validmind.BivariateScatterPlots

BivariateScatterPlots

@tags('tabular_data', 'numerical_data', 'visualization')

@tasks('classification')

defBivariateScatterPlots(dataset):

Generates bivariate scatterplots to visually inspect relationships between pairs of numerical predictor variables in machine learning classification tasks.

Purpose

This function is intended for visual inspection and monitoring of relationships between pairs of numerical variables in a machine learning model targeting classification tasks. It helps in understanding how predictor variables (features) interact with each other, which can inform feature selection, model-building strategies, and identify potential biases or irregularities in the data.

Test Mechanism

The function creates scatter plots for each pair of numerical features in the dataset. It first filters out non-numerical and binary features, ensuring the plots focus on meaningful numerical relationships. The resulting scatterplots are color-coded uniformly to avoid visual distraction, and the function returns a tuple of Plotly figure objects, each representing a scatter plot for a pair of features.

Signs of High Risk

  • Visual patterns suggesting non-linear relationships, multicollinearity, clustering, or outlier points in the scatter plots.
  • Such issues could affect the assumptions and performance of certain models, especially those assuming linearity, like logistic regression.

Strengths

  • Scatterplots provide an intuitive and visual tool to explore relationships between two variables.
  • They are useful for identifying outliers, variable associations, and trends, including non-linear patterns.
  • Supports visualization of binary or multi-class classification datasets, focusing on numerical features.

Limitations

  • Scatterplots are limited to bivariate analysis, showing relationships between only two variables at a time.
  • Not ideal for very large datasets where overlapping points can reduce the clarity of the visualization.
  • Scatterplots are exploratory tools and do not provide quantitative measures of model quality or performance.
  • Interpretation is subjective and relies on the domain knowledge and judgment of the viewer.