Taste Neighbors (continued)

Comments (Archived):

1. Fred, I’d like to see Foursquare working on it rather than anyone else.

   1. that’s interesting. care to explain why?

      1. Briefly (because I’m going out for lunch!) it would be a great feature and it would get me to open the Foursquare app on my iphone and use it more often. I don’t want another app on my iphone that utilises the data that Foursquare already have and I think it would complement the existing service nicely.

         1. I second that. It would extend Foursquare’s game mechanics.

         2. Agreed. It’s nice to have everything in one app rather than bits of data/functionality scattered across many.

      2. I think check-out is the key to 4sq being able to support the information gathering to really make this use model work – closing the loop and giving users the opportunity to evaluate their experience in a lightweight (Netflix stars etc) way helps to assign “value” to a check-in. Not sure if 4sq needs to support this info gathering themselves – third-party like Yelp may work as well. I think a tie-back to the 4sq user’s profile is necessary in order to “weight” recommendations from your friends etc.A really neat side-feature to this (requires live geo which may have privacy concerns) would be to pick X friends and a type of meeting place and have 4sq suggest meet-up locations based on current geo location of each participant as well as reviews. Helping solve “Where should we go?” in a way that incorporates geo.Obviously check-out would need to be incorporated into the game mechanics as well – points are low-hanging fruit, super reviewer badges, and so on.

      3. this seems like it might be an obvious feature that foursquare would build at some point. so working on it would have significant risks of foursquare blowing you away down the road. it reminds me of the idea that plugging holes in platforms (like building a twitter client) might not be the best thing to work on.on the other hand if you integrated with other data as others have talked about, that might reduce those risks.

   2. I personally wouldn’t spend the time working on an app like this, simply because it’s such an easy feature for 4sq to include themselves.

2. Sounds like an interesting addition to Foursquare – but the problem I see with it is that Foursquare is very much a neutral arbiter of what you just did. I was in Hawaii, I checked in at a number of restaurants. There is no follow up by me to determine how much I really liked the restaurants, what dishes were great, whether you should sit inside or outside, what are the specials tonight etc… In music, because the check-in is essentially a song listened to, and they are 4 minutes long, the amount of data coming into the dataset quickly becomes meaningful – and songs that you don’t like tend not to get repeated – for places it is much harder to build in that level of self selection I would think – although you could probably extrapolate peoples home turf pretty easily. If I go back and look through my check-ins in my home town of New Canaan – its pretty clear what I like – and you can probably infer that because Aloi is my favorite Italian restaurant, I like it better than Sole or Cava – but I’m not sure how much more you can infer.

   1. Yes, and I would add that a check-in is not the same as a vote, even for places with multiple check-ins by the same person.I regularly meet some friends in a restaurant I don’t actually like that much. Should I stop checking-in because I don’t want the system to infer I like the place?It could all get very silly, very quickly.Votes should be explicit, not inferred.

      1. With enough data, I think inferences can be valid. For example, if a place is checked into primarily by people with families, you can assume it’s kid friendly. (If you were using FB data, you’d potentially know this.)It is also something that foursquare or others could add to their game mechanics. A few days after a visit, you could be prompted to enter a star rating. (Similar to what Netflix does when you return a movie.) This could also be positioned as a service for businesses. It’s a lot simpler than a comment card.

      2. At crāv (http://www.doyoucraveit.com) we are working on just that. Utilizing “satisfactions” for this very purpose in the recommendation realm.Many people (including myself) have checked in to places that they either knew they didnt like, or later discovered they didnt like.

      3. Again., this is why my friends and I don’t tend to check in everywhere, we like the mutiple badges approach, and why we like leaving tips. I’ve left negative ones in the past.

   2. I think the real answer will be a mash-up that incorporates both Quantitative and Qualitative sources i.e. Number of check ins from FourSquare (Quantity) and level of reviews from Yelp or other sources (Quality)Being a numbers orientated guy, I would probably expect FourSquare have the best starting point to convert raw data into semi-qualitative matches of similar tastes or requirements for a certain type of restaurant. Examples could include- Time of Day; identify popular spots for Breakfast, Lunch or Dinner.- Day type; identify popular spots for business meals (workdays) vs personal(weekends)- Proximity History; Could be used to identify local hot-spots from users who have a strong history of other nearby check-ins (i.e. they are probably locals). Equally could identify when users are in areas where they have little local knowledge.- Inconvenience Factor; Could be used to identify highly rated locations worth the effort to visit. i.e. users who regularly check in at a location that is just outside their normal locality.- Time at location; Identify good sit down locations vs fast and cheerful based on time before next check in (this would only work for a subset of serial users who check in very regularly)- Frequency; A regular spot. People keep coming back- Intimacy; A good location for large groups or individuals based on numbers checking in together…. you get the picture! The more data the better.

