Counterfactual Explanations for Oblique Decision Trees:Exact, Efficient Algorithms
Authors: Miguel Á. Carreira-Perpiñán, Suryabhan Singh Hada6903-6911
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments Our algorithm is exact, as we have shown. It is also very efficient for both axis-aligned and oblique trees, even if there are categorical variables; solving a counterfactual explanation in our experiments takes usually milliseconds. We show 3 types of results: an example illustrating how the algorithm can be used interactively; summary results in several datasets, comparing with other algorithms; and, in Carreira Perpi n an and Hada (2021a), results on MNIST digit images, which can be visualized. Table 2: Comparison of counterfactuals generated by our exact algorithm and by searching over the training & test set, over a variety of datasets, using an oblique classification tree. |
| Researcher Affiliation | Academia | Miguel A. Carreira-Perpi n an, Suryabhan Singh Hada Dept. Computer Science & Engineering, University of California, Merced {mcarreira-perpinan, shada}@ucmerced.edu |
| Pseudocode | No | The paper describes algorithms and mathematical formulations but does not contain any structured pseudocode blocks or figures explicitly labeled as "Algorithm" or "Pseudocode". |
| Open Source Code | Yes | Code implementing the algorithms is available from the authors web page. |
| Open Datasets | Yes | We consider the UCI Adult dataset for binary classification... and ...results on MNIST digit images, which can be visualized. |
| Dataset Splits | No | The paper mentions using "training & test set" in the context of comparison methods and selecting "test instances" for generating counterfactuals. However, it does not explicitly provide details about the specific training/validation/test dataset splits (e.g., percentages, sample counts, or defined validation sets) used to train the models in the paper. |
| Hardware Specification | No | The paper states that its algorithm is "very fast" and has a "runtime of a few milliseconds per instance" but does not provide specific details about the hardware (e.g., CPU, GPU models, memory, or cloud resources) used to conduct the experiments. |
| Software Dependencies | No | The paper mentions the use of "CPLEX or Gurobi" as mixed-integer optimization solvers but does not provide specific version numbers for these or any other key software dependencies. |
| Experiment Setup | No | The paper describes the mathematical formulation of the counterfactual problem and various components like cost functions and constraints. However, it does not provide specific experimental setup details such as hyperparameter values (e.g., learning rates, batch sizes, number of epochs) or specific optimizer settings for training the tree or solving the counterfactual optimization problem. |