Recommendation system approaches

Implementation, test and comparatives of different recommendation models (Collaborative filtering like).

Models

• Embedding dot model

  • Learn both users and movie embedding.
  • Dot product between these to predict user ratings.

• Embedding biases dot model

  • Add a bias to each user and movie. Similar to the bias in a fully-connected layer or the intercept in a linear model. It just provides an extra degree of freedom.
  • Also pass dot product output through a sigmoid layer and then scaling the result using the min and max ratings in the data. This technique introduces a non-linearity into the output and results in a loss improvement.

• Embedding dense model

  • Instead of performing a dot product between users and movies, embedding adds a fully connected layer(dense) after input embedding layers.
  • This technique is more flexible than the previous ones because it allows adding more feature columns as model input.

• User/Movie/Gender embedding dense model

  • Same Embedding dense model approach.
  • Add genders features columns.
  • Use a sigmoid layer and then scale the result using the min and max ratings in the data.

• Wide and deep model

  • A mix between two models, linear regression and Embedding dense model.
  • This model learns to combine memorization and generalization like humans do.
  • Liner model learning to memorize.
  • Deep model learning to generalize.
  • See: Wide & Deep Learning: Better Together with TensorFlow

• Deep Factorization Machine Model

  • Pending. See: Deep Factorization Machine Model for CRT prediction.

Notebooks

• Recommendation system: Approaches
• Recommendation systems: Deep Model only
• User/Movie/Genders Deep Model
  • 1. Dataset preprocecing
  • 2. Dataset preprocecing and train/validation/test split
  • 3.1. Model training/validation (Pandas Ray version)
  • 3.2. Model training/validation (Spark version)

Requisites

• anaconda / miniconda

Getting started

Step 1: Clone repo.

$ git clone https://github.com/adrianmarino/recommendation-system-approaches.git
$ cd recommendation-system-approaches

Step 2: Create environment.

$ conda env create -f environment.yml

Step 3: Enable project environment.

$ conda activate recommendations

Open notebooks locally

Step 1: Enable project environment.

$ conda activate recommendations

Step 2: Under project directory boot jupyter lab.

$ jupyter lab

Jupyter Notebook 6.1.4 is running at:
http://localhost:8888/?token=45efe99607fa6......

Step 3: Go to http://localhost:8888.... as indicated in the shell output.

Note: Use Pycharm community to edit source code for more comfort.