      1. More data can certainly make for a better service – but the key is getting the right data and correlating it correctly – which is I assume what Hunch is trying to do (see the comment on this thread from Chris Dixion)If you look at something as simple as music – at Pandora – every time they play a song, one of four things can happen:1. You hit Thumbs Up2. You hit Thumbs Down3. You skip the song4. You do nothing and listen.The first two are pretty clear indications of your taste in music – very much like a highly positive Yelp review – a 5 star rating on Netflix (or is it 4 star) or a 0 or 1 star rating – etc… Very unambiguous.The skip is harder to deal with because there’s no contextual information – did they skip the song because they don’t like it and it was easier to skip than thumbs down? did they skip because they are tired of hearing the song? (Pandora actually has a way of dealing with this song fatigue but it is rarely used) did they skip it because they like it, but it just didn’t fit the mood of the station they were looking for at the time? Very difficultOf course the listen is the hardest of all – because it provides absolutely no context. Could be you liked it or hated it – but just decided not to add to the data set with a click.Once you get out into the real world and try to aggregate the same things for restaurants, bars, shops, other places – it gets more and more difficult because the level of input necessary to make reasonable correlations gets much higher.I’ve always found zagat an interesting service in this regard. They have all sorts of online filters you can use to find outdoor – kid friendly restaurants in Manhattan for example – but the more constraints you put on the selection, the slimmer the data set becomes and the less relevant the answer is. You look at their reviews in NYC (which is the market I am most familiar with) and you start to understand the profile of the reviewers relative to yours.Compare that to Yelp – which is a far younger demo for reviews – and you get very very little overlap.That’s the real challenge.

         1. Hi Harry – agree that Pandora is a great example for finding music and I love their simple voting mechanism. However I’m beginning to think that the Pandora style analogy may not be ideal for tackling this problem as the user behaviors can be quite different.1 – Cost of participation. Participating in Pandora or Last.fm only requires a users time. All the hard work of sourcing content and delivering it is done by Pandora and the audio prompts a reaction. Flying overseas and checking in to a hotel in Dublin, or even the local restaurant in NY requires more cost and effort on behalf of the user. They need to research the location, decide to go there and then remember to check in without being prompted. People who check in have typically made a higher personal investment (along with time) when they check-in. This must carry considerable more weight in my mind.2 – The timing of the participation. Typically a user will at least have ‘tasted’ some of the music before they vote. It may only be 5-10 seconds but it is enough for them to make an informed vote. Check-ins could happen while outside parking the car or after paying the bill. It is hard to know how much interaction the user has had with the location before they check in (unless they checked in at the location before)3 – Social Reach / Status. Music preferences are typically kept private and are therefore probably a truer insight into personal taste. 4Square check ins are mainly broadcast to social networks to project an image based on where users go to eat, work, party or vacation. This is basic human nature (i.e. our egos) and desire to impress our social circle. i.e. How many check-ins do you see from McDonalds compared to updates from Cool locations or Airports/Conventions/Museums etc… Where as with music I probably would not skip the “Old McDonald had a farm” song if I secretly like it :)I haven’t used Zagat or Yelp too much here in Ireland but I do know that I’m now starving after talking about all this checking in to great restaurants 🙂

      2. better yet, the number of positive reviews in yelp by people who go to the same places i go to on foursquare

   3. dennis has said recently that tips and to-dos are under utilized and he wants to fix that

3. Hi Fred, I’m actually building something very similar to this. Aggregate my tastes from around the web (my Foursquare checkins, Last.fm listens, Goodreads.com books, etc) and then show me recommendations of similar people and stuff I might like. It represents a pivot from a company I launched last year called Gelato Dating (http://ge.la.to). I realized the opportunity was much larger than just dating and built version 2. It’s looking awesome.Could I give you a demo?Steve

   1. yes, how do we do that?

      1. It didn’t look like you were going to respond, so I went ahead and recorded a video demo and emailed the video link to you. Let me know if you don’t receive it. If you’d prefer, we could do a demo over skype.

4. Hey Fred – As you probably know, this is exactly what we are doing at Hunch. Our thesis, however, is that the data is much more powerful when you combine it across multiple sources. Democrats tend to like Vegan food, Vegans like deals at vintage clothing stores, etc. Keeping the data siloed by category is what has caused existing recommendation systems to disappoint. As an excellent article in the NYTimes magazine about the Netflix challenged noted out, researchers couldn’t predict whether a user likes Napoleon Dynamite based on her previous movie ratings. But by combining data from lots of different sources, Hunch can predict it with 95% accuracy. You’ll be seeing Hunch taste functionality on a bunch of websites soon.

   1. Chris – once you get this taste functionality up and running – will I as a user have to go to the Hunch website – or will you be integrating these features into other sites that I might already be on. For example, if I am walking in the West Village and am looking for a restaurant, I can go to Zagat, or Yelp, or any number of other sites – but what would be fantastic would be to have the functionality built in so that if I give you permission, you scrape all of my check ins and any other data you can collect – and use your algorithms to determine the best reference set for me personally – and thus narrow down the recommendations. Thus, I might end up in a much smaller quieter place that was equally reviewed than a more crowded noisy place that was very popular.

