Data-Dependent Rectangular Bounding Processes.

Area of Science:

• Machine Learning
• Data Mining
• Computational Geometry

Background:

• Stochastic partition processes segment multi-dimensional spaces for data homogeneity.
• Existing methods create unnecessary divisions in sparse data regions.
• This inefficiency hinders accurate data description, especially in dense areas.

Purpose of the Study:

• Introduce a parsimonious partition model, the Rectangular Bounding Process (RBP).
• Develop a data-dependent RBP (data-RBP) for efficient, sequential partitioning and online learning.
• Demonstrate the RBP and data-RBP's applicability and performance.

Main Methods:

• Employ a bounding strategy using rectangular boxes to enclose data points.
• Extend the RBP to handle infinite spaces and develop the data-RBP for sequential, data-bound partitioning.
• Prove distributional equivalence between RBP and data-RBP (excluding empty boxes).

Main Results:

• The RBP efficiently partitions multi-dimensional spaces, avoiding unnecessary divisions.
• The data-RBP effectively reduces model complexity and enables online learning.
• Validated applications in regression trees, relational modeling, and random feature construction.

Conclusions:

• The RBP offers an efficient and parsimonious approach to multi-dimensional data partitioning.
• The data-RBP enhances model efficiency and supports online learning capabilities.
• Both RBP and data-RBP demonstrate strong performance in accuracy and efficiency across diverse applications.