Duplicates
Tests dataset for duplicate entries, ensuring model reliability via data quality verification.
Purpose
The ‘Duplicates’ test is designed to check for duplicate rows within the dataset provided to the model. It serves as a measure of data quality, ensuring that the model isn’t merely memorizing duplicate entries or being swayed by redundant information. This is an important step in the pre-processing of data for both classification and regression tasks.
Test Mechanism
This test operates by checking each row for duplicates in the dataset. If a text column is specified in the dataset, the test is conducted on this column; if not, the test is run on all feature columns. The number and percentage of duplicates are calculated and returned in a DataFrame. Additionally, a test is passed if the total count of duplicates falls below a specified minimum threshold.
Signs of High Risk
- A high number of duplicate rows in the dataset, which can lead to overfitting where the model performs well on the training data but poorly on unseen data.
- A high percentage of duplicate rows in the dataset, indicating potential problems with data collection or processing.
Strengths
- Assists in improving the reliability of the model’s training process by ensuring the training data is not contaminated with duplicate entries, which can distort statistical analyses.
- Provides both absolute numbers and percentage values of duplicate rows, giving a thorough overview of data quality.
- Highly customizable as it allows for setting a user-defined minimum threshold to determine if the test has been passed.
Limitations
- Does not distinguish between benign duplicates (i.e., coincidental identical entries in different rows) and problematic duplicates originating from data collection or processing errors.
- The test becomes more computationally intensive as the size of the dataset increases, which might not be suitable for very large datasets.
- Can only check for exact duplicates and may miss semantically similar information packaged differently.