A topic that I didn’t discuss in my post about recommendation engines was the concept of exploitation versus exploration in terms of providing recommendations. Some unfortunate terminology here, but basically “exploitation” refers to using a person’s activity to recommend products that are similar in nature. Exploration refers to recommending products that is a bit out of the way. There is less correlation in terms of interests between the product that is recommended and what the person may have searched, viewed or bought in the past.
Exploration is very important to increase your information about a user’s profile. Without exploration, a recommendation engine will have a very limited view of the user. For example, if a person always searches for books and you only present them with similar books, the only information you gain about the user is based on the kind of books they view. You may surmise that they are interested in cooking or travel and maybe recommend something along those lines.
On the other hand, if the recommendation engine actively seeks to provide unrelated suggestions, it can make big leaps in understanding. Instead of drawing conclusions from what the user has done in the past, you can suggest entirely new domains and see if the user is interested. For example, the lover of cooking may also be a cycling enthusiast, but has been always searching on other sites for cycle products and accessories (and maybe relies on cycling blogs instead of books for information). You would never find out, unless you suggest and find out if the user is interested.
Take a look at the Recommendations List image below. This is from GoodReads on their “The Adventures of Huckleberry Finn” page. They use information from other readers who liked the same page. But as you can see, all the recommendations are stories that appeal to children or young adults, most of them also adventures. The recommendation engine could try to add 1-2 books that fall outside this audience and see what happens.
One problem with exploration is that your misses can be more numerous than your hits. And that can sometimes turn off users. They may be confused and may feel that the recommendations are worthless and stop paying attention to the recommendation widgets or emails. So the majority of recommendations must be based on similar items, so that the suggestions stay useful. This introduces the possibility that the user may not even see that exploratory suggestion.
A possibility is to tie the exploratory recommendation in the form of a reward. Perhaps a bargain sale, or some contest. For example, Amazon had this Gold Box, where there was a big deal for one item every day. You could visit it and then based on your actions (clicking, further viewing, etc.), Amazon had the potential to understand whether you liked items of that kind. I don’t think these recommendations were different for each individual, but even then, if a million people viewed the Gold Box item and some of them were found to be new interests in that kind of product, it was a net win for Amazon.
This is not only for product sales. If you are running a media site (say a newspaper), you would like readers to spend more time on your site and visit more content. So while the person visiting the op-ed pages may be more interested in business or politics, you could also try luring them to the arts and culture section, or the travel section, while still keeping the focus on their primary interests.