In vgherard/neuralcf: Neural Collaborative Filtering

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)

neuralcf

This package provides support for implementing recommender systems using the R Keras interface.

Installation

You can install the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("vgherard/neuralcf")

Example

library(neuralcf)

This example shows how to train and evaluate a neural network with the NeuMF architecture, for explicit feedback recommendations, using the MovieLense-100k dataset:

data <- ml_100k
head(data)
c(num_users, num_items) %<-% c(max(ml_100k[["user"]]), max(ml_100k[["item"]]))

We define our model using the sequential Keras API (all the steps explicitly illustrated here can be performed automatically through the utility neuralcf::neumf_recommender()). Our input is a pair of (user, item) integer ids:

input <- list(
    layer_input(1L, name = "user_input", dtype = "int32"),
    layer_input(1L, name = "item_input", dtype = "int32")
    )

A Generalized Matrix Factorization (GMF) layer embeds 'user' and 'item' in the same latent space and performs element-wise multiplication:

gmf_output <- layer_gmf(
        input,
        n = c(num_users, num_items),
        emb_dim = 64,
        emb_l2_penalty = 1e-4,
        name = "gmf"
    )

In parallel, we perform two independent embeddings of 'user' and 'item', and process the concatenated input through a Multi-Layer Perceptron (MLP) with four hidden layers:

mlp_input <- lapply(1:2, function(i) {
        layer_embedding(input[[i]],
                input_dim = c(num_users, num_items)[[i]] + 1,
                output_dim = 32,
                embeddings_initializer = "glorot_uniform",
                embeddings_regularizer = regularizer_l2(1e-4) 
                )
    }) %>% layer_concatenate %>% layer_flatten
mlp_output <- layer_mlp(
        mlp_input,
        units = c(64, 32, 16),
        activation = "relu",
        l2_penalty = c(1e-4, 3e-5, 1e-5),
        dropout_rate = c(0.3, 0.2, 0.1),
        name = "mlp"
    )

Finally, we concatenate the GMF and MLP outputs and project to a single output dimension (a positive number, rating of 'user' to 'item'):

gmf_mlp_output <- layer_concatenate(list(gmf_output, mlp_output))
output <- layer_dense(
        gmf_mlp_output,
        units = 1L,
        activation = "relu",
        kernel_regularizer = regularizer_l2(1e-4)
    )

We finalize our model using:

model <- keras_model(input, output)

We train the model on 90% of the MovieLense data using Adam:

compile(model,
    loss = "mean_squared_error",
    optimizer = optimizer_adam(),
    metrics = c("mean_squared_error", "mean_absolute_error")
    )

train_index <- 1:(.9 * nrow(data))
train <- data[train_index, ]
h <- fit(model,
    x = list(train[["user"]], train[["item"]]),
    y = ml_100k[["rating"]],
    batch_size = 256L,
    epochs = 20L,
    validation_split = .1,
    callbacks = callback_early_stopping(monitor = "val_mean_squared_error"),
    )

Here's a view on the loss and metric functions during training:

plot(h)

We finally evaluate on the remaining of the data:

test <- data[-train_index, ]
evaluate(model, x = list(test[["user"]], test[["item"]]), y = test[["rating"]])

vgherard/neuralcf documentation built on Dec. 23, 2021, 3:08 p.m.