Parameterized exact fair decision tree learning with small sensitive group regimes

Fair decision tree learning under fairness constraints presents a challenge due to the computational complexity of exact optimization. This paper introduces a memoized exact-search framework for decision trees that optimizes under explicit statistical-parity constraints. The memoization technique reduces redundant computations while maintaining exactness.

Through extensive evaluations encompassing scalability, caching ablations, comparisons with other algorithms, benchmark repeats, and real-world dataset validation, the authors demonstrate that the proposed method achieves high test accuracy even when faced with small sensitive group regimes. By employing the memoized exact search, the test accuracy surpasses that of full-training CART, same-120-subset CART, and a demographic-parity reduction baseline when using a depth-3 CART.

The study also confirms that test-set statistical-parity differences remain comparable, and training-set feasibility is preserved. Supported by grants from the Fujian Provincial Department of Science and Technology and Quanzhou Normal University, the authors contributed equally to this work, with Zhigao Huang, Miao Pan, and Shiyan Zheng as the primary contributors.

The findings provide a practical approach to balancing runtime, accuracy, and parity control for scenarios involving small encoded optimization subsets. This research is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, requiring proper attribution to the original authors and source.