(See pfd versions for more information.)

Using autoencoders to build recommender systems

Task description

We need to build recommender systems that:
1.	Can be speed effective in the model using phase. 
2.	Will be speed effective while processing new data. It will not need to rebuild the model for every new rating or new user or product.   

Decision

Used data

We use the next data structures.
1.	The user data (i.m. for example sex, age for the MovieLens 1M dataset).
2.	The product data (i.m. for example year of a film production for the MovieLens 1M dataset).
3.	The rating data (i.m. for example rate day of a film for the MovieLens 1M dataset).
4.	The rate (or rates) itself.
5.	The user id vector.
6.	The product id vector.

The phase of model training

First we need to train autoencoders. There are to autoencoders in system one for the products space and one for users space. 

Let the first autoencoder takes M (5..9) records of rating some product by some users.  One record for the first autoencoder contains: 1) the user id vector; 2) the user data; 3) the rate (or rates) of a user for the product and 4) the rating data (by users for some product). All that records (for one train step) are for one product. And first autoencoder encodes the products id vectors. 

Let the second autoencoder takes M (5..9) records of rating products by some user. One record contains: 1) the product id vector: 2) the product data; 3) the rate (or rates) of the user for a product and 4) the rating data (by the user for a product). All that records (for one train step) are for one user. This autoencoder encodes the users id vectors.

In 100..1000 cycles after we have trained autoencoders, we start getting products and users id vectors.

For one product we get (24..64) encoded values by running the first autoencoder. (We set autoencoder input as we did while training.) Then we get average of those encoded values. And so we get new product id vector to move to. We repeat that process for some products. So we get new products id vectors to move the old products id vectors to. Same we do for some users. And will get new users id vectors to move the old users id vectors to.

But I should mention that we need to take into account the “breathing” of encoded values of autoencoders. I use getting the average offset of id vectors to correct such “breathing”. See code for details.

Then we repeat training of autoencoders for corrected id vectors. And then we repeat getting of new id values once again. And so on. 

So we get products and users id vectors. That id vectors we use in prediction of rating value(s). We train neural net that takes 1) a user id vector; 2) a user data; 3) a product id vector; 4) a product data; 5) a rating data to predict rating(s) of a product for a user.

After building the model we can use it.
1.	We predict ratings of a product for a user.
2.	We correct id vectors after getting new ratings from users. We do it same as we did while getting id vectors in the training phase.
3.	We add new id vectors of new products and users. In same way. 

When we use model we do not need to train neural nets, we just compute the id vectors. That gives us opportunity to optimize calculations.