Skip to main content

Recommendation Systems

Aims to find the and suggest items of likely interest based on the users' preferences

Association based

Content Based

Collaborative Filtering

  • Memory/neighbor based: user based, item based
  • Model based: matrix factorization

Community Based

Knowledge Based

Terms

  • Long tail: selling niche products to many customers
  • Cold start: a new user that doesn't have any available data
  • Serendipity: recommend an item that a customer likes, even though they didn't ask for it
  • Implicit: inferred data from user behavior but we will need to derive the rating. Often use [[Jaccard Similarity]] to compute this
  • Explicit: directly given (related to revealed preference survey), but can be subject to sparsity, erroneous, unreliable
  • Sparsity: recommending items in the long tail is hard because of few user rating

Measures:

  • Mean absolute error
  • Mean average precision at K: calculate the mean average of number of relevant items within the top k recommendations
  • Lift: the number of hits of the model at top k recommendations over the number hits at random