Expose Dataset API

[folmos-at-orange]

Description

Currently the Dataset class is an internal utility for the sklearn module. The idea is to render this class public so it is an utility to create multi-table datasets.

Questions/Ideas

This feature would ease many tasks:

• Splitting a dataset in train/test
• Sorting a whole dataset
• Building core api parameters (notably additional_data_tables)
• Simplify tutorials and samples by having a method get_dataset_sample("Accidents", type="pandas")

Main design element: A builder pattern.

• Add mutator methods to construct the dataset from an empty one (Dataset())
• (Partially) Implemented in prototype:
  • Classes
    • PandasDataset
    • FileDataset
  • add_table(self, name, source, key=None)
    • key mandatory for multi-table
    • source will be different in each Dataset subclass
  • train_test_split (implemented in PandasDataset only)
  • sort sorts the dataset by their keys (implemented in FileDataset only)
  • create_khiops_dictionary_domain
  • create_additional_data_table_param
  • add_relation(self, parent_table_name, child_table_name, one_to_one=False)
• Not in prototype:
  • remove_table(self, name)
    • Removes all relations asociated the the table
  • remove_relation(self, parent_table_name, child_table_name)
  • check(self):
    • Raises warnings and exceptions
      • errors:
        • Non-existent table names
        • No main table set in multi-table datasets
        • No key set in multi-table datasets
      • warnings:
        • Dangling tables
  • add_external_relation(self, parent_table_name, key, another_dataset)

Design questions:

• Immediate consistency checks:
  • That is , should check be called at each mutator call ?
    • I'm inclined to this one since the target audience are not only dev's
  • or the user should check the consistency before using it ?
• Should we accept mono-table datasets ?
  • This adds many edge-cases
• What about helper functions using the FileDataset:
  • train_predictor_ds(ds, target_variable_name, output_dir, <kwargs without additional_data_tables, header_line, field_separator>)
  • deploy_model_ds(model_kdic, ds, output_dir, <kwargs - additional_data_tables, header_line, field_separator> )

[folmos-at-orange]

Waiting for input after the first version of the spec: https://github.com/KhiopsML/khiops-python/wiki/Dataset-Spec-Proposal