推荐系统九大原则(英)

Datetime:2016-08-23 01:12:55          Topic: Recommender System           Share

Strands published a great vision paper about The Big promise of recommender systems . In this post I break down the paper into a few useful requirements; Recommender systems should:

  1. be easy to integrate and easy to remove . Vendor lock-in is a negative so having a minimal impact on the clients system is a big plus
  2. complete with internal marketing teams. Generated recommendations will compete for space and resources with internal marketing departments as their core goals are identical. So reuse the marketing metrics and methods in recommender systems to measure and drive real business value
  3. first collect enough data to avoid the cold start problem. Don’t release the recommender systems functions till there is enough data to provide good recommendations
  4. scale to the businesses needs, number of users and number of items
  5. be a hybrid of approach as these create a robust solution that solves many problems of exclusive algorithms
  6. be a balance between algorithms and UX. Focusing entirely on algorithms or UX will not create value, it must be a mixture of both
  7. use implicit instead of explicit feedback. Explicit ratings or reviews can be manipulated, however implicit (e.g. if they actually bought or watched the item) are more difficult to manipulate and will provide better data
  8. differentiate their products as recommender systems have become commoditized. If a system if difficult to evaluate and looks similar to other systems it will fail
  9. have contextual awareness of where the recommendations are being shown. For example, recommendations on mobile are different to web and different to email

Check out my Good Enough Recommender (GER) to see how it can be used to implement the requirements in this list and check out the List of Recommender Systems and see how they implement these requirements





About List