NN Models

Models are in the src/models/ folder of the repository. For each model, there is a struct that is a subtype of NCFModel. The constructor ensures the proper folder_name member parameter setting.

Generalized Matrix Factorization

NeuralCollaborativeFiltering.DotProductModel — Type

DotProductModel(df_train, df_test, embeddings_size, flux_model)

Recommender system based on Generalized Matrix Factorization: 2 Embedding layers with a dot product. Sizes of both embedding layers depend on a number of movies and on a number of users respectively.

Fields

df_train::DataFrame: Training set to be learnt on.
df_test::DataFrame: Testing set to be learnt on.
emb_size::Int64: Size of a single embedding vector for one user or for one movie.
model::Chain: Flux.jl Chain object representing the NN.
folder_name::String: Assigned to dot_product_ncf in the constructor automatically.

References

Source: Figure 1 (left) from Neural Collaborative Filtering vs. Matrix Factorization Revisited

source

Multi Layer Perceptron

NeuralCollaborativeFiltering.MLPSimilarityModel — Type

MLPSimilarityModel(df_train::DataFrame, df_test::DataFrame, emb_size::Int64, model::Chain)

Recommender system based on Multi Layer Perceptron: concatenate 2 Embedding layers and pass the result to the MLP. Sizes of both embedding layers depend on a number of movies and on a number of users respectively.

Fields

df_train::DataFrame: Training set to be learnt on.
df_test::DataFrame: Testing set to be learnt on.
emb_size::Int64: Size of a single embedding vector for one user or for one movie.
model::Chain: Flux.jl Chain object representing the NN.
folder_name::String: Assigned to mlp_similarity_ncf in the constructor automatically.

References

Source: Figure 1 (right) from Neural Collaborative Filtering vs. Matrix Factorization Revisited

source

Combination of GMF and MLP similarity

NeuralCollaborativeFiltering.GMFAndMLPModel — Type

GMFAndMLPModel(df_train::DataFrame, df_test::DataFrame, emb_size::Int64, model::Chain)

Recommender system based on a combination of Generalized Matrix Factorization and Multi Layer Perceptron similarity. Sizes of both embedding layers depend on a number of movies and on a number of users respectively.

Fields

df_train::DataFrame: Training set to be learnt on.
df_test::DataFrame: Testing set to be learnt on.
emb_size::Int64: Size of a single embedding vector for one user or for one movie.
model::Chain: Flux.jl Chain object representing the NN.
folder_name::String: Assigned to gmf_and_mlp_ncf in the constructor automatically.

References

Source: Figure 3 from Neural Collaborative Filtering 2017

source

build_model() function

"Wrapper" function build_model() uses Julia's multiple dispatch to create a neural network model based on the type of arguments:

NeuralCollaborativeFiltering.build_model — Function

build_model(x::Type{DotProductModel}, df_train::DataFrame, df_test::DataFrame; embeddings_size=50, share_embeddings=nothing)

Constructs and returns an instance of DotProductModel using training and testing data sizes along with specified model parameters.

Arguments

x::Type{DotProductModel}: The type of the model to be created. Multiple dispatch allows to define build_model for every model type differently.
df_train::DataFrame: The training data frame. Used to get the size of the embedding layers based on the numbers of movies and users.
df_test::DataFrame: The testing data frame. Same as df_train.
embeddings_size::Int=50 (optional): The size of the embedding vectors for both users and movies.
share_embeddings::Any=nothing (optional): Redundant here. Used in GMFAndMLPModel.

Returns

DotProductModel: An instance of DotProductModel with the initialized Flux Chain.

Example

m = build_model(DotProductModel, df_train, df_test, embeddings_size=60)

source

build_model(x::Type{MLPSimilarityModel}, df_train::DataFrame, df_test::DataFrame; embeddings_size=50)

Constructs and returns an instance of MLPSimilarityModel using training and testing data sizes along with specified model parameters.

Arguments

x::Type{MLPSimilarityModel}: The type of the model to be created. Multiple dispatch allows to define build_model for every model type differently.
df_train::DataFrame: The training data frame. Used to get the size of the embedding layers based on the numbers of movies and users.
df_test::DataFrame: The testing data frame. Same as df_train.
embeddings_size::Int=50 (optional): The size of the embedding vectors for both users and movies.
share_embeddings::Any=nothing (optional): Redundant here. Used in GMFAndMLPModel.

Returns

MLPSimilarityModel: An instance of MLPSimilarityModel with the initialized Flux Chain.

Example

m = build_model(MLPSimilarityModel, df_train, df_test, embeddings_size=60)

source

build_model(x::Type{GMFAndMLPModel}, df_train::DataFrame, df_test::DataFrame; embeddings_size=50, share_embeddings=false)

Constructs and returns an instance of GMFAndMLPModel using training and testing data sizes along with specified model parameters.

Arguments

x::Type{GMFAndMLPModel}: The type of the model to be created. Multiple dispatch allows to define build_model for every model type differently.
df_train::DataFrame: The training data frame. Used to get the size of the embedding layers based on the numbers of movies and users.
df_test::DataFrame: The testing data frame. Same as df_train.
embeddings_size::Int=50 (optional): The size of the embedding vectors for both users and movies.
share_embeddings::Bool=false: If set true, the embeddings will be shared between the GMF and MLP parts of the model. Otherwise, two separate sets of embeddings (4 Embedding layers in total).

Returns

GMFAndMLPModel: An instance of GMFAndMLPModel with the initialized Flux Chain.

Example

m = build_model(model_type, df_train, df_test, embeddings_size=60, share_embeddings=true)

source