Implicit Recommender Systems

Based on Alternating Least Square

Alternating Least Square is a method to find the matrices X,Y given R The idea is to find the parameters which minimizes the L^2 cost function,

while regularization factor controls the speed of converge

Step:

1.fix X, optimize Y

2.fix Y, optimizr X

3.repeat until converge or reach the iteration number

Some algorithms about ALS

Implicit Feedback:Link

http://activisiongamescience.github.io/2016/01/11/Implicit-Recommender-Systems-Biased-Matrix-Factorization/#Implicit-Feedback

The basic approach is to forget about modeling the implicit feedback directly. Rather, we want to understand whether user u has a preference or not for item i using a simple boolean variable which we denote by pui.pui. The number of clicks, listens, views, etc, will be interpreted as our confidence in our model.

While for the implicit feedback, the formula changes:

where C_ui is our confidence in P_ui. That is, the more a user has interacted with an item, the more we penalize our model for incorrectly predicting pui

LightFM use Stochastic Gradient Descent

DIfference between ALS and gradient descent

Need fewer iterations to reach the convergence,because every step is actually minimize the cost function