Recommendation Algorithms

• Two Approaches
  • Collaborative Filtering: Predict based on past customer-item relationship data
  • Content-Based: Predict based on information of each customer and item itself
• Assumption: Similar customers tend to like similar things overall
  • Based on this assumption, there is a method to do it with collaborative filtering
  • GroupLens: Early prediction algorithm
    • Measure similarity between customers using correlation coefficient
  • Matrix factorization of item-customer table
    • $y_{i,j}=U_i^T{v}_j$
      • Fill in the unobserved parts based on the constraint that the table can be represented by the product of two low-dimensional matrices
    • Further extended to n-dimensional tensor factorization
    • Applicable to various situations such as item-customer-time or person-person-time, not just the two-dimensional relationship between items and customers
    • Calculation involves applying machine learning methods such as gradient descent (Information Science Expert)

#recommendation