dataSet {
trainSource 'train.yml'
testSource 'test.yml'
}Data for Train-Test Experiments
The easiest way to configure data for a train-test experiment is to use the crossfolder and register it as a dataSet, as in the first example in the train-test evaluator section.
However, sometimes you need to prepare one or more data sets externally and register them for with the train-test evaluator. An individual train-test pair can be registered by specifying their YAML files:
However, when generating many train-test splits this can be quite cumbersome.
Under the Hood
The LensKit crossfolder communicates with the train-test evaluator by generating a YAML file that pulls together all of the train-test data, along with any other data, into a single file describing a set of data sets. This file looks like:
name: DataSetName
datasets:
- train: training-data-1.yml
test: test-data-1.yml
- train: training-data-2.yml
test: test-data-2.ymlThe train and test keys can either be the names of files containing data manifests, or they can be the actual data manfest itself as a YAML object.
In addition to train and test, individual data sets can have an entity_types key identifying one or more entity types that will be used as the user’s test data. The default is rating; this key can either have a string naming a single type, or an array of type names.
If you have one of these files, you can add it directly as a data set in a train-test evaluation:
dataSet 'data-set.yml'Single Set Case
If the collection of data sets only has a single train-test pair, the datasets list is not required:
name: DataSetName
train: training-data.yml
test: test-data.ymlAdditional Settings
In addition, the train-test specification can have the following keys:
isolate-
A boolean; set to
trueto instruct LensKit to operate on each data set separately. This can reduce memory consumption for large data sets.