Utilities

Data utilities

Recommenders.leave_one_out_split — Method

leave_one_out_split(table; col_user = :userid, col_time = :timestamp)

Leave-one-out split for the input table. For each user, whose column is specifed by col_user, the items are sorted by col_time, and the last one is split into the test set. The others remain in the train set.

Returns

train_table
test_table

source

Recommenders.ratio_split — Function

ratio_split(table, train_ratio = 0.7)

Split the table randomly, with the train set ratio specifed by train_ratio argument. Current implementaion assumes table object that can be converted to DataFrame.

Returns

train_table
test_table

source

Callbacks

For models trained by Stochastic Gradient Descent (SGD), one can give callback functions to the fit! method. Callback is any callable (functions or callable structs) that takes model, train_loss, epoch and verbose as inputs. If StopTrain exception is raised by the callback, training loop stops before the completion.

Currently only the following callbacks are implemented

Recommenders.EvaluateValidData — Type

EvaluateValidData(valid_metric::MeanMetric, valid_table, early_stopping_rounds, name = "val_metric")

Callback to monitor the validation metrics during training, and raise StopTrain exception if early stopping is requred.

Constructor arguments

valid_metric: monotring metric. See Evaluation metrics for the available ones.
valid_table: any Tables.jl-compatible object for validation dataset.
early_stopping_rounds: If the validation metric does not improve more than this epochs, the early stopping is invoked. If set to be less than 1, no early stopping is applied.
name: metrics name to show on logger.

Example

Use in the matrix factorizaion training.

ndcg10 = MeanNDCG(10)
cb = EvaluateValidData(ndcg10, test_table, 1, "val_NDCG")

model = ImplicitMF(16, true, 0.01)
fit!(
    model,
    train_table,
    10,
    callbacks = [cb],
    col_item = :movieid,
    n_epochs = 20,
    n_negatives = 1,
    learning_rate = 0.01,
    verbose = 1,
)

source