   2. you know that i am a big fan of getting Hunch on a lot of web servicesbut i also think that actions speak louder than wordsyou have to get to what people actually do in addition to what they say

      1. I think this is a very logical extension of “Deals Nearby.” As the rewards ecosystem grows on the Foursquare platform it would be great to see these deals served up based on my taste graph and not just based on my proximity.It would be a fantastic problem for Foursquare to have… but at what point do check-ins start having lower correlations to actual spending / tastes? i.e. as more rewards programs get launched on the 4sq platform, am I going to be more incentivized to check-in to simply earn loyalty points (instead of actually spending money)?Could be really interesting to see mashups using the Foursquare API and transaction APIs like Yodlee or Mastercard’s new program…

      2. This is exactly what we saw when building recommendations at Amazon. Recs driven by purchases were very compelling, while recs driven by page views were terrible (and at times accidentally controversial). User-based / Taste profile-based solutions were avoided because of the cold start problem – you were forcing new users (and almost everyone was a new user) to fill out a 10-20 question survey before they could get any results, a huge amount of friction. Item-based recs provided relevant results with only a single data point.Where this approach breaks down is when you are trying to recommend products witha very short shelf-life. Apparel, for example, goes out of style after only a few months, so all of the purchase data you’ve collected is suddenly a lot less valuable. (The fact that purchase data is super sparse thanks to so many variations in color, size, cut, etc just makes the problem worse).I’d be really interested to see if Hunch could solve Etsy’s search / product recommendation problems.

   3. Kind of interesting- I am a democrat and I do like vegan food, and occasionally I go for a vintage buy. Gah. I get the feeling over time, Hunch would just creep me out.

      1. I think it would be a rude awakening for many people. Too big brotherish to have a piece of technology reveal that you really aren’t all that unique and special. Kind of like the kids who all dress goth to express their individuality (actually to belong). Might slow the pace of adoption if key typical early adopters are turned off.

5. I wrote a piece on this in response to the launch of Facebook Places. The context was could Facebook Places overtake Yelp as a source of recommendations. My bet is yes. The same goes for anyone that gets enough sufficient check in data.The key here is volume of data. I love Yelp and use Yelp pretty much every time I’m looking for a recommendation. But it’s got a few issues:* Only a very tiny proportion of people write reviews. I used to be a heavy reviewer, but as Yelp has gotten more popular, I review less. (What’s the incremental value of me writing the 426th review of a place?) * The reviewers aren’t my real friends or people whose tastes I can interpret. Even among my own friends, I will interpret recommendations differently. There are some people with whom I have a negative taste relationship.Besides all of that, it’s a lot of work to read through reviews of multiple restaurants. Usually I’m looking for a good solution not the perfect solution.With check in data, you can get a lot more volume just because the LOE to produce it is a lot lower. (Like there are more tweets than blog posts.) While each nugget isn’t as data rich, with enough nuggets you can make inferences.The full post is here: http://blog.agrawals.org/20…

   1. cool. i will go read it. thanks for the link

6. Foursquare’s current API makes it very difficult to pull this sort of thing off right now…mostly because of privacy issues and what data gets exposed…so basically you can only get your own and some friends data (you could argue that friends data is more useful for this sort of service, but in reality the beauty of something like last.fm is it can find strangers at scale that like music just like me and it’s that scale that makes the algorithm work/useful).But again there are privacy issues that need to be figured out with foursquare before any of this can really be done via their API (though internally they could def. do some interesting things like this)…What they would need to expose is more of a firehose of checkins (I think they could do this without sharing specific usernames by just giving internal ids to accounts that aren’t easily/directly mapped back to users via any other API calls…so anyone could grab the checkins for user xyz even though they don’t actually know who user xyz is)…So IMHO, until this sort of data is exposed…it’s really only companies like Foursquare (or Facebook) that can pull off a last.fm for places…(in the meantime stuff like my http://hotaround.me service can only poke around this general idea but can’t nail it perfectly) 🙂

   1. thanks for explaining that Kevinvery helpful and useful information

      1. One of the things I’ve bugged Naveen for in the past is better access to friends’ history via the API.The Foursquare-piggybacked app I want is one that maps my friends’ historical checkins — basically treating them as implicit reviews. I’m rarely in midtown these days, but I was there this morning and wanted a cup of coffee after an early meeting; if I could see a map of where Alex Lines, Andy Weissman, and Mike White have checked in around midtown, I’ve got a list of places I can likely find a good cup of coffee. [Last time I checked, you could only access a short history of your friends’ checkins through the API, so you’d need to build up a checkin DB outside of Foursquare to do something like this.]Oddly (or perhaps not), you’re close to the three companies that fascinate me most with regards to API potential: Foursquare, Disqus, and Tumblr. They’ve all made a lot of useful data and functionality available, but they also all frustrate me at times with what *isn’t* possible via the API. In all cases, the issue is not that the APIs are bad, rather that the potential is so significant that it’s tough to keep up with it.Chris Dixon and others have been writing recently on the intersection of biz dev and APIs, which reminded me of a conversation I had a couple of months ago with a wise and sage gentleman: he tossed out the idea that APIs may benefit from being viewed as products in and of themselves, very much complementary to — but distinct from — the core product. It’s an interesting thought to play with.

