restaurants as you do, will you ask him out or something? Probably not. If I were to receive a hello message from my “style doppleganger” who happens to be a *guy*, I’d probably feel a bit creepy. If this style buddy is a girl, well that would be a much better thing :), but then on second thought, if the girl has pretty similar cultural tastes as you – a guy – chances are she could be less feminine, or could even be… wacky/macho. I mean, do you want to go out with a girl who, just like you, likes to watch UFC fighting championship matches or Monday Night Football? I doubt it. So the biggest merit of this “Taste Neighbor” system we can think of would be the “Stumble upon” effect i.e. finding other items also liked by your Taste Neighbor, which you weren’t aware of before. But then, when it comes to this “Stumble upon” effect, wouldn’t it be better to just have your friends display their favorite items, rather than allowing the system explicitly or implicitly import and analyze your personal data, which might lead to potential privacy concerns?
As you point out, social networking and relationships online need to be dynamic. You’re never going to get everyone to contribute in a meaningful way, or to even get them to join. Anything you can do to dynamically “force” them into a network in a meaningful way is key. The best way to do that is to use existing data from services you already use or would want to use, not to request that people contribute for the sake of building up a profile. Sorry to plug, but I really do talk about this in some depth on a blog posting of my own from earlier today: http://www.bschoolers.com/2… (starting mostly with Implications of broadcasting and sharing through social networks).
I think the answer to the problem raised in Fred’s post above is “social-search” which effectively outputs results that are filtered through your trusted network of “friends.” True, not every recommendation suggested by your friends works for you; but you’re a heck of a lot more likely to act upon a recommendation of a friend who knows and understands you rather than a complete stranger (with whom I share few commonalities).The example Mark uses in his blog posting is spot on: if I am looking for a camera through my social search engine, it’s similar to sitting around a dinner table with my friends asking for their input as to which camera I should buy. If five of my friends bought and/or recommend the Cannon Elph, this camera is likely to be a very good option for me as well. Clearly, social search is a very powerful form of trust-based recommendations.And the infrastructure is there and evolving at an incredibly rapid pace, albeit coming from two unique (and hopefully interconnected) approaches. Facebook has its active/user-defined network forming approach while Google uses its dynamic analysis tools to “create” (or identify) your relevant network. While Mark takes a position in the comment above, I am not fully convinced that one approach is clearly superior to the other (at least given current technology), but my intuition tells me the most powerful applications will incorporate the best elements of both the user-defined and dynamically extrapolated networking approaches.
Ben, that’s a good point about the two approaches – and actually to back off of my comments above, I do argue both sides in my aforementioned blog posting. You want to capture as many people as you can, and some love the Facebook style “active networking.” The benefit there is that you encourage active engagement. But I would still say on the whole the Google style dynamic model is at least more interesting to me because it could reach really nearly everyone that the active participation model doesn’t. Completing the pie and getting everyone involved is a huge piece of this puzzle.
Fred, Howard, Andrew, and Darren:On Yelp you want the content (the restaurant listing) to be contextual and the past user engagement (the review/rating) to increase the value of that content. Case in point: I have one friend who knows restaurants really well, and her recommendation means 100X that of some Joe on Yelp (whose review is nearly valueless to me).That sentiment makes a lot of sense, and I think it can be said for all types of content. Getting the perfect amount of input from the global community and the best input from your community is what it’s all about. That’s about 1/3rd the power of BricaBox.
The Gotham Gal is renowned for her wisdom.
Hey Fred,At one of the meetups last month (think web2NY) I saw a company present their business for a site where users upload video reviews of products and places. At the time I thought it was OK, and while reading this post this company, again, popped in my head.All the things you say in the post make sense, but I think we are hardwired to look/listen to someone for 3 seconds and we know if they are our taste neighbors.
When Wesabe launched its API, I realized that by leveraging it, any review system — from Yelp to Citysearch to Menupages — could confirm if you truly made a purchase at a particular merchant such as restaurants. It is probably difficult to enumerate how many reviews are simply biased (think Whole Foods CEO), but if someone spends $40 at a restaurant, wouldn’t his opinion matter more than the person who spent $5? Or, what if the latter person frequents the restaurant four times a month on average, and the former only once.The mashup possibilites of the Wesabe API can bring that added confirmation so that when I am linked up with a fellow reviewer with similar tastes, it’s a person that truly has transacted at the merchant they reviewed. Even better, with Image Capture from Wesabe, Billeo, and the distant NeatReceipts, I can be connected with individuals who love specific foods (e.g., coffee, vegetarian food), because transaction level details are much more powerful.
Taste of course can be fickle to articulate. PowerReviews is trying to offer something similar to what you wrote about on their ASP service for retailers. I don’t know how popular that particular feature is with their partners.
Fred, the approach we’re taking with a project about to launch (a series of guide to cities by business people and for business people) is to emphasize a common purpose and generally perspective – i.e. in our case that of business people.So for us a great restaurant is not just judged by food – but as importantly by fit for certain business functions (i.e. Shake Shack is indeed great – but an atypical place for a business lunch – and any discussion held in the middle of a public park is not exactly private).Our thesis is that this commonness of focus should help our reviews & recommendations to be useful for our target audience.That said, we also plan on having an API and lots of potential for mashups – so I’ll take a look at what something like the Wasabi API might allow in conjunction with our API’s. (confirmation of a pattern of spending at a given place might, for example, give a reviewer’s review extra weight).The problem is not limited to restaurants – friends and I have had almost this exact conversation over hotel review sites online – noting that there in particular the needs of a business traveler, a budget travelers/college student, and a honeymooner are very different – and rarely captured by the reviews of any given site.