shapiq: An Interactive Feature Explanation Tool for Machine Learning Models

shapiq is a Python package for calculating Shapley interaction values of any order. It extends the traditional SHAP method, enabling the quantification and visualization of synergistic effects between features to help understand complex interaction relationships in model predictions. It fills a gap in the field of machine learning explanation tools, deepening from single-feature contribution analysis to the study of synergistic effects of feature combinations, and is applicable to high-risk decision-making scenarios such as medical diagnosis and financial risk control.

In the field of machine learning model explanation, SHAP has become a standard tool, but traditional SHAP mainly focuses on the main effects of individual features and ignores the interactions between features. In real-world scenarios, the synergistic effects of feature combinations often better explain the model's decision logic. The shapiq project was born to address this limitation, extending SHAP's capabilities to support the calculation of Shapley interaction values of any order.

Shapley interaction values are an extension of Shapley values in game theory. Features are regarded as game players, and model predictions as payoffs. Traditional Shapley values calculate the marginal contribution of individual features, while Shapley interaction values quantify the synergistic effects of feature combinations (order 2 and above). For example, in house price prediction, the positive interaction between "number of bedrooms" and "house area" will increase the value more than the sum of their individual contributions, and shapiq quantifies such relationships through indices like k-SII.

shapiq provides multi-level functions: end users can calculate and visualize interaction values through the concise shapiq.Explainer API; researchers can use the approximator and games modules to develop new methods. It supports multiple indices (k-SII, FBII, FSII), and users can specify the interaction order (order 2 and above).

Usage flow: Prepare the trained model and reference dataset → Create an Explainer (specify index type and maximum order) → Call explain to get InteractionValues results. Visualization methods include force-directed graphs (showing features and interactions for a single prediction) and network graphs (presenting the topology of second-order interaction strengths). When interpreting, positive values indicate positive synergy, and negative values indicate inhibitory effects.

For high-dimensional scenarios, shapiq provides the ProxySPEX approximator, which reduces computational complexity through sparse approximation; it supports TabPFN model explanation, providing TabPFNExplainer that uses the "remove-recontextualize" paradigm to calculate Faithful Shapley values, demonstrating adaptability to cutting-edge models.

shapiq deepens model explanation from "which features are important" to "how feature combinations work together", which is crucial for high-risk decision-making scenarios. It improves model transparency, discovers non-intuitive patterns and potential biases, guides feature engineering optimization, and as a supplement to the SHAP ecosystem, provides strong tool support for responsible AI applications.