         1. with foursquare i am fairly certain the issue is privacy and their need tobe careful about thatbut i agree about all three companiesthere is so much valuable data in those three services

7. The great thing about Last.fm is that your tastes are recorded without any effort (no “checkin” required). I’ve been using the Latitudie iPhone app to record my location in Google Latitude. It runs in the background all day long. I’m not interested in sharing that data publicly, but I keep thinking Google (or Apple) could create a killer app by aggregating anonymous location data from strangers and use it to power recommendations like this… without the need to checkin.I would also checkin to Foursquare a lot more frequently if they offered recommendations.

   1. yupi think reminding people to checkin is a big opportunity

8. This may not be the place for this post – but what the heck – since we are talking about Foursquare.I just got back from a vacation with the family to Hawaii and Northern California.I checked in when I remembered to – which is not always (can Foursquare write an extension that takes voice commands form the iPhone – double click the button and say check-in?) and now that I am back home sitting in front of the computer – there is a lot of opportunity for Foursquare in knowing where I’ve been.For example – they could team up with Trip Advisor to prompt me to review all of the hotels and restaurants I’ve checked in at (I’ve always been terrible about doing that – and I use Trip Advisor enough that I feel I should give back) Foursquare could then take and display some of the data from Trip Advisor in its tips section.Same could be said for Yelp – or any other location based service.Why not have Foursquare be the engine that drives customer service reviews back to the establishment that you checked in at.I stayed at the Rosewood Sand Hill Road hotel in California. Great place. They just sent me a customer satisfaction survey – which I can fill out for them – but wouldn’t it be more meaningful for everybody if that survey were essentially public – with all the data collated?Maybe I’m technologically ignorant – but I would love it if I could attach a picture to my foursquare check-in. Delicious porchetta at Madera is a good check in – but the truth is, I’d already checked in to the restaurant before I ordered, and I see no easy way to add a photo to my check in – or revise the check-in without repeating it.Just thoughtsAm I missing something?

   1. you aren’t missing anything harryyou rarely dothat’s why i was posting the maps of my foursquare checkins from my time in europesame basic desirei think this is important

9. Curious, nobody seems to be giving Yelp the chance/encouragement to do this, whereas it has a platform with far greater data to mine and in-place reach. They have already added FB open graph (so you get to see what your friends have written), they have created check-in (rather than a new food app, if we all agree Yelp is currently the most informative), and they are working on offers. They also seem to be doing very useful stuff with rating trends and context not to mention crowd-rated photos, which hasn’t been mentioned but is very useful content.I’m 100% for the power of small teams and API leverage, but I don’t hear anything that seems to create new value that can’t be more effectively created on Yelp, given their active extensions…What am I missing?

   1. the funny thing is the original taste neighbors post was sort of all about yelp

10. Love how Fred gets behind his portfolio companies to the point where he seeds ideas to others which effectively reaffirms the companies position as the centre of the universe for their industry.The question is which came first -Did investing in 4SQ lead to him deciphering the location market and then work out how to position the pieces to put them in the centerOrDid he have it all sussed out from the beginning and then look for a company and a team to put in the center of it….over to you.

    1. a little of bothi wrote the taste neighbors post three years agoback then i thought wesabe would be central to solving thisthat didn’t happennow i think foursquare might be

       1. Curious Fred – I would think a service like Wesabe that gets all of your transaction data would be able to solve this too … why did that not happen?I’m asking b/c I’ve been noodling on an idea that is similar in that it allows a consumer to aggregate their purchase history and take control over their purchasing power for (1) rewards & incentives from companies, (2) better service treatment, and [what you are referring to in this post …. ](3) recommendations & influence in common interest communities. All is demonstrated by purchase behavior (what I did, not what I say).I’m an enterprise apps entrepreneur by background and the inspiration for thinking about this comes from the perpetual multi-decade discussion about providing companies with the ultimate customer database so they get one view of their customer. For a lot of reasons this is never going to happen in and it occurred to me the consumer can solve this problem for themselves now with a clever web service + mobile app and then use the data to assert their own influence when needed AND get great recommendations.Blippy and Swipely are doing some of this in aggregating purchase data, but not in the value prop of leveraging purchase power for incentives, service and referrals/recommendations. Do you see any candidates in this arena that have promise?

