Deep learning for latent factor collaborative filtering

Instead of using SVD, you can leverage deep learning methods to derive the user and item profile vectors of given dimensions.

For each user i, you can define a user vector u_i ∈ R^k through an embedding layer. Similarly, for each item j, you can define a item vector v_j ∈ R^k through another embedding layer. Then, the rating r_ij of a user i to an item j can be represented as the dot product of u_i and v_j as shown:

You can modify the neural network to add biases for users and items. Given that we want k latent components, the dimensions of the embedding matrix U for m users would be m x k. Similarly, the dimensions of the embedding matrix V for n items would be n x k.

In the The deep learning-based latent factor model section, we will use this embedding approach to create a recommender system based on the 100K Movie...