Is Usable Discovery The Holy Grail?

Over the past few days we have seen Twitter get criticized heavily over the #fixreplies fiasco. Much of the complaining was due to the communication of the changes, or the lack of timely communication. However, there was another large group of voices that used the @replies as a way of discovering other interesting users. This is important because there really is no way to find new users on the service.

We also saw a new recommendation widget for Google Friend Connect. I know these may not seem related, but the Google Social Web blog has a good quote that may help:

You spend a lot of time creating great content for your site, but are you ever curious about which parts of your website your community likes best?

OK, the connection may not be obvious to many people, but what is recommendation? Recommendation is really just a form of discovery that has additional variables. For the recommendation widget, users may be able to “discover” new content on your blog because other people have liked it. The idea here is that the users in the community have similar tastes, given they are reading the same blog, and probably like some of the same posts.

Twitter highlighted the fact that their @reply feature did not work in the manner that they wanted, and will likely bring the concept back in a new feature. This is important for Twitter as they need a discovery feature of some sort because Facebook and FriendFeed do have this in some manner. Facebook has a sidebar widget of “people you may know” based on your demographic information and your current network of friends. The widget works well enough, but it could be much better. FriendFeed uses the “friend of a friend” (FOAF) model, where you may see content that your friends liked and you are not currently subscribed to. This also works fairly well, but sometimes there can be a lot of noise in the FOAF data.

Blog search is another area where discovery has been woefully lacking. Technorati tried with “fans” of blogs and the tags. The tags generally work for categorization, but there is little help in discovering new blogs or content.

On the recommendation side, we have Amazon. Their recommendation feature has made them a lot of money. However, if you have ever bought gifts for people, your recommendations start to skew in odd directions. Amazon has taken steps to avoid this, like not using a purchase for recommendations, but fine tuning the recommendations requires a lot of user input.

So, what is the problem? First, real discovery is hard. Amazon has been doing this for years and still has some problems. So, newer services will likely have problems with it as well. Discovery also takes knowledge of advanced topics like statistics and machine learning. People well versed in these topics typically are not well versed in usability. The other side of the discovery problem is time. Typically, discovery takes a lot of time and CPU cycles. We are starting to see a lot of data in the social media space, which means it is ripe for mining and applying discovery. Some of the newer technologies like BigTable, Hadoop and many of the cloud related initiatives could help us along the way. I have a feeling we could be primed for a breakthrough in discovery, and I wish I knew who would provide it.

14 thoughts on “Is Usable Discovery The Holy Grail?

  1. I think the problem with usable discovery is not only does it have difficulty discerning my interests from my actions, but it doesn’t which of my past actions are related to my current concerns.

    There needs to be some way for the searcher setting parameters themselves for usable discovery to be useful.

    Case in point: Purchases on Amazon does not work as a discovery mechanism because the fact that I regularly buy children’s books for family members does not mean I have an interest in seeing information about children’s books except near their birthdays and Christmas.

    For that matter I don’t want to see web design books in my recommendations when I’m looking a new novel to read.

    Similary with Facebook. I am glad to see what my nephew is up to, but I don’t want to have to wade through every child in the Barrington, RI school system to see if there are friends of friends that I might want to connect to.

    Like

  2. Daniel,

    Your Amazon and Facebook examples are exactly what I am talking about. Amazon has been trying to allow people to refine their interests and recommendations, but it takes a decent amount of work for the user. I like to assume that users are generally lazy or busy, meaning they will not take much time to do something.

    Facebook attacked it from a different angle, with the location and school networks. That fails as well because you don’t know everyone in the area.

    Like

  3. Jean-Marc

    The signal to noise ratio is important, but what if the stream of information is noisy? Using Daniel’s example of buying children’s books as gifts, what if he has several nieces and nephews? The number of children’s books could be larger than other purchases, and thus skew the signal. “False positives” like this make the problem domain very difficult to do very well.

    Like

  4. I agree that noise comes in different qualities – the engine rumble that ruins your conversation is less interesting than that conversation overheard in a party or that book you stumbled upon while looking for another.

    Rather than noise, we should probably talk about serendipitous discovery – an old concept that Elaine G. Toms was to my knowledge the first to put in the context of computer supported information systems. Here is the conclusion of her 2000 “Serendipitous Information Retrieval” paper found at http://www.ercim.org/publication/ws-proceedings/DelNoe01/3_Toms.pdf

    “While significant evidence exist to support the value of serendipitous experiences, few information systems support such an method. Yet, their value is generally assumed. Serendipitous retrieval demands approaching information retrieval in an unorthodox manner, one that does not tightly couple the explicit match of query with result, but instead takes a fuzzy approach to the problem”.

    Like

  5. Jean-Marc,

    I do not want to get too deep into this, but the paper is interesting and is close to the “discovery” that most people talk about. Using methods like “k nearest neighbor” or even naive bayes methods can get closer to the discovery that many of us desire. Fuzzy approaches are the “next step” after the standard machine learning methods fail. Many of the methods from my gentle introduction to data mining (https://regulargeek.com/2009/05/01/a-really-gentle-introduction-to-data-mining/) can be used for discovery. This is also where the applications can get really interesting.

    Like

  6. I love the idea of serendipitous search. To some degree you can see how that might work at a site like the NY Times. I am often pulled to other articles by the 10 most e-mail/blogged box.

    I haven’t looked at Amazon recommendations recently so I don’t know if the format has changed, but it strikes me that one of the problems with their recommendation feature was one of categorization. Novels were interwoven with children’s books and web design books. It would have been far more useful if I can have gone done the branch I was interested in (or even suppressed a whole category).

    NetFlix has had far better results with its recommendation system at least as far as I am concerned. It had the feature of showing people with similar lists with you and giving the percentage of overlap. You could then go look at that person’s list and see what else they had on the list that might be of interest.

    I think the one limitation of Netflix’s viewer matching system was a failure to weight it in any way. My guess would be that the fact that I shared an interest with the someone in a highly popular film like “Star Wars” or “Big” is of less significance than the fact that we shared an interest in an unusual film like “Saragosso Manuscript” or “The Visitor”.

    Like

  7. Daniel

    Netflix is really a pioneer in the recommendation arena. They have their competition to improve their algorithm which has received a lot of attention. I am not sure how I forgot about them.

    Like

  8. Bryant

    I loved your overview of the Netflix prize competition. Napolean Dynamite is a great example of how recommender systems “get confused”. One day I may try something for the netflix prize, but it takes time, and I don’t have much of that.

    Like

Comments are closed.