Skip to content

Latest commit

 

History

History
227 lines (200 loc) · 8.93 KB

README.md

File metadata and controls

227 lines (200 loc) · 8.93 KB
Raw

Dataset Similarity

What is Datasets Similarity?

The Dataset Similarity project deals with the issue of comparing tabular datasets. The idea of the project is that we will have a set of datasets that we want to compare with each other and find out their similarity or distance. This project mainly focuses on comparing only two tables. The final similarity is calculated according to the similarity of individual columns based on their metadata. Columns are compared by type and by content.

For testing, we have prepared two sets of data, the main set (training) on which the program is tuned, and a validation set for validating the results.

Definition of table similarity:

img_1.png

Parameter important columns is user input.

Parameter k is also user input.

Approach

You can see two options for implementation in the pictures below. This implementation is only for comparing two tables. In both implementations, we first create metadata for each table. MetadataCreator creates the metadata and implementation of the creator is modular. After metadata are created for both tables, they are used as input for Comparator. Comparator compares metadata and it computes distance. We should test which one is better.

  1. img_2.png
  2. img_3.png

Metadata creator

MetadataCreator has:

  • constructor that fills fields:
    • size
    • column_names
    • column_names_clean(lowercase, only numbers and letters)
    • column_incomplete(if a column has more than 30 % missing values, it is marked as incomplete)
  • Methods for set creator
    • set_model: sets word transformer model
    • compute_column_names_embeddings: computes embeddings for clean column names
    • compute_column_kind: computes kind
    • compute_basic_types: compute types on top level
    • compute_advanced_types: compute types on middle level
    • compute_advanced_structural_types: compute types on smaller level (user should pick only one of these 3)
    • compute_correlation: correlation between columns
    • create_column_embeddings: create embeddings for columns
  • Getters
    • get_column_by_type: it returns all names of columns with a specified type
    • get_numerical_columns: it returns names for all numerical columns
    • get_metadata: it is the main method. It returns created metadata.

Usage: firstly we call constructor, then we can call any set methods (but for types we should pick just one), then we can get metadata

Comparator picture 1

This comparator creates several matrixes, each matrix represents a comparison for two columns of the same type. Matrix's could represent different aspects.

For example, for type int we will create:

  • a matrix comparing column names
  • a matrix comparing max values
  • a matrix comparing range
  • ...

For type string we will create:

  • a matrix comparing column names
  • a matrix comparing embeddings
  • a matrix comparing the most used word

Then we will create one matrix for string and one matrix for int by using built-in function to unite matrix's.

From each of these two matrixes we will compute a distance number. Then these distances will be merged.

Comparator picture 2

This comparator will create one big matrix for all columns regardless of the type. Each element in the matrix will be computed from several aspects (for int: column names, max value, range ...). Then we create one number from this huge matrix, which is the distance of these two tables.

Column2Vec

Column2Vec is a module in which we implement word2Vec based functionality for columns. It will compute embeddings for columns, so we can compare them. More about this module can be found here.

Types and Kinds

We have decided to split columns by type. We can compute types or kinds for each column. Types define the real type of column. Some you may know from programming languages (int, float, string) and some are specific (human generated, word, sentence ...). Kinds represent higher categorization.

Types have some hierarchy as you can see on picture 3. In the previous lines we named it: top level, middle level, smaller level. Explaining some types:

  • Human generated: with more than three numbers after decimal point. All others are computer generated.
  • word: string without a space
  • sentence: string starting with an upper case letter and ending with fullstops(or ! ?). It contains only one fullstops.
  • phrase: string with more than one word
  • multiple: string that represents not atomic data or structured data
  • article: string with more than one sentence
  1. img.png Kind has only for "types" plus undefined. You can see all types on the picture 4. Explaining kinds:
    • As Id will be marked column that contains only uniq values
    • As Bool will be marked column that contains only two unique values
    • As Constant will be marked column that contains only one unique value
    • As Categorical will be marked column that contains categories. Number of uniq values is less than threshold % of the total number of rows. Threshold is different for small and big dataset.
  2. img.png

Applicability

  • merging teams
  • fuze of companies
  • found out which data are duplicated
  • finding similar or different data

Structure

  • Source code is in folder similarity. More about similarity folder structure in README.md
  • Source code for column2Vec is in folder column2Vec.
  • Tests are in folder test
  • Data are stored in folders data and data_validation.
  • Main folder contains: folder .github, files .gitignore, CONTRIBUTING.MD, LICENSE, README.md, requirements.txt, constants.py and main.py
  • Folder images contains images for README.md

column2Vec folder contains all files for column2Vec feature. More about the structure of this folder can be found here.

Datasets for testing are stored in data and data_validation Corresponding link, name and eventual description for each dataset is stored in DatasetDescription.md in belonging folder (data, data_validation). Both folders contain file DataShow.md with metadata information for each dataset (data, data_validation).

How to run

You can compare two or more tables by running main.py. You can use both comparator and comparatorByColumn, change the comparator in compare_datasets The Result will be distance between tables.

 pip install -r requirements.txt
 python main.py # for fix files
 python main.py data/imdb_top_1000.csv data/netflix_titles.csv # for specific files

You can disable or enable warnings in main by adding these to line for disabling:

warning_enable.change_status(False)
warning_enable.disable_timezone_warn()

Enable by:

warning_enable.change_status(True)
warning_enable.enable_timezone_warn()

DataShow

Is generated by file Dataset_description

How to run tests

Tests are in folder test.

For running tests, you have to switch to a test folder and then run test by using pytest.

cd test

pytest types_test.py #test name to run

Or you can run all the tests by running this:

 python -m unittest
 #or
 pytest
 # or for specific file
 pytest test/test_types.py

Please be aware that some tests in the test_column2Vec module may take a long time.

80 passed, 2 skipped, 500 warnings in 1887.83s (0:31:27)

Static tests

For static tests, we use pylint and black.

Pylint

You can run it by using this command:

pylint $(git ls-files '*.py')
# or for specific file
pylint similarity/metadata_creator.py

You can change pylint settings in .pylintrc file.

Black

You can run it by using this command:

black $(git ls-files '*.py')
# or for specific file
black similarity/metadata_creator.py

You can change black settings in pyproject.toml file.

Coverage

You can run it by using this command:

PYTHONPATH="./similarity:./similarityRunner:$PYTHONPATH" \                      
           coverage run --source='similarity,column2Vec,similarityRunner' -m \
           pytest test/test_similarity* test/test_runner* test/test_column2VecCache.py

How to contribute

Please see our Contribution Guidelines.