In most online content services today, recommendations accompany any purchase. A section titled You may also be interested in invites deeper browsing of the content library, and increases sales as a result. The technology behind this is known as a content recommendation engine or content discovery engine a software app that uses a […]
In most online content services today, recommendations accompany any purchase. A section titled You may also be interested in invites deeper browsing of the content library, and increases sales as a result. The technology behind this is known as a content recommendation engine or content discovery engine a software app that uses a set of algorithms, rules and business logic to filter the library and suggest movies, TV programmes, music, books and so on that are likely to be of interest to a specific user.
Although this presents a simple face to the user, it requires a lot of technologies and tuning under the lid, and there are significant differences between the way a good recommendation engine works in different circumstances. For instance, it may use different criteria to recommend video on demand (VOD) to a TV viewer as opposed to suggesting book purchases to someone browsing a bookstore.
Recommendation engines also come with varying degrees of sophistication. At a simple level, the recommendations may be generic suggestions offered to all users. A more nuanced approach would be to add a degree of broad-based personalisation, using groupings of users derived from demographic data. At the advanced level, precise personalisation makes use of explicit and implicit user preferences that may be combined with collaborative filtering and social media data.
Content-based recommendation uses analysis of the content items based on metadata such as author/director, actors, or genre. Collaborative filtering is based on preferences expressed by other users either from direct relationships between the target user and the other users, or in more advanced solutions, hidden correlations between sets of users and items.
Not all of these approaches are equally successful or appropriate for every type of content. Some algorithms are too computationally-intensive to be useful on a large scale in digital media apps like TV where there is a rapidly changing inventory and quick response is part of the user experience.
Classifying user profiles and content catalogues based on mood and personality have given return mixed results.
Some algorithms can also be very sensitive to small changes in tuning parameters, so these again do not suit a rapidly-changing content inventory where changes can skew the recommendation results produced.
Using contextual data from open web sources also generates noise in the system and has yet to prove a useful enhancement of recommendation quality.
These factors show that one recommendation engine is not necessarily like another, and not all are appropriate for digital media such as TV and movies. A good recommendation engine should be available in variants that have been tuned to the different requirements of individual applications such as online book retailing or VOD. It can take years of working with digital media customers to fine-tune the algorithms so that they perform well on specific quality problems such as weighting the relevance of live TV programming, VOD or eBook recommendations.
What are the benefits if a recommendation engine uses the appropriate algorithms tuned optimally for the application? For the content owner, there are some interesting effects on user behaviour, which have substantial implications for sales performance. Without a recommendation engine, users find things through browsing. This approach works well for those items the user already has in mind, and may be effective enough in selling a retailers best-selling content inventory. But when the inventory grows beyond a few hundred to a thousand items, the less well-known items become harder to sell. A well-tuned recommendation engine raises the chance of the discovery and consumption of lesser known content. Sales of mid-tail items typically items ranked 1500 to 3000 in the inventory rise sharply, while the tail of items beyond 3,000 typically contributes 15% of the sales. At Netflix, as reported by the New York Times, 70% of demand is from the backlist of older or small, independent movies.
Recommendation engine are also effective when used proactively to monitor and pilot recommendations as part of a promotional effort. Using business rules, the engine can be used to push content according to specific criteria: these can be a combination of demographic, price-based, behavioural, and time-based.
The ROI for operators of content recommendation systems has proven to be in the region of a 35% increase in sales, but good recommendations also have a beneficial effect on engagement (up between 15% and 25%) and customer churn (down by around 40% on average), indicating that the user experience is enhanced as well as the bottom line.
Sylvain Girard is general manager at ContentWise.