v1.3.2

Runtime improvements, bug fixes, improved visualization for instance-cost-sensitive

• equivalent points bound for cost-complex-accuracy
• purification bound for sparse objectives
• precompute feature pairs per instance to speed up d2-search
• filter features for d2-search
• skip zero-cost labels for d2-search
• improve sim.lb execution
• progress tracker for verbose execution
• search for trees of increasing depth
• precompute the equivalence of two feature vectors with unique vector ids
• improved the visualization of instance-cost-sensitive trees
• bug fix: check if a feasible solution exists (not optimal means still feasible)
• bug fix: computing the combination of the left and right lower bound previously ignored the cached left and right lower bound
• bug fix: ignore the time limit in the reconstruct phase
• bug fix: pass the optimization task in the STreeDRegressor (was previously ignored)
• bug fix: do not recompute the hash for dataviews
• bug fix: off-by-one error for the subtract UB procedur

v1.3.1

This release fixes a bug in the cost-complexity accuracy objective

Previously, this objective mistakenly rounded down the cost-complexity to an integer. This is fixed now.

v1.3.0 Regression

This release adds regression, both piecewise constant and piecewise (simple) linear regression

This version is used in the experiments of the ICML-24 paper on regression trees

• Van den Bos, M., Jacobus G. M. van der Linden, and Emir Demirović. "Piecewise Constant and Linear Regression Trees: An Optimal Dynamic Programming Approach." In Proceedings of ICML-24, 2024.

v1.2.3

This release adds the instance-cost-sensitive classifier

v1.2.2

This release adds the script to upload pystreed to pypi

v1.2.1

This release introduces a small bug fix in the computation of the lower bound within the loop of finding the best branching feature.

While maximizing accuracy, in our tests on 93 different data sets, for max_depth=4, we found one data set for which the bug gave a misclassification score one higher than optimal. This release fixes this bug.

v1.2.0

This release adds survival analysis

This version is used in the camera-ready version of

Huisman, T., Jacobus G. M. van der Linden, and Emir Demirović. "Optimal Survival Trees: A Dynamic Programming Approach." Proceedings of AAAI-24. 2024.

v1.0.1

This release is the version used for the experiment result in the camera-ready version (Oct. 24, 2023) of the following paper:

Van der Linden, Jacobus G. M., Mathijs M. de Weerdt, and Emir Demirović. "Necessary and Sufficient Conditions for Optimal Decision Trees using Dynamic Programming." Advances in Neural Information Processing Systems (accepted, not published). 2023.

Updates since v1.0.0:

• removed a bug in the python bindings for the gcc and clang compilers
• added github workflows for cross-platform building with cmake and pip

v1.0.0

This release is the version used for the experiment result in the camera-ready version (Oct. 24, 2023) of the following paper:

Van der Linden, Jacobus G. M., Mathijs M. de Weerdt, and Emir Demirović. "Necessary and Sufficient Conditions for Optimal Decision Trees using Dynamic Programming." Advances in Neural Information Processing Systems (accepted, not published). 2023.