YADL: Yet Another Data Lake

YADL (Yet Another Data Lake) is the benchmark data lake we developed for the paper "Retrieve, Merge, Predict: Augmenting Tables with Data Lakes".

We use this repository to "unpack" the YAGO 3 tables in order to go from few, very large tables to a collection of smaller tables that contain some of the same information, in different formats.

Preparing the base tables

Depending on the method used (refer to the paper for more details), tables are prepared in one of the scripts

• generate_yadl_long.py
• generate_yadl_long_variants.py
• main_yadl_construction.py

Implementing the YADL variants

We provide the YADL variants we used in the paper on Zenodo.

If YAGO 3 files are available, additional variants may be constructed by using the scripts mentioned above. Alternatively, it is possible to build them starting from the YAGO 3 files, provided they are available, by using the script main_yadl_construction.py for YADL Base and YADL Binary, and the scripts generate_yadl_long.py and generate_yadl_long_variants.py for YADL 10k and 50k.

Building YADL Base and YADL Binary

usage: main_yadl_construction.py [-h] [-d DATA_DIR] [--top_k TOP_K]
                                 [--min_count MIN_COUNT] [--explode_tables]
                                 [--comb_size COMB_SIZE [COMB_SIZE ...]]
                                 [--min_occurrences MIN_OCCURRENCES] [--debug]
                                 {wordnet,binary}

positional arguments:
  {wordnet,binary}      Strategy to use.

options:
  -h, --help            show this help message and exit
  -d DATA_DIR, --data_dir DATA_DIR
                        Root path to save new tables in.
  --top_k TOP_K         The top k types by number of occurrences will be selected.
  --min_count MIN_COUNT
  --explode_tables      After generating the tables, generate new synthetic subtables.
  --comb_size COMB_SIZE [COMB_SIZE ...]
                        Size of the column combinations to be generated in the explode
                        stage. Defaults to 2.
  --min_occurrences MIN_OCCURRENCES
                        Minimum number of non-null values to select a pair. Defaults to
                        100.
  --debug               If set, downsample the yago tables to 100_000 values to reduce
                        runtime.

To construct the YADL Base variant, use the following command and parameters:

python main_yadl_construction.py -d destination/path/to/use --explode_tables --comb_size 2 3 --top_k 200 wordnet

To construct the YADL Binary variant, use the following command and parameters:

python main_yadl_construction.py -d destination/path/to/use binary

Building YADL 10k and 50k variants

This operation is done in two steps, first by creating the long form tables using generate_yadl_long.py, then by building the subtables using generate_yadl_variants.py.

usage: generate_yadl_long.py [-h] [--dest_path DEST_PATH]
                             [--max_fields MAX_FIELDS]

options:
  -h, --help            show this help message and exit
  --dest_path DEST_PATH
                        Path where the tables will be
                        saved.
  --max_fields MAX_FIELDS
                        Number of argument values to keep
                        when building tables.

dest_path is the path that will contain the "seed tables": this is the same path that should be provided to the next script to generate the subtables. max_fields is used to limit the length of the seed tables. Whenever a subject has multiple relations (e.g., a director with many films), the number of rows in the resulting table will be multiplied by the number of relations; to avoid having scalability issues and keeping the size of the seed tables "contained", max_fields is set to 2. It can be increased to build much longer tables.

usage: generate_yadl_long_variants.py [-h] [--col_resample COL_RESAMPLE] [--row_resample ROW_RESAMPLE]
                                      [--sample_fraction SAMPLE_FRACTION] [--minimum_arity MINIMUM_ARITY]
                                      [--minimum_rows MINIMUM_ROWS]
                                      base_path

positional arguments:
  base_path             Path with base tables to replicate.

options:
  -h, --help            show this help message and exit
  --col_resample COL_RESAMPLE
                        Number of subtables to generate.
  --row_resample ROW_RESAMPLE
                        Number of resamplings for each subtable. Defaults to 2.
  --sample_fraction SAMPLE_FRACTION
                        Fraction of lines to keep for row resampling. Defaults to 0.7.
  --minimum_arity MINIMUM_ARITY
                        Minimum number of columns.
  --minimum_rows MINIMUM_ROWS
                        Minimum number of rows of columns.

col_resample is the number of random projections that should be generated for each seed table. row_resample will generate row_resample new additional tables for each random projection by executing a random sampling of the rows of size sample_fraction (so, if row_resample==2, two additional subsamplings will be added on top of the single projection). minimum_arity is the minimum width of columns that a seed table should have to be considered for generating subtables (other seed tables will be ignored). minimum_rows removes all subtables with fewer rows than this number.

Estimating the size of the resulting data lake

For the long variants (YADL 10k and 50k), the generation procedure may be time-consuming and lead to a large disk footprint. We provide a notebook (notebooks/Estimate number of tables.ipynb) to help the user with estimating the resulting size of the data lake based on the parameters that are provided.

Due to the fact that there is a high degree of randomness involved in generating the tables (some tables may be filtered out, the fraction of null values is not constant etc.), the values provided are not exact; we observe that they are however within 10% of the actual values resulting from preparing the data lake.