Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpretable Mesomorphic Networks for Tabular Data

Arlind Kadra, Sebastian Pineda Arango, Josif Grabocka

TL;DR

This paper proposes a new class of interpretable neural networks for tabular data that are both deep and linear at the same time (i.e. mesomorphic) and optimize deep hypernetworks to generate explainable linear models on a per-instance basis.

Abstract

Even though neural networks have been long deployed in applications involving tabular data, still existing neural architectures are not explainable by design. In this paper, we propose a new class of interpretable neural networks for tabular data that are both deep and linear at the same time (i.e. mesomorphic). We optimize deep hypernetworks to generate explainable linear models on a per-instance basis. As a result, our models retain the accuracy of black-box deep networks while offering free-lunch explainability for tabular data by design. Through extensive experiments, we demonstrate that our explainable deep networks have comparable performance to state-of-the-art classifiers on tabular data and outperform current existing methods that are explainable by design.

Interpretable Mesomorphic Networks for Tabular Data

TL;DR

This paper proposes a new class of interpretable neural networks for tabular data that are both deep and linear at the same time (i.e. mesomorphic) and optimize deep hypernetworks to generate explainable linear models on a per-instance basis.

Abstract

Even though neural networks have been long deployed in applications involving tabular data, still existing neural architectures are not explainable by design. In this paper, we propose a new class of interpretable neural networks for tabular data that are both deep and linear at the same time (i.e. mesomorphic). We optimize deep hypernetworks to generate explainable linear models on a per-instance basis. As a result, our models retain the accuracy of black-box deep networks while offering free-lunch explainability for tabular data by design. Through extensive experiments, we demonstrate that our explainable deep networks have comparable performance to state-of-the-art classifiers on tabular data and outperform current existing methods that are explainable by design.
Paper Structure (28 sections, 3 equations, 12 figures, 16 tables)

This paper contains 28 sections, 3 equations, 12 figures, 16 tables.

Table of Contents

  1. Introduction
  2. Proposed Method
  3. Shallow Interpretability through Deep Hypernetworks
  4. Interpretable Mesomorphic Networks (IMN)
  5. Explainability Through Feature Attribution
  6. Proof-of-concept: Globally Accurate and Locally Interpretable Classifiers
  7. Related Work
  8. Interpretable Models by Design:
  9. Interpretable Model Distillation:
  10. Interpretable Deep Learning via Visualization:
  11. Experimental Protocol
  12. Predictive Accuracy Experiments
  13. Baselines:
  14. Protocol:
  15. Explainability Experiments
  16. ...and 13 more sections

Figures (12)

  • Figure 1: The IMN architecture.
  • Figure 2: Investigating the accuracy and interpretability of IMN. Left: The global decision boundary of our method that separates the classes correctly. Right: The local hyperplane pertaining to an example $x'$ which correctly classifies the local example and retains a good global classification for the neighboring points.
  • Figure 3: The critical difference diagram for the white-box interpretable methods. A lower rank indicates a better performance over datasets.
  • Figure 4: Black-box methods comparison with critical difference diagrams. Top: The average rank for the binary datasets present in the benchmark. Bottom: The average rank for all datasets present in the benchmark. A lower rank indicates a better performance. Connected ranks via a bold bar indicate that performances are not significantly different ($p > 0.05$).
  • Figure 5: Performance analysis of different interpretability methods over a varying degree of feature correlation $\rho$. We present the performance of all methods on faithfulness (ROAR), monotonicity (ROAR), faithfulness, and infidelity (ordered from left to right) on the Gaussian Linear dataset for $\rho$ values ranging from [0, 1].
  • ...and 7 more figures