          1. wesabe wasn’t adopted by enough users

    2. I think the Pandora comment makes sense, but I would “decode the genome” based on the restaurant’s “user experience” — not just the food ingredients… eg: loud/quiet, dark/light, high-traffic neighborhood/isolated, spicy food, etc…you get the picture.

11. Have not seen this “wisdom of the crowds” type personalization in local services, but we sure are working on it with regard to what content consumers might want to see. (www.verticalacuity.com) With all the focus on localization, its pretty certain this predictive local recommendations will be here shortly.

12. Fred – We are developing something like that in Montreal, Canada. We own http://www.restaurant.ca since 96 & Gastronomie magazine http://bit.ly/acazReIt's a lifestyle social network and management tool that will allow users to connect with “buddies” with similar affinities based on their 3-dimensional cultural & lifestyle profiling. It will be standalone and available through facebook & possibily foursquare. It even has a very nice non-advertising business model associated (sorry can’t same more at this moment)

13. Whrrl does something similar to taste neighboring but unfortunately not many people use the service and aside from having a (very) confusing UX for the iPhone, they hype up the points game way too much, in my opinion a feature that provides no benefit to the user and obfuscates the purpose of the application.A friends and location-based groupon app would have broader appeal, stoking followee narcissism while simultaneously rewarding the follower with financial incentives, something that has yet to be explored.

14. Hi Fred,Your initial blog post is awesome and I totally feel your pain.Interestingly, we’re pretty much doing EXACTLY what you’ve described – we are taking check-ins, using that to learn about our users and then giving them personalized recommendations for where to go next. If you check out http://www.getSpotOn.com/about you’ll see your blog post summarized (almost verbatim)!While we understand this could be a natural extension to foursquare, we feel that there is a LOT of merit to remaining platform agnostic as checkins can come from multiple sources and a foursquare recommendation engine would be, well, primarily for foursquare users.We’re still very much in product development phase. We’re gathering data primarily by allowing users to give us their foursquare checkins and will (very) soon be letting them give us their Facebook places checkins and other LBS checkins.We should have a demo-worthy product ready in a few weeks – could we demo to you then? We’re in NYC btw.Thanks

    1. Great comment: perfectly relevant, well-written, and it closes with a call to action.I hope your understandable desire to remain platform agnostic doesn’t lessen the chances of getting a meeting with Fred, but either way, keep plugging.

    2. i’d love to try your service once you feel it is ready for that

15. Foursquare could certainly benefit from further qualitative reviews where the data can be rolled up and searched for, by location, type of food, ratings, etc…But the issue remains one of slicing the social graph accordingly, and the Gotham Gal hit the nail on the head when she said: “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.” So, if we slice the results by my friends who I’ve tagged as foodies, then that would be more valuable, or even better, go across the “foodies out there” if such a segmentation accurately existed. Speaking of food, this sort of selective reviewing has been around for a while with Chowhound, eGullet and Opinionated About, but they don’t have the location-friendliness that a Foursquare or mobile phone App might have. An exception is the Michelin Guide (viamichelin.com) who does an outstanding job (for Europe primarily- note: data is scant outside of Western Europe) for pinpointing restaurants/hotels at the zip code level with great accuracy. Frankly, I don’t need any friend to tell me that such a restaurant is highly rated. If Michelin found it and rated it,- I trust them blindly, although I might read comments from recent visitors. Michelin has an iPhone app btw, but I’m not sure if its’ tied with location-based data & check-in where it’s all sync’ed in- that would be a killer. Example search on Michelin: http://www.viamichelin.com/web/Restaurants#resu…

    1. i don’t trust michelini’ve had a number of bad experiences

16. Hi Fred,I would say you almost have to link the choice someone makes to the checkin data. By that I mean use the fact that someone uses a service like Yelp to find somewhere, then Yelp ‘sees’ that same person has checked in from somewhere that they looked up earlier – this gives Yelp the chance to remind the person about that visit and possibly post a review or even just give it a mark out of 5.I think foursquare could also implement this with retailer specific review services – ie Starbucks can purchase the opportunity to ask 4sq users – say 1 in 1000 – how they’re last visit to a Starbucks was when they checkin from one. As to the side of the privacy line this is, well that would be a discussion for another day.As people have said, I think the pure fact that someone has checked in from a place a couple of times is not necessarily a pointer to a good review. It could be that they’re visiting the proprietor who they happen to know well, or it’s an easy place to eat on the way home from work. Neither point to the fact that it’s good or bad.The context of multiple checkins in one place is yet to be harnessed as a valuable piece of info, yet it could be extremely powerful if privacy concerns can be addressed.Mat

17. I actually built something like this during TechCrunch HackDay a few months back. My initial prototype uses a combination of Foursquare Check-ins and Yelp data. I think the most interesting thing about this problem is that there’s so many ways that it can be approached or weighted. Thanks for great post and reminder to continue on my explorations of the idea!

