12 Comments

Summary:

Hunch, a startup trying to build a “taste graph” of people’s like and dislikes that can act as a recommendation engine, has partnered with Gifts.com to make suggestions about what kinds of presents your Facebook friends might like, based on their Facebook profiles.

Hunch recommendations3x2

With the holidays approaching, we’re entering that time of year when desperate people grab things like ties, inappropriate books and goofy toys in a pathetic attempt to bring joy to their friends and loved ones. Can algorithms help? New York-based startup Hunch thinks they can, and Gifts.com, the shopping site that is part of the IAC empire apparently agrees. Gifts.com has been using Hunch’s recommendation engine for almost two weeks now, and the company says its conversion rate (i.e., the number of people who go from being shoppers to buyers) is as much as 60 percent higher than it was before the site started using the tool to make recommendations.

When you go to the Gifts.com section that’s using Hunch, you’re prompted to log in with your Facebook account, since Hunch uses your Facebook “social graph” or friend connections, to power the feature. When you log in, you see a list of your friends (including those who have birthdays coming up). Clicking on any of them brings up a suggested list of presents, along with the question “would this person like this?” and a button for yes and no. You can click those buttons to refine the search, or you can answer some Hunch questions to fine-tune the profile of that person.

Just as on Hunch’s website, the service asks questions ranging from the prosaic (does the person tend to vote liberal or conservative, etc.) to the somewhat bizarre — including “would this person think it’s wrong to train dolphins and keep them in captivity for aquatic shows?” and “Would this person believe that alien abductions are real or fake?” Hunch founder Chris Dixon says answers to those kinds of questions can indicate where people fall on other spectrums of likes and dislikes, and can be used to make guesses about what they might be interested in.

The alien abduction question, for example, “tends to correlate highly with political orientation,” Dixon says, and the answers that you provide to those kinds of questions allow Hunch to make “cross-category inferences” about whether you might like a certain sports team, or whether you like Italian food better than French cuisine. Gifts.com is one of the companies taking part in the rollout of Hunch’s recommendation platform, along with ShopStyle, Milo and FanBridge — and Hunch has also been talking to other companies, including some of the major cellular carriers.

The most obvious comparison to Hunch’s partnership with Gifts.com is Amazon’s gift-suggestion service, which also looks at your Facebook social graph in order to make recommendations — something Amazon launched in July. I haven’t used Amazon’s feature that much, but after trying it out a few times, I found it was just barely better than a guess, or a random pick from one of the site’s top most-recommended items. In many cases, it didn’t do any personalizing, because the friend’s profile didn’t have enough information about their likes and dislikes.

Dixon say this is one of the things Hunch’s database of correlations tries to do: namely, fill in the gaps or make connections between the things people like on Facebook and the things they might like elsewhere. It learns over time — Hunch’s recommendations, for example, seemed fairly off the mark for many of my friends, even after I did some training of the algorithm by answering questions, but Dixon said the overlap between the kinds of products at Gifts.com and the kind in Hunch’s database is not very large, and so the system is still learning to make connections between the two.

“I think it probably needs another two weeks or so before it becomes mature,” Dixon said. The good part of such relationships for Hunch, he says, is that the more data it pulls in about what people like or don’t like — the average user at Gifts.com is giving feedback on 20 items, which is higher than expected — the better the system gets at making recommendations. Maybe some day it will be better than just randomly picking up a pair of cufflinks or slippers for that person on your holiday list.

Related GigaOM Pro content (sub req’d):

Post feature image courtesy of Flickr user Randy Robertson

Related stories

  1. this science project is nearing the end of its life. still cannot figure out their business model. add to deadpool!

    Share
    1. Well as far as business models go, I assume they either are generating revenue from these partnerships or are planning to do so. I think there are lots of businesses that would like to provide recommendations in the same way, but can’t be bothered to create their own database or don’t have the wherewithal to do so.

      Share
  2. Worked pretty well for me, especially my regular non-blogging-foursquare-blippy.tv-using friends. Looks like the inventory of gifts.com is definitely targeted toward “normals”.

    Share
    1. Yes, that’s a good way of putting it, Jake.

      Share
  3. You, or bloggers, may not be the right demo for Gifts.com items. For personalized recommendations, it is partly dependent on the inventory of products that a site like gifts.com offers to determine how relevant the gifts will be for your friends. I mean, if you are looking on Sephora for a gift for your wife and you know that she doesn’t like perfume, you may have to realize that you are not the right demo for the site. The 60% conversion lift speak for itself, in my opinion. Gifts.com is a really sophisticated IAC property and they have optimized the site, so 60% increase in conversion means their demo of users must be seeing great value.

    Share
    1. I agree — that was the point I think Chris was trying to make as well. And you are right that 60 percent conversion growth means that users must be finding value there.

      Share
  4. are the recommendations more accurate … or is the simple act of having the shopper go through the process more likely to lead to converions

    there’s 2 simple ways to test it

    - a a/b comparison test against a different recommendation method on the same site … the Hunch method should produce results that are greater than the MOE vs. another method

    - increasing gains in conversion rate over time for the same set of users … as it learns and recommendations improve … that improvement should lead to increasing conversion … for example a recommendation at signup + 30 should have a greater conversion than one at signup + 5 on a linear time scale

    does it matter? … well yes … if its the process that causes the conversion increase, then you can isolate and maximize it … if its the Hunch, then that can be isolated and maximized … you ought to know whats driving the behavior

    Share
    1. That’s a good point, Eric — although in the end I suppose Gifts.com doesn’t care so long as engagement (and the conversion rate) increases.

      Share
      1. it does matter for hunch … if its the process, then anybody can duplicate it with a bit of work

        if its the accuracy of the algorithm, then its unique for now

        it also matters for the companies, if you don’t isolate the cause, then the results might not be sustainable and could change at any time

        Share
  5. I’ve recently launched a site with a similar idea, called WishGenies. The difference between Hunch, and WishGenies is that I realized that generating good recommendations in an automated way was going to be hit or miss without a lot of data. Instead, what I’m doing is asking other users with similar interests for recommendations. It’s a “takes one to know one” kind of approach. It’s more social (it’s a facebook app). You don’t get immediate results yet, as you do with hunch, but you do get results that are from other real users with shared interests.

    http://apps.facebook.com/wishgenies

    Share
  6. [...] then provides a short list of places that match a person’s interests. This is reminiscent of Hunch’s taste profile questionnaire, but it felt a little cumbersome having to write in the answers of my favorite places, which is [...]

    Share
  7. [...] is personalized for each person based on their preferences and tastes. Companies like Gravity and Hunch are also part of this wave, trying to take signals from users and build recommendations that are [...]

    Share

Comments have been disabled for this post