Abstract
SLISEMAP is an open-source Python library that integrates manifold visualization with local explanations for black-box models. By simultaneously learning a low-dimensional embedding and local interpretable surrogate models, SLISEMAP provides a unified global view of how a complex model's decision-making logic varies across a dataset. This paper outlines the SLISEMAP methodology, its PyTorch-based implementation, and demonstrates its utility using the AutoMPG dataset.
Introduction
In Explainable AI (XAI), researchers often use local, model-agnostic explanations (like LIME or SHAP) to interpret individual predictions. Separately, dimensionality reduction tools (like t-SNE or UMAP) are used to visualize high-dimensional data clusters. However, a gap exists between these two approaches: visual clusters do not necessarily correspond to shared decision logic. SLISEMAP addresses this by creating a 2D embedding where proximity indicates that data points share similar local "white box" explanations, offering a global perspective on local model behavior.
Methodology
The SLISEMAP algorithm takes a data matrix and the predictions from a pre-trained black-box model as inputs. It performs a simultaneous optimization of a low-dimensional embedding and a set of local surrogate models (e.g., linear regression).
The core mechanism involves a "push and pull" logic: if two data points require significantly different local models to explain the black box's behavior, they are pushed apart in the embedding. Conversely, points explained by the same local model are clustered together. This distance-weighted optimization ensures that the visual space directly reflects the logic of the surrogate models.
As shown in the parameter table above, the library allows for fine-tuning of embedding spread (radius) and regularization (lasso/ridge). Built in Python using PyTorch, SLISEMAP supports automatic differentiation and GPU acceleration, interfacing seamlessly with the broader scientific Python ecosystem.
Case Study: AutoMPG Dataset
To demonstrate the library's capabilities, the authors applied SLISEMAP to a random forest regressor trained on the AutoMPG dataset to predict fuel consumption.
The resulting visualization (left) and coefficient analysis (right) reveal distinct clusters of decision logic. For instance, Cluster 0 (blue) identifies a group of heavy, US-made cars where 'weight' is the primary negative predictor of fuel economy. In contrast, Cluster 2 (green) represents lighter, non-US cars where 'model year' and 'horsepower' are the dominant factors. This macro-level view allows users to identify how the black box's logic shifts across different subpopulations of the data.
Conclusion
SLISEMAP provides a powerful framework for bridging the gap between local explanations and global data visualization. Beyond static analysis, the library includes features for projecting new data items to detect concept drift or find alternative explanations. While the simultaneous optimization of embeddings and local models is computationally demanding, the library's support for subsampling ensures it remains practical for real-world datasets. SLISEMAP is available as an open-source package under the MIT license.
Summary updated