A recommendation system, is a system that can predict a rating for a user-item pair, where this user has never interacted with the item before.
In other words, is users of a user can give preferences for items/products in the sense of a star system (1-5 stars like NetFlix) or a thumbs-up/thumbs-down like YouTube, one can generate a rating matrix consisting of users and items. And each cell then has a rating value if the user has rated this particular item.
The table above, is called a ratings matrix, where a value above 0 is a rating defined by the user. A rating of 0 means the user has never encountered the product before.
We now wish to calculate all the cells with 0’s – This is where Matrix Factorization is useful.
Let’s do some maths’
We have a set of users and a set of items . Let of size be the ratings matrix that the users have assigned their items. We want to discover latent factors
We now need to find two matrices and such that their product approximates :
So each row of would represent the strength of the associations between a user and its factors. To get the prediction of a rating of an item by we can calculate the dot product of the two vectors corresponding to and :
We just need to find and – There exist several ways to do this.
I’m going to use Gradient Descent – initialize the two matrices with some values, and calculate how different their product is to and try to minimize the difference iteratively.
The squared error can be calculated by:
We want to minimize this error, and we want to know in which direction this error goes.
We now know in which direction to go, to minimize the error (i.e. the gradient) for both and
Where is a constant that determines the rate of approaching the minimum, and should be a small value like . If this value is too large we might step over the minimum, and maybe oscillating around the minimum.
Using the above update rules we can iteratively perform the operation until a convergence.
Now we have the factorization done, which is fine. – We can although run into a problem with over-fitting the model.
To avoid that, we introduce a regularization step:
Where is used to control the magnitudes of the user-factor and item-factor vector such that and would give a good approximation of without having to contain large numbers. In practice $latex $beta$ is set to some value in the range of
So the new update rules are:
Implementation in PHP
Running the above PHP implementation you will get the output:
Where all bold numbers was 0’s in the ratings matrix.
The only thing you need to do, is to use some kind of similarity function (Tanimoto, Pearson, K-NN) to find which items to recommend.
(I have to stop this guide now – it’s already too long) – I hope you gained some knowledge from this, or at least you can use my implementation in your system.
The PHP implementation is written by me, but taken from a Python implementation from here