Recommender System: Recommendation Algorithms

5th July 2017 at 10:07am
Recommender System
  • Non-Personalized Summary Statistics
    • Mean-based:
      • Background: In a 5-star scale rating system
      • Symbols:
        • UiU_i for users who rated item ii
        • ruir_{ui} for rating by user uu to item ii
        • Ui|U_i| for number of user who rated item ii
        • α\alpha for damping factor, larger would cause the rating floats more smooth
        • μ\mu for global rating across all items and users
        • s(i)s(i) for score of item ii
      • Mean: s(i)=uUinruiUis(i) = \dfrac{\sum_{u \in U_i}^n r_{ui}}{|U_i|}
      • Damped Mean: s(i)=uUinrui+αμUi+αs(i) = \dfrac{\sum_{u \in U_i}^n r_{ui} + \alpha\mu}{|U_i| + \alpha}
    • Association-based:
      • Background: typically in a shop that customs buying products
      • Symbols:
        • P(ij)P(i|j): probability of buying i when already buying j
      • Basic Association, measuring probability by counting:
        • Story: How many percentage of people who buying i also buying j among the whole people who buying i?
        • Formula: P(ij)=P(ij)P(j)=UiUj/UUj/UP(i|j) = \dfrac{P(i \land j)}{P(j)} = \dfrac{|U_i \cap U_j| / |U|}{|U_j| / |U|}
        • Bad case: if j is popular, then the result is bad
      • Bayes's Law: P(ij)=P(ji)P(i)P(j)P(i|j) = \dfrac{P(j|i) P(i)}{P(j)}
      • Lift Association, measuring score by counting:
        • Story: people who bought i and j together more often means i and j are more associative
        • Formula: s(ij)=P(ji)P(i)P(j)s(i|j) = \dfrac{P(j \land i)}{P(i) P(j)}
  • Content-Based Filtering
    • Information Filtering
    • Knowledge-Based
  • Collaborative Filtering
    • User-User
    • Item-Item
    • Dimensionality Reduction
  • Others
    • Critique / Interview Based Recommendations
    • Hybrid Techniques