What makes a good recommendation?

At MediaUnbound, we’ve been analyzing what makes a “good” and “successful” recommendation since before the first dot-com meltdown. Since then, one thing we've definitely learned is that there is no single magical algorithm, a secret combination of content features, or a simple recipe that will provide the right answer every time. What makes a “good” recommendation is a complex question that changes based on context, audience, and customer requirements.

Our technology follows some basic guidelines in our quest to provide successful recommendations for each customer's unique products:

• Focus on the User: Understanding and learning about the user is central to MediaUnbound’s recommendation philosophy. Instead of attempting to solely model the media itself, we seek to replicate the complex cultural and aesthetic qualities of taste and human preferences. Our approach has resulted in formally modelling not users or media, but the complexity of the experience. Most other personalization or recommendation systems typically have their basis in categorization — extracting and describing the features of songs, video, pictures, and so on — in order to help users find similar media. We however have found from the beginning that recommendations based on categorizing media don’t reflect the complicated ways real people choose and react to what they see, read, and hear. Focusing on the user produces a more relevant set of data and analytic techniques for recommending media.
• Use human media analysts to complement computer technology: Computers are very good at cataloging and arranging large sets of data, but despite these strengths can not on an experiential level understand why people like what they do. By blending the input of our expert human analysts with our advanced computer technology, we can create a far more enjoyable and exciting experience for users, rather than just telling people what they already know.
• Seek out inefficiencies in user’s and world’s knowledge: When creating recommendations, MediaUnbound strives to understand the boundaries of the user’s knowledge so that we can provide a mixture of known and new material. By using human analysts, we can create a system that better understands what new material to present to users. This can include both selections that there is little data about in the world, or which are relatively unknown to the general population.
• Build trust: Taste and preference are based much more on emotional affinities than formal connections. We have found that the source of a recommendation is sometimes as important as the the recommendation itself. Our products demonstrate to users specific knowledge of both the content that is being viewed and the individual’s unique preferences. Trust is built by intelligently exposing users to both familiar and new material, and by avoiding bad selections of material that is merely globally popular. Sometimes it is more important to know what not to recommend.
• Silver-bullet solutions do not exist: There is no single algorithm or monolithic piece of software that can make the best recommendation every time. MediaUnbound uses multiple techniques and is not afraid to try new approaches for new problems.