ERExplain is a package to explain any Entity Resolution model, using a Black-Box approach
The explanations can be provided into two different ways:
- Attribute-Level: the system shows to the user a histogram reporting the importance of the attributes with respect to the prediction. The importance is a number in [0,1]
- Value-level: the system shows to the user, for each chosen attribute set/single attribute A, the most influential values for A, with respect to the prediction
To use this package you need some simple requirements:
- A model M per Entity Resolution (ER), for example Deepmatcher, but also simpler models, for example a decision tree
- A Pandas dataframe D that contains labeled record pairs for ER and that as in this example:
| label | ltable_product_title | ltable_manufacturer | rtable_product_title | rtable_manufacturer |
|---|---|---|---|---|
| 0 | Nvidia rtx 2060 | Asus | Nvidia Rtx 2070 | MSI |
| 1 | Asus rog strix rtx 2080 super | Asus | Asus Geoforce rtx 2060 s | Asus |
The label 0 indicates a non-matching pair, while label 1 indicates matching pairs. Ltable_ and rtable_ are the prefix for the left and the right attributes respectively. You can use any prefix you prefer
- A list of Pandas dataframes containing the sources of records, each of one read in the following way:
source1 = pandas.read_csv('source1_path',dtype=str)
- A wrapper function that takes in input three parameters in this exact order: the dataframe D, the model M, and the list columns to ignore in D
ERExplain uses some well-know libraries:
To install the package simply run
$ pip install erexplain