15_Recommender systems
Carpe Tu Black Whistle

Recommender systems

what recommender system do is looking at users and predict how they would have rated other movies that they have not yet rated.

Motivation(Reasons

  • important application of machine learning.
  • it’s a typical big think of machine learning

Problem Formlaiton





Other notation

Content Based

image

For each user, learn a parameter. Predict useras rating moviewithstars.

To learn:

For convenience, we are going to get rid of the termwithout changing the value of thetathat it can be taken out of the optimization.

Optimizaition objective:
To learn(parameter for user):

To learn:

Gradient descent update:

Collaborative Filtering

feature learning

Given, to learn:


Given,to learn:

Algorithm

Minimizingandsimultaneously:

  1. Initializeto small random values.
  2. Minimizeusing gradient descent( or an advanced optimization algorithm).E.g for every:


  1. For a user with parametersand a movie with(learned) features, predict a star rating of.

Vectorization: Low rank matrix factorization

image

Mean normalization

If there is a user who haven’t rate any item before.
The Collaborative Filtering Algorithm will predict all the user’s rated score ZERO. So the idea of mean normalization will fix the problem.

image

For each product, we learn a feature vector.

Some features are over human being’s comprehensive ability

How to find moviesrelated to movie

5 most similar movies to movie
Find the 5 movieswith the smallest

后记

协同滤波算法的一大特点就是,学习一个多用户多目标的大系统。 不适合单个用户的喜好推荐(大材小用)
协同滤波是一种学到的第一个半监督算法