Source-linked AI summary
ALIME: Autoencoder Based Approach for Local Interpretability
Sharath M. Shankaranarayana, Davor Runje
TL;DR
Opaque deep learning models, including those used in healthcare, are difficult to interpret, and LIME explanations can be unstable or lack local fidelity. The paper modifies LIME with an autoencoder-based weighting function and reports improved stability and local fidelity in experiments.
Problem
Opaque deep learning models are difficult to interpret, while local explanation methods face questions about stability and local fidelity.
Source: §1 · pp. 1, 3Method
The paper modifies LIME by using an autoencoder as the local model's weighting function.
Source: §1 · p. 3Results
The authors report that their method yields better stability as well as local fidelity in empirical experiments.
Source: §4 · p. 9Takeaways & Limitations
The paper presents an autoencoder-based approach for explaining predictions on tabular data within the LIME framework.
Source: §4 · p. 9Using training examples for local explanations can overfit and may not reflect black-box behavior on test instances.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Machine learning and especially deep learning have garneredtremendous popularity in recent years due to their increased performanceover other methods. The availability of large amount of data has aidedin the progress of deep learning. Nevertheless, deep learning models areopaque and often seen as black boxes. Thus, there is an inherent need tomake the models interpretable, especially so in the medical domain. Inthis work, we propose a locally interpretable method, which is inspiredby one of the recent tools that has gained a lot of interest, called localinterpretable model-agnostic explanations (LIME). LIME generates singleinstance level explanation by artificially generating a dataset aroundthe instance (by randomly sampling and using perturbations) and thentraining a local linear interpretable model. One of the major issues inLIME is the instability in the generated explanation, which is caused dueto the randomly generated dataset. Another issue in these kind of localinterpretable models is the local fidelity. We propose novel modificationsto LIME by employing an autoencoder, which serves as a better weightingfunction for the local model. We perform extensive comparisons withdifferent datasets and show that our proposed method results in bothimproved stability, as well as local fidelity.
1 Introduction and related work
The paper asks whether local explanations can gain stability while retaining fidelity, and proposes an autoencoder-based modification to LIME, evaluated on three healthcare datasets. It situates this goal amid the interpretability needs of opaque deep models and challenges in choosing local neighborhoods.
1 Introduction and related work
- Opaque deep neural networks are increasingly used in healthcare tasks such as diagnosis and survival analysis, motivating efforts to assess whether they are correct, fair, unbiased, and ethical.Source: §1 · p. 1
- Local explanations fit an interpretable model around a test instance; choosing the right neighborhood is non-trivial and can affect surrogate fidelity [8].LIME generates perturbed points and fits a local linear model; alternatives use training-data clusters or hierarchical clustering.Source: §1 · p. 2
- Post-hoc, model-agnostic methods such as LIME explain already-trained models without requiring access to their inner workings.The section also distinguishes example-based, local, and global explanations.Source: §1 · pp. 1–2
- Random perturbation can make LIME's feature weights vary across runs, while simply enlarging the sampled set may improve stability but reduce local fidelity.The paper asks whether stability can improve while local fidelity is maintained.Source: §1 · pp. 2–3
- Using training examples for local explanations can overfit by matching training instances to their classes and may not reflect black-box behavior on test instances.The paper contrasts this concern with methods that use training-set clusters.Source: §1 · pp. 2–3
- ALIME modifies LIME with a novel weighting function to address stability and local fidelity, and the authors compare it with LIME on three healthcare datasets.The paper frames these as its central method and experimental contributions.Source: §1 · p. 3
2 Methods
LIME explains individual predictions by fitting a weighted local surrogate to black-box predictions on generated samples. ALIME replaces the ordinary proximity weighting with an autoencoder-based latent-space procedure that selects and weights nearby points.
2.1 LIME
- LIME samples and perturbs data, queries the black box for predictions, then fits an interpretable model weighted by proximity to the instance being explained.The surrogate should approximate black-box predictions locally, not necessarily globally; its objective uses a loss, model-complexity term, and neighborhood measure, though LIME optimizes only the loss part.