Explaining Groups of Points in Low-Dimensional Representations

Ruud van Bakel, Abhijith Chintam, Andreas Hadjipieris, Roel Kuiper

Keywords: Reproducibility Challenge, Reproducibility, Explainability, Transparency, Explainable AI, Transparent AI

Abstract:

Our paper:

The main claims of the paper Explaining Groups of Points in Low-Dimensional Representations include an introduction of a new type of explanation - Global Counterfactual Explanation (GCE) which is relatively sparse and is consistent i.e., symmetrical and transitive among all the groups.The paper also claims that the explanations based on Transitive Global Translations (TGT) algorithm are better than Difference Between the Mean (DBM) baseline for varying degrees of sparsity. The TGT algorithm is also claimed to be capturing the real signals in the data.
For reproducing the paper we decided to convert the Tensorflow 1.0 implementation of the authors to a PyTorch implementation. We also reproduced the results from the original code base to verify the primary claims of the paper. We also replicated the results for a new Crop mapping using fused optical-radar dataset.
We were successfully able to reproduce the results of the paper. The results of TGT's coverage, correctness and sparsity in comparison with DBM were similar to those claimed by the authors on all datasets.

It was reasonably easy to run the authors' original Tensorflow code. The authors have have clearly separated the experiments on different datasets from the core implementation of the algorithm which helped greatly while re-implementing in Pytorch. Additionally, the notebooks provided for each of the experiments streamlined reproducing the results.

A Readme file explaining the steps to setting up and running the code would have made it a much better experience. We faced two challenges in particular while running the authors' code. Since the authors had some fixed paths to files in the code base, running the code required minor adjustments. Additionally, the python library dependencies were not clearly mentioned, these could have been provided as environment details. For example the scvis library was used but not mentioned in the readme of the code base.

Communication with the original author's was not necessary to reproduce the code.

Original paper:

A common workflow in data exploration is to learn a low-dimensional representation of the data, identify groups of points in that representation, and examine the differences between the groups to determine what they represent. We treat this workflow as an interpretable machine learning problem by leveraging the model that learned the low-dimensional representation to help identify the key differences between the groups. To solve this problem, we introduce a new type of explanation, a Global Counterfactual Explanation (GCE), and our algorithm, Transitive Global Translations (TGT), for computing GCEs. TGT identifiesthe differences between each pair of groups using compressed sensing but constrains those pair-wise differences to be consistent among all of the groups. Empirically, we demonstrate that TGT is able to identify explanations that accurately explain the model while being relatively sparse, and that these explanations match real patterns in the data.

Paper Url: https://openreview.net/forum?id=MFj70_2-eY1