18. You’ve brought up a great topic Fred, and there’s some great comments and interesting looking companies in the comments.This is a problem I’ve been thinking a lot about. As many have pointed out, check-ins alone are a worthy signal (and like your delicious example could be generally construed to be positive votes), but I don’t think they are enough input alone to be able to build great personalized local recommendations. It definitely would be interesting to see recommendations based on Foursquare check-ins alone, but I think a better solution can be built that uses check-ins as one of many key inputs. This may be why a company that isn’t already so tied to one type of data becomes a strong entrant.Hunch has started to piece multiple things together and done a nice job, but I think we’re still very early in the recommendation era of the web.We’ll see more players, and people who specialize in specific areas (search and vertical search comes to mind). Also similiar to search, if recommendations are the next version of search as I’ve seen many say, is there a next generation of monetization to go along with it? Let’s call it “recommendation monetization”. Sure, there’s the natural and obvious steps of monetizing a recommendation through an affiliate relationship, but I think there may be some new and interesting things we see in that space that makes it different based on the difference in mindset and actions between getting a recommendation and searching for something more specific on a search engine.There’s another interesting dynamic to doing dining/travel recommendations that I haven’t seen pop up in the comment thread that I think is a key ingredient to building a winning product in this space. I’ll save it for now as I’m working on it….

19. Fred,I’m a co-founder at map.pr, where we are working on this very idea, and just put out a new version of our iPhone app that introduces our take on things. We use Foursquare (and soon Facebook et al) checkins and tips as signals for ranking places (with a great map-based UI). But to capture even more high quality signals, we have brought in a group concept, which lets users add a group to their checkin, effectively tagging a place for that group. I like to think of it as hashtags for Foursquare. Anyone can create a group around their interests, and we let you discover active nearby groups easily.My quick demo of mappr highlights how this works, this link takes you directly to that point in the video:http://www.youtube.com/watch?v=cU5pL8lvWcE#t=03…We hope to get to the type of collaborative filter you talk about, but as mentioned elsewhere, the API makes it difficult to get much of the data needed to do this really well. We think the added signal in groups can do great things right away, especially for finding places catering to more niche interests.We will soon extend to the web some great features, so for example if you had checked in with your “European Vacation” group, you could then have an embeddable map of all those places, with tips specific to that group. And yes, mobile web version also coming, supporting all you Android users.

20. Fred — I think the most important aspect of what you’re talking about is that you can get implicit social recommendations without the social group having to take any actions. No writing a review, no rating something, no nothing. Just based on behavior you can at least get items to put into a discovery set and then layer in external qualitative data to make that discovery more robust. There are some issues in addressing privacy and permissions with all the social APIs (not just four square, but you need to add in facebook, twitter, brightkite, gowalla, flickr — anything that has location tagging), but there are ways to work within a framework that maintains the privacy of the end users. I’m going to have some of our mutual contacts at Cooley reach out to you later this week if this is something you want to look at in greater detail.

21. I am maybe missing something but I do not see how a foursquare check-in makes it a recommendation. Why not a yelp or Zagat restaurant page visit? I might check-in into a restaurant with Foursquare and hate the experience or become the mayor of a place just because it is close to my office. On the other hand, I might avoid checking into a great place where I’m having a romantic date, just to avoid being spotted. There are so many other elements that make recommandation relevant to me: means (can I afford the place?), lifestyle (am I into great chefs or great steaks?), ambiance (what type of atmosphere am I in the mood for?) etc. I don’t see how Foursquare or even Hunch for that matter can answer these questions with an appropriate recommendation.

    1. Michel,I think one of the things that makes the check-in so interesting is the upending of the traditional ratio of Power user / casual user / lurker (as Mark Suster recently wrote about http://www.bothsidesoftheta…, which typically breaks down to 1% / 9% / 90%). With the checkin, every single discrete checkin is a signal, so the signal to unique user ratio is very high indeed.Of course, qualitatively, a Yelp review is worth much more than a Foursquare checkin, but taken as a whole, checkins (plus tips, a sort of lightweight review) can grow to be a very powerful set of signals to be mined.The checkin + your social graph can also be much simpler to model algorithmically, and as such is our first tool in surfacing interesting places. Moreover, I think we will see ways these can be built into simple checkins (I’m of course partial to the groups concept we are developing, described in a separate comment I made here…).Then adding more editorial or long form UGC (Yelp, Zagat) can help confirm to the user that the checkin recommendations surfaced indeed to seem worthwhile.

