DI2

DI2 discretizes heterogeneous clinical and molecular variables using an unsupervised, prior-free, multi-item approach to produce robust discrete representations for downstream associative modeling and classification.

Key Features:

• Unsupervised, prior-free discretization: Performs discretization without predefined distributional assumptions or prior parameters.
• Multi-item assignment: Assigns multiple items to values near discretization boundaries to enhance robustness around border values.
• Skewed-distribution handling: Explicitly manages variables with arbitrarily skewed distributions common in biological datasets.
• Outlier-aware, data-regularity preservation: Considers underlying data regularities and the presence of outlier values to respect natural data structure.
• Compatibility with associative models and classifiers: Produces discretizations suited to state-of-the-art associative models and classifiers capable of handling border values.
• Statistical validation: Statistical tests reported demonstrate that DI2 generally outperforms well-established discretization methods with statistical significance.

Scientific Applications:

• Associative modeling and biomedical data mining: Provides discrete inputs compatible with state-of-the-art associative models used for biomedical data mining.
• Classification and predictive modeling: Improves predictive accuracy in classification tasks by accommodating border values and dataset heterogeneity.
• Analysis of clinical and molecular variables: Handles the diversity and heterogeneity of clinical and molecular variables within biomedical datasets.
• Method benchmarking and comparison: Supports comparative evaluation of discretization methods via statistical performance assessment.

Methodology:

Unsupervised, prior-free multi-item discretization that assigns multiple items near boundaries and accounts for arbitrarily skewed distributions, underlying data regularities, and outlier values.