Book Recommender

Types of recommendation systems:

• Collaborative filtering model:

  • Based on feedbacks:
    • explicit feedback(ratings, thumbup/thumdown)
    • implicit feedback(book in shopping cart, numbner of episode watched in a TV show)
  • User-user similarities (people like you, also like "The Hobbit")
  • Item-item similarities (if you like "The Hobbit", you will also "The Lord of Rings")
  • User-item similarities : combined both approaches above and use matrix facotrization -> Best trained with Singular Value decomposition (SVD)

• Content-based model:

  • Similarities are calculated based on content based features (like the genre, duration of the movie)

• Hybrid recommender model:

  • Combined collaborative filtering with content-based
  • Because Hybrid Deep Learning algorithms are non-linear, they are less prone to over-simplify a user’s tastes. They allow us to learn much finer interactions between users and items.
  • Deep learning models can reveal more complex tastes over cross-domain data (predict a user's taste on music, movies and even dishes!?)

In this project:

Clean data:

• book data
• user data
• rating data

Data analysis - extract insights

• explict rating (non-0s) vs. implicit rating (0s)
• calculate sparsity of the matrix

Create user-item matrix

• fillna
• transpose the matrix

Build recommender

• Collaborative filter: KNN approach
  • Item-based recommender (CF)- find similiar books and recommend to users
  • Use KNN to find k most similar books
  • Default params:
    • metric = 'cosine'
    • algorithm = 'brute'
    • k = 5
• Content based model using Natural Language Process (NLP)
  • content here is the book title
  • NLP steps:
    • remove NonAscii, make words lower cases, remove stop words, punctuations and html links in the title
    • convert book title in to vectors by TfidfVectorizer (bigram)
    • use consine similarity for similarity measurement
    • score the book titles based on similarities and get top k most similar book titles by the given book title.