The personalized music service Pandora is a huge hit in my family, where almost everyone has his or her own custom "channel" of Internet radio. Over time, Pandora learns the kinds of tunes we like -- and don't like. It requires minimal and only occasional interactivity that amounts to voting a thumbs-up or thumbs-down on whether the company's music personalization algorithms match our tastes properly.
For the most part, I can lean back and let it play a nice collection of tracks I already know and love, and a good selection of stuff I didn't know but come to like. This approach, a combination of Pandora's internal "music genome" classification engine and my own occasional interactions, comes closer to the ideal of digital personalization than most models. It learns and reacts to my taste but also has an element of serendipity and true discovery built in.
The builders of the Worio engine content recommendation engine I wrote about last year are trying something like a Pandora model for content. The original Worio model paired a search engine with content recommendations. Now they are working on the Zite project (named after "zeitgeist") that tries to leverage social networks and at the same time get over the limitations of relying on social media solely as a recommendation tool.
As the team explained the model in a recent press preview, media sharing platforms like dig and Twitter are limited by the community itself, and what your friends and like-minded types are reading. Zite takes a broader view of content gathering news related to the topics you are interested in from around the Web. They call it "A personal view of things that are in the zeitgeist."
The system is in private beta, but the company is opening it to the first 50 readers of this column at http://zite.com/?code=minonline
When you start, the system asks for your Twitter address so it can scan your stream for clues about your interests. Basic to both Worio and Zite's personalization is the idea that users don't want or need to be too proactive in shaping their own profiles if the system is smart enough to read their tastes and behaviors. And you need a starting point rather than building a profile from zero. The content aggregation engine crawls for connections among the content pieces and uses semantic analysis to see how topics overlap and connect to one another.
The interface looks very much like a classic RSS reader, except you highlight a set of topics rather than singular news sources. A series of newer headlines appear at the top half of the screen and more evergreen older content appears below. Mousing over the headline renders its source, as well as the volume of retweets and digs it has received.
According to Zite, the engine is surfacing and prioritizing content in part by looking at the conversations occurring about the topic and the content piece on blogs, links, comments, etc. It crawls over 100,000 blogs, for instance, to identify the topics of interest to you. It is trying to take a holistic view of zeitgeist by looking for the "high velocity" items that are then filtered according to the user's demonstrated interests. If a like-minded Twitter user tweets a link to a story, then that might the story up on your list.
Zite creates a "Follow" list of topic that you click on as you would an RSS source. And then it gives you a series of more precise topic suggestions to tap into. For my general "movie reviews" topics it offered a "parent" category of "movies" and a number of "siblings," including "movie new moon" and "movie box office" etc. And then it also suggests a set of terms based on formats like reviews, trailers and interviews. And there is even a set of topic suggestions just around story "tone," including "movie humor," "Movie fun," and "movie cool." At the same time it lets you drill into some of the most popular news sources.
Zite seems to work well in surfacing stories that are of interest, but of course there is no telling what is being filtered out and why. For instance, I am following "games" and get a list of a dozen or so stories that occurred over the last three days, surely not reflective of the volume of games-related content on the Web. But somewhere in there the system is choosing the hot topics and the ones it considers most relevant to me, though it is less clear why it made these choices.
A personal content filter for the real-time Web is a good idea, but ultimately it can leave one wondering what you may be missing in the cut of headlines the system is giving you and how you might tweak it to render more or fewer results.
That transparency is not an easy thing to achieve in an engine that is so complex in its variables. Compare it to the Apple Genius filter that is on an iPhone. The Apple handset has a Genius tab in the App Store that suggests more apps based on what is already on your phone. Apple makes the filter transparent by telling you it is recommending one title because of this other one that it names in the process. Zite, which seems to be considerably smarter than the iPhone "Genius," would have a harder time explaining its reasoning.
Arguably, the Zite engine uses this content filter really as a starting point for discovery, in that it gives you a topic taxonomy with which to drill into the real-time Web as you wish on your own. Still, it raises an interesting conundrum for personalization at such a high level. As the technology becomes sophisticated enough to perform such high-level information-filtering tasks (which we claim we want and need), it requires increasing levels of trust on our part in the "genius" in the machine.