    2. why would you go to a place you hate?my checkin history is almost a perfect representation of the places i love

22. Fred, very interesting post. I think you’re right, any form of “useful information discovery” presents a huge business opportunity and specially, when it reaches critical mass as in foursquare, it becomes hugely timely.You might find the work form the Sense Networks [1] guys interesting, they’re working on “implicit checkins” and that’s where I think it gets really interesting.[1] http://www.sensenetworks.com/

23. Hi Fred – I saw Chris Dixon’s tweet earlier this week (http://twitter.com/cdixon/s… and wanted to let you know that there’s at least one more company out there that owes a debt to you and your Taste Neighbors meme. We started AppStoreHQ (http://www.appstorehq.com) to apply the insight to smartphone apps and have learned a ton from Last.fm in particular. Since you’re an Android user now we also have a native app in the works that carries the idea much farther than we can on Web data alone – we’d love to add you to the test group if you’re game…

    1. i’d love toapp recommendations is a big issue

24. (Another) great post! Completely agree that when designing a recommendation engine, it’s important to incorporate not only users’ explicit preferences, like thumb-ups or check-ins, but also users’ implicit preferences. E.g., StumbleUpon’s content recommendation engine considers not only users’ thumb-ups or thumb-downs but also time of day, media type, dwell time, sharing and commenting activity, semantic patterns, and dozens of other factors – and re-mixes the relative importance of each of these factors on a user-by-user basis. If one has enough feedback from enough individuals, one can analyze these signals across clusters of algorithmically determined like-minded users to predict with a high degree of relevance what else a given user would like. It’s possible that generating high-quality recommendations against a more structured, finite data set (e.g., locations, things) may not require as expansive a set of inputs and as much computing power. At StumbleUpon, however, we’ve found that successfully helping users serendipitously discover new material they wouldn’t have predicted or found through traditional search offerings is only possible with an approach like ours.Marc Leibowitz, VP – Partnerships, Sales & Marketing, StumbleUpon

25. I agree that something which can “recommend” similar places that you might like is a great idea, but it has been implemented (with varying degrees of success), at various other “review” sites. One of my favorite sites, Qype, does it quite well and it seems fairly accurate. Looking through the list of recommended restaurants for me, it throws up many that I have gone to already (but have not reviewed), and also many that I am wanting to go to. It does, however, not go into enough detail to calculate the recommendations.For example, it would be good if it could factor in distance. I live close to central London, but it throws up restaurants about an hour commute on the other side of London, which I would hesitate to make the long trip for. It doesn’t take into consideration the price range which I like, throwing up Michelin star restaurants and really cheap places instead of more middle-of-the-road.

    1. Good points and I agree it isn’t perfect yet!It should be better with the new Qype app on the iphone as it tries to provide recommendations near to the user’s current location. On the website, things could certainly be improved although one can narrow down the location via filters to get recommendations only within a local area (e.g. West End and so on).Regarding the price sensitive recommendations, there is definitely room for improvement here and it’s something on our radar. Price is a tricky subject though because it varies from occasion to occasion, so it’s not just about taste, but also intent and timing.If you have other ideas or suggestions, drop us a line!

26. Hi Fred,Great idea here, and you were clearly spot on already a few summers ago.As pointed out in the comments above, there are many challenges in doing really good recommendations engines: the type and volume of data, the diversity and amount of users and also the scalability and performance of the system. I just wanted to add that in our experience, for companies like Foursquare, Yelp and other major sites (eg online travel sites) that sit on huge amounts of data, it primarily comes down to 1-understanding the type of personalization that will deliver the best results based on their service and the data they already have and 2-planning for the data they should start gathering to improve personalization services in the future. We think of personalization as a journey, not a single feature. After all, Amazon’s been working on recommendations for 10 years and is still making improvements…That says something.Our company, http://www.LikeCube.com, is expert in personalization, recommendation and discovery solutions for the travel and leisure online and mobile markets. In addition to helping our customers defining their personalization strategy, we also enable exactly what you are talking about, as a B2B offering. At SXSW in March, we launched our new checkin based recommendation solution where it received a lot of attention. It allows predicting the likelihood a user will checkin in a place, recommending places a user might like to go to and finding taste neighbours of that user (eg to filter tips). We can also use a combination of user data (ratings, reviews, bookings and wish-lists) as well as semantic data to create personalized experiences that will make a real difference in terms of trust and loyalty to a service. LikeCube powered recommendations were tagged as “quite spooky, bang on the money and an incentive to join”. These comments are from users of http://www.qype.com, the biggest local business review site in Europe (equivalent of Yelp in Europe), with over 16 million monthly uniques and powered by LikeCube.It would be great to further the discussion with you. If you wish, we can arrange for a demo.

27. One main problem, that has been touched on in several comments, is that meals are not all eaten in the same context. I love The Breslin in New York and have recommended many times to my friends when they go to New York. (Marcoullier loved it) But it takes times to get in and it is a big meal. If you are not willing to wait or not into pig, I would probably not recommend it. (Music is often sampled over and over again in a similar context – not as large of an issue. )I see this issue when trying to use Yelp. Having a small winery, I end up traveling to my vineyards often during harvest. I often take my 5 year daughter. We use Yelp to help pick places to eat – the success ratio is very low. The children filter eliminates nearly every place we would consider so we stopped using. In Sacramento, we have been disappointed with nearly every place we have tried using Yelp.One other factor to consider would the power of the recommendations. There is some fascinating research on wine and tasting that has been done in Italy and also replicated by Cal Tech. When you tell a wine drinker that what they are about to taste is great, the neural network paths are different then when you tell them it is okay or cheap. The expectation of something great causes the brain to process the information in a different manner.All of that said, I think it is a great idea. Time of day, frequency of check-in etc can provide some of the required context.If a stand alone app was built using the FourSquare API, I would be willing to answer 2 or 3 binary/multiple choice questions about a place I was dining. Thumbs up/thumbs down and perhaps questions to better understand my context for the rating. For example, price could be easily established with a few users answering it. As could noise level etc.I do hope someone builds it.

    1. me toothe best meal i had this year was at The Breslinapril bloomfield is an amazing chefi had a meal at the Spotted Pig (another of her restaurants) a couple nightsago that was also fantastic

28. Great topic and ideas, Fred and others. I’m working on something along these lines in the Bay Area. I think the combo of explicit and implicit data is most powerful, and building trust with the user is critical to getting explicit data (and sometimes access to implicit data, too, e.g. credit card account info). Building trust over time will enable greater access to data.

29. Hi Fred, great post and liking the active discussion that has followed. With enough data, I think inferences can be valid. For example, if a place is checked into primarily by people with families, you can assume it’s kid friendly. (If you were using FB data, you’d potentially know this.)I realized the opportunity was much larger than just dating and built version 2. It’s looking awesome.thanks for sharing…

30. Hi Fred, this is essentially what Rummble was created to do – as we discussed when we met last Summer. Our challenge has been to provide a UI and first user experience which is not only simple enough but also communicates the difference between our trust graph and that of regular recommendations. New versions of our website and apps are on their way – I’ll ping you when they’re done.

31. I and my friends use tips for the like/dislike function. It’s never enough with a location to just mark it, I want to know why, and what to do there.

32. about five or six years ago now, i asked Joshua Schachter to think about adding ratings to delicioushe replied “we have ratingsif you post a link to delicious, you are voting for itif you don’t, you are not”i feel the same way about a checkini’ve checked into Joseph Leonard something like 30 timesi think you can tell that i really enjoy eating there

33. Hi Fred, great post and liking the active discussion that has followed.The question boils down to what inputs are required to form a great recommendation engine and how important is cross-category data for any given vertical, publisher or specific search. Who builds it, push vs pull, etc, I believe, are all secondary until we solve the question of what inputs a system has at its disposal.What’s missing here is passive or active user intent as it specifically relates to deciphering tastes, interests and preferences. I could bookmark an article on delicious because i love it and want to save it or because I want to remember to read it for research. The prior is effectively a “like”, but the latter is similar to a single foursquare check-in … an action with unknown interest coordinates. I check in to Pastis, but it was unclear whether I enjoyed it or not. A check-out feature with a simple rating would do wonders for taste data as a node to connect a user to a venue. Check-in data alone, even with velocity and volume, will likely yield recommendations, but likely with marginal value. Add in the rating system and now we are talking positive interest signals.Also, building a recommendation system in-house is very difficult to do. Even some of the largest ecommerce companies in the world have decided to outsource the function after years of trying to build it. The good news is that Dennis has recruited excellent talent at 4sq. The big question, which we will only find out over time, is whether or not 4sq’s silo’d data will be sufficient enough to yield the necessary inputs to recommend venues with a high enough degree of accurately. This becomes a data question and not an expertise/talent discussion. To Chris Dixon’s point, the Napolean Dynamite Syndrome is not only real, but scientifically proven.Thanks again Fred!

34. I concur. Voting with your feet works.And our experience at LikeCube is that the amount of checkins that services like Foursquare has can be sufficient to start delivering meaningful personalized recommendations and compute taste neighbours.Of course, adding ratings, reviews and buying history on top will increase the accuracy and effectiveness of the solution.

35. Great point. And that is a mighty fine restaurant!

36. I love Pierre Gagnaire too. Then, if you like it, why make your search so complicated,- go to his website and you’ll know he’s got Sketch in London, another one in Hong Kong, and look at the Michelin guides if you’re interested at that level. I think the issue that the Foursquare’s can solve is not for the top 1000 restaurants in the world which are well known and documented- it’s the next 10,000 or so that are very decent but more difficult to find out about.

37. I think the explicit recommendation is pretty important. I go to some great places with my family that if I could, I would frequent but I cant afford it. needless to say, my check-in count would be low.we’ve been trying to tackle this problem for a bit. We’re making the assumptions that the implicit information carried by a personal relationship has vastly more value than any sort of document able data. Our approach has been to let you use attributes of the recommenders as filters/facets until you build up your social graph.Here is an example of what we’re trying to do “Recommendations for restaurant near new york within 3.75 miles from users whose occupation is Technology consultant or Architectural designer or Graphic designer and live in New york or Brooklyn and are interested in design or startups or technology.” Here is the link for that search http://wby.im/diVz8pIf you click the link, try playing around with the facets on the right.