I’ve recently decided to make an entrepreneurial bet on the belief that there is high value in being able to provide recommendations on the future performance of Content – in particular for news media companies!  This might sound like any other retail recommendation firm out there, but I wanted to draw attention to the complexity of doing Predictive Analytics for Media versus that of doing Predictive Analytics for Retail. Largely a task (and ability) of analyzing and modeling on current and historical data (facts) to make predictions about a future outcome.

The difference becomes clear as we think of the following three time intervals; the past, now and the future. If you look at the following two items, It becomes almost obvious that the time interval is dramatically different.

When an Article is born, in particular in news media, the value of it decreases from that moment on forward – and not only that, the lifespan is exceptionally short. Compare this to a retail product where you don’t necessarily have to promote the product instantly and reap in the cost over e.g. 10 hours.

This provides an interesting challenge when doing predictive analytics for media, which is that you have to provide a first set of recommendations very rapidly and on very little training data. Take the example from above; one would have to push the off-duty-cop story instantly while the value is high, knowing that there is little or no past data. You have to push the Article aggressively (with accurate recommendation on where to promote it) in real-time, knowing that e.g. 10 hours later, the story dies and any data value in the last 10 hours disappears (note: there is of course a long tail play in content, which benefits from past data). In Retail you have a plethora of past data, and the present (now) is important, but not necessarily overly important compared to 1 hour from now, and finally, there is a long and prosper future ahead, where we have plenty of past data to work with. A simple illustration of this challenge could look like this:

This concept of time abundance in retail optimization, explains why you can allow yourself to do landing page optimization on products pages, but rarely on individual article pages (layout yes, content no) etc.

In conclusion, if you want to provide recommendations on how well content is going to perform in the future, you have to accept that you have very little data to model on initially and that the recommendation itself expires (decrease in value) rapidly. Knowing which Article should have been pushed on the Front Page of the New York Times Yesterday has little or no value today!

The Visual Revenue Recommendation Platform – takes this into consideration and I would love to chat more about this subject if you are a Publisher in the need of knowing your future article performance.

Cheers :-)
http://about.me/dennismortensen (cool about.me page tool from @TonySphere)