R: janus

janus {janus}

R Documentation

janus

Description

Coarse-to-fine optimization of a recommending system based on deep neural networks with Tensorflow/Keras back-end

Usage

janus(
  data,
  rating_label,
  rater_label,
  rated_label,
  task,
  skip_shortcut = FALSE,
  rater_embedding_size = c(8, 32),
  rated_embedding_size = c(8, 32),
  layers = c(1, 5),
  activations = c("elu", "selu", "relu", "sigmoid", "softmax", "softplus", "softsign",
    "tanh", "linear", "leaky_relu", "parametric_relu", "thresholded_relu", "swish",
    "gelu", "mish", "bent"),
  nodes = c(8, 512),
  regularization_L1 = c(0, 100),
  regularization_L2 = c(0, 100),
  dropout = c(0, 1),
  batch_size = 64,
  epochs = 10,
  optimizer = c("adam", "sgd", "adamax", "adadelta", "adagrad", "nadam", "rmsprop"),
  opt_metric = "bac",
  folds = 3,
  reps = 1,
  holdout = 0.1,
  n_steps = 3,
  n_samp = 10,
  offset = 0,
  n_top = 3,
  seed = 999,
  verbose = TRUE
)

Arguments

`data`	A data frame including at least three features: rating actor, rated item and rating value.
`rating_label`	String. Single label for the feature containing the rating values.
`rater_label`	String. Single label for the feature containing the rating actors.
`rated_label`	String. Single label for the feature containing the rated items.
`task`	String. Available options are: "regr", for regression (when the rating value is numeric); "classif", for classification (when the rating value is a class or a factor).
`skip_shortcut`	Logical. Option to add a skip shortcut to improve network performance in case of many layers. Default: FALSE.
`rater_embedding_size`	Integer. Output dimension for embedding the rating actors. Default: coarse-to-fine search (8 to 32).
`rated_embedding_size`	Integer. Output dimension for embedding the rated items. Default: coarse-to-fine search (8 to 32).
`layers`	Positive integer. Number of layers for DNN. Default: coarse-to-fine search (1 to 5).
`activations`	String. String vector with the activation functions for each layer. Default: coarse-to-fine search ("elu", "selu", "relu", "sigmoid", "softmax", "softplus", "softsign", "tanh", "linear", "leaky_relu", "parametric_relu", "thresholded_relu", "swish", "gelu", "mish", "bent").
`nodes`	Positive integer. Integer vector with nodes for each layer. Default: coarse-to-fine search (8 to 512).
`regularization_L1`	Positive numeric. Value for L1 regularization of loss function. Default: coarse-to-fine search (0 to 100).
`regularization_L2`	Positive numeric. Value for L2 regularization of loss function. Default: coarse-to-fine search (0 to 100).
`dropout`	Positive numeric. Value for dropout parameter at each layer (bounded between 0 and 1). Default: coarse-to-fine search (0 to 1).
`batch_size`	Positive integer. Maximum batch size for training. Default: 64.
`epochs`	Positive integer. Maximum number of forward and backward propagation. Default: 10.
`optimizer`	String. Standard Tensorflow/Keras Optimization methods are available. Default: coarse-to-fine search ("adam", "sgd", "adamax", "adadelta", "adagrad", "nadam", "rmsprop").
`opt_metric`	String. Error metric to track for the coarse-to-fine optimization. Different options: for regression, "rmse", "mae", "mdae", "mape", "smape", "rae", "rrse"; for classification, "bac", "avs", "avp", "avf", "kend", "ndcg".
`folds`	Positive integer. Number of folds for repeated cross-validation. Default: 3.
`reps`	Positive integer. Number of repetitions for repeated cross-validation. Default: 1.
`holdout`	Positive numeric. Percentage of cases for holdout validation. Default: 0.1.
`n_steps`	Positive integer. Number of phases for the coarse-to-fine optimization process (minimum 2). Default: 3.
`n_samp`	Positive integer. Number of sampled models per coarse-to-fine phase. Default: 10.
`offset`	Positive numeric. Percentage of expansion of numeric boundaries during the coarse-to-fine optimization. Default: 0.
`n_top`	Positive integer. Number of candidates selected during the coarse-to-fine phase. Default: 3.
`seed`	Positive integer. Seed value to control random processes. Default: 42.
`verbose`	Printing specific messages. Default: TRUE.

Value

This function returns a list including:

pipeline:
model:
- configuration: DNN hyper-parameters (layers, activations, regularization_L1, regularization_L2, nodes, dropout)
- model: Keras standard model description
- recommend: function to use to recommend on rating actors
- plot: Keras standard history plot
- training_metrics: tracking of opt_metric across folds and repetitions
- test_frame: testing set with the related predictions, including
- testing_metrics: summary statistics for testing
time_log

Author(s)

Maintainer: Giancarlo Vercellino giancarlo.vercellino@gmail.com [copyright holder]