Recommender System


Amazon Music Unlimited



Overview



Amazon Music Unlimited is a service that charges a monthly fee of 10 dollars ( or 8 for Prime members ).
Its home page consists of a main banner entitled My Likes and More and underneath that, 4 ribbons consisting of 7 ablums, which the user can scroll left and right to see more.



Recommendations


The recommender system needs to decide on what banner to show, what themes for the ribbons, and then what albums specifically to offer inside the ribbons.


Note, the user can scroll down to see even more banners and ribbons, including “New Releases For You”, “Albums for You”, “Ultra HD Albums” and many others.



Scenario Design



Who are your target users ?
   Music lovers. Specifically those who would want to purchase a music service and never cancel.

What are their key gaols ?
   To explore. There is a certain buzz to exploring music. They also want to discover new music that they love.

How can you help them accomplish their goals?
   The web site should be fun. It should help the music lover find that buzz.



Reverse Engineering





I think Amazon has a nigthly offline-generated list of albums with a calculated, numerical factor for each category.

For example, “You might like” and “Albums for You”, based on similar customers (collaborative) , and similar music choices (search-based methods) would produce a list of 500 recommended albums, order by the most recommended to least recommended.

The degree of recommended is represented by a calculation using many factors including …

  1. User history of clicking on new recommendations
  2. User history of clicking on old recommendations
  3. User history of playing music in entirety
  4. User history of playing music partially
  5. Similar User history of clicking on recommendations
  6. Similar User history of playing music in entirety
  7. Similar user history of playing music partially
  8. User history of Variety across disparate genres
  9. User history of Variety across musical eras



The algorithm would return factors that yield a probability matrix to determine the likelhood of …

  1. Changing the recommended albums (how often)
  2. Changing the ribbon titles (how often)
  3. The genres of the recommendations
  4. The specific albums



Its important to point out. The recommendations dont need to change every day. The consumer may stare at a recommendation for a week and get convinced to click on it and enjoy it.

Lastly, the randomizer will select 8 albums from the large list, using a probability function that uses the factors to randomly select the recommendations.



Suggestions

I would suggest gap analysis, for example to gratify a consumer with a history of listenging to 70s, 80s and current, offer the 90s and 00s under the same genres.

I would also suggest focused surveys with target consumers to help validate your interpretation of the results.