Recommendation engines need context and personalization

A post from ReadWriteWeb about recommendation engines featuring Baynote, a software as a service company that helps sites target their recommendations, covered some interesting ground. Baynote argues recommendations based on current context produces superior results than those based on historical actions. The post cites Amazon as the leader for the opposing approach of recommendations based on historical views and transactions. Since I've had “bubble bath” recommended to me by Amazon, I understand historical data can lead to offers wide of the mark. (I’d made a 1-time purchase of Soap products as a gift for my mother).

The ideal employs both methods. An understanding of historical patterns and current context and intent will deliver more precise recommendations than either method. To return to my bubble bath example, for 330 days of the year showing me bubble bath is ridiculous. But, approximately one year after my original purchase and knowing it shipped to someone else, an offer of bubble bath related items might work.

Going forward a broader definition of user context than Baynote offers will improve recommendation engines. Context seems to be limited by onsite behavior. Other inputs like current real world activity, location, and friends interests all improve the results. Read Write Web astutely points out that room exists for multiple approaches in the marketplace.