Explain-Da-V is a framework aiming to explain changes between two given dataset versions. Explain-Da-V generates explanations that use data transformations to explain changes.

Explaining Dataset Changes for Semantic Data Versioning with Explain-Da-V Roee Shraga, Ren'ee J. Miller, PVLDB (to appear), 2023

BibTeX: TBD

1. Anaconda 3
2. Tabulate
3. Featuretools

0. Download and extract the Semantic Data Versioning Benchmark
   0.1 For BYOD (bring your own data), please follow the format of SDVB
1. Clone Explain-Da-V repository
   1.1. Add three empty directories named results (stores functional dependencies), temp, and output (will hold the results of Explain-Da-V)
   1.2 Download Metanome runnable.
   1.3 Rename metanome-cli-1.1.0.jar as metanome.jar and add to the Explain-Da-V repository
2. (optional) locate the dataset folder in the repository

1. Configuring Explain-Da-V is done via the config file
   1.1. (required) update the following entries to be consistent with local machine:
   - dataset_name: the name of the dataset (also the name of the folder, e.g., IMDB)
   - problem_sets_file: the location of the problem_sets file (e.g., 'Data/Benchmark/{}/problem_sets.csv'.format(dataset_name))
   1.2 (optional) update other parameters, e.g., CATEGORICAL_UPPER_BOUND(the number of unique values to be considered as a categorical type).
2. Run main
   2.1. Use main_with_problem_sets for default setting
   2.2. Other settings are used for ablation study (e.g., use_fd_discovery=False) and baselines (e.g., main_with_problem_sets_baseline_original(extend_for_auto_pipeline=False, extend_for_plus=True)
3. The output will be generated in the directory output
   3.1 The output file documents the problem_set, nature of change (e.g., adding columns), the resolved trasformation and its evaluation (please see paper for more details).

The code framework uses two existing systems, namely metanome and Foofah (both are included in the repository) :

• We find functional dependencies using Metanome. See Functional Dependency Discovery: An Experimental Evaluation of Seven Algorithms and Repeatability - FDs and ODs for additional details.
• We adopt and extend Foofah to resolve textual transformations. See Foofah: Transforming Data By Example for additional details.

Explain-Da-V was developed at the Data Lab, Northeastern University by Dr. Roee Shraga and Prof. Ren'ee J. Miller.

The repository also contains Ablation Study Plots (Figure 11, Section 7.3):

• Validity
• Generalizability
• Conciseness
• Concentration