The Cold Start Problem

People who bought this item also bought...

If you've ever shopped online, then you're familiar with recommendations that rely on collaborative filtering, predicting a user's interests from the past behavior of other users with similar interests. There are countless varieties of collaborative filtering -- we've come a long way since researchers at Xerox PARC proposed the method two decades ago.

But collaborative filtering depends on the availability of past data. What about recommendations for new products that people have not yet had a chance to purchase, let alone recommend?

Here we encounter what's known as the "cold-start" problem. For the product recommendation case, there's often a work-around -- namely, making inferences about new products based on the historical purchasing behavior of related products.

But there's a more extreme version of the cold-start problem: launching a new data product, like when LinkedIn first introduced the world to "People You May Know". For this scenario, there's no past data to draw on at all, at least in theory.

In practice, there's usually a way to bootstrap. We can apply what Monica Rogati calls "data recycling" -- using data from other contexts to bootstrap our initial statistical models until we can collect live data. We can also use an "explore / exploit" strategy to optimize the speed of learning.

Developing a creative solution for a cold-start problem is a great test of a data scientist's prowess -- it also makes for a great interview problem.

But cold-start problems aren't just for data scientists. They come up in everyday life.

For example, consider the human tendency to overgeneralize from first impressions. Or the reliance on stereotypes instead of evidence. Both are attempts to overcome cold-start problems. At one time they may have provided evolutionary advantages, but today they can be crippling cognitive biases.

Still, as human beings, we have to make decisions all the time, and we don't always have the luxury of obtaining the best decision from a model built with representative training data. So we should all get better at facing cold-start problems. At some level, we are all data scientists. So let's try to be good ones.