Content-Based Recommendation

KAIST GSDS 대학원 박찬영 교수님의 수업인 추천시스템 및 그래프 기계학습 수업 필기입니다.

 

Goal: Recommend items similar to those the user liked

 

When is it useful?

     -> Useful when ratings of other users are not available

 

How to defines similarity?

• Profile is vector of features and calculate computing similarity

• How to pick important features?
       -> TF-IDF (Term frequency X Inverse Doc Frequency)

 

Term frequency-Inverse Doc Frequency(TF-IDF)

• Bag-of-Words: Ecach document is represented by a binary vector
  idf: 특정 term이 등장한 문서의 수
  We can consider the number of accurrences of a term(i.e., Term frequency)
  Cons: Order of words/terms is ignored. Recently, word embedding methods are used e.g., word2vec

• Term-document count matrix
  tf: 특정 documents에서 특정 term 등장 횟수
  We should also consider the document frequency

 

 

 

Content-based approach: Pros/Cons

• Pros
  No cold-Start or parsity
  Able to recommend to users with unique tastes
  Able to recommend new and unpopular items
  Able to provide explanations
• Cons
  Requires content that can be encoded as meaningful features
  Difficult to implement seredipity
  Easy to overfit
  Effective for providing recommendations for new items, but not for new users
  
  

Pure content-based systems are rarely found in commercial environments