Which is not to say that personalized services have not thrived, but the success cases occur almost always outside of the realm of news and information. Amazon has hands-down the best personalized service online. It leverages the usage patterns of you and other buyers to create the closest thing to a concierge experience we have online. Netflix, while not as sophisticated as Amazon, effectively pushes relevant recommendations to DVD renters in a way that uncovers unknown gems and forgotten must-sees. But also in this mix is Pandora, which shapes Web radio stations by matching the characteristics of a track against your own declared taste and recent responses to previous tracks.
Amazon, Pandora, and Netflix have a simple personalization principle in common, says Eduardo Hauser, CEO of DailyMe: "They require very little effort or input but give a lot of output." The value delivered is wildly disproportionate to the effort required. Traditionally, this has been the missing element when it comes to personalizing news: real interest.
Hauser discovered this firsthand. DailyMe launched in July as a personalized aggregation of newspaper and wire content across all categories. DailyMe licenses content from 500 publishers to process 20,000 stories a day. Users registered with the service and declared their preferences in order to create a tab of custom headlines on the front page. But as Hauser admits, "performance has been less than we would like" -- about 100,000 uniques a month and 25,000 who registered and formed profiles.
The original model was based on "strict filtering," whereby a user asks for content type X and gets content type X. Even at this rudimentary level of personalization, however, DailyMe illustrated the clear benefits of personalization when it comes to engagement. "We discovered that people who went through the process of creating profiles are pretty good users. They will look at seven pages per visit and spend 45 minutes a month and come back three times a week." The problem was scale. "The number of people willing to give you that much information didn't scale very well," he says. "We have to do it differently."
While Amazon, Pandora and Netflix suggested a general principle for making personalization work (little input, big output), news content is different in kind from music, book or DVD shopping. First, news has an issue of "maturity" some of these other content types don't have. "News cannot be re-recommended until you buy it," says Hauser. Moreover, a personalization engine for real-time aggregated content never knows exactly what content will be in its inventory at any given time. And finally, news is consumed differently from these other media types. People flit across providers, and it is inherently difficult to get the user to invest time even to register with a site, let alone personalize it.
"The way to fix this is to engage in dynamic personalization," says Hauser. "We are going to deploy a system that will require very little input and give a lot of output, be suitable for news, learn from the user, and not require registrations."
Scheduled to launch in May, the next iteration of DailyMe will use a cookie-based approach to tracking incoming users ties to their email login. The engine will map a user's history onto a histogram or frequency distribution chart. The content will be cataloged from tagging at the source as well as DailyMe's own engine that will map 141 categories. A number of things go into the content analysis, from examining the number of images in a piece or even characters involved to determine its depth. A rank of sources will try to determine which of two similar pieces is coming from a more authoritative publisher on that particular topic.
The user's histogram will be populated by a series of data points that go far beyond the content types a user views. Hauser says that elements like time spent in categories, entry and exit pages, etc. will help profile the user's level of interest. Whether the user emailed or voted on a story exposes his level of engagement with the topic as well. "All of these data points are first normalized against the entire audience and then divided into time spans to distinguish between short terms of 3 to 7 days or longer term interest," says Hauser. An algorithm will make predictions about news likely to be of interest to particular users. The system will apply machine learning technologies from computer science to the task of building a profile on that user in background.
On the business side, what becomes interesting about this news personalization engine is that it can also serve as an ad personalization engine. The same data about usage, interests, etc., that is passed to the content engine for determining the editorial mix could also be passed on to the ad network.
Of course, creating a more passive personalization model that responds to behavior rather than profiling is only part of the larger problem. News is commoditized online and most of us either use a search engine or multiple sources. Expecting anyone to invest large amounts of time with any one general news source is outmoded. And so a secondary realization for DailyMe is that it had to become a technology for licensing to others and not rely on a destination model. The new cookie-based technology makes it possible now for the engine to be licensed out to publishing partners.
There has been a lot of discussion lately about how over-personalizing the media experience creates narrow ruts of interest and knowledge for us and robs us of serendipity and discovery. This is true, and Hauser is keeping that prospect in mind by looking for ways to make DailyMe a blended experience, an editorially driven home page and perhaps behaviorally driven article pages. Personalized news could occupy a box amid a standard news home page, in much the same way Forbes.com has its Attache customized news bar on the side of its site. In some sense, over-using past behaviors to predict future behaviors can actually undermine the users' attempts to move their news gathering habits online. Personalization without the opportunity for greater discovery over-solves one problem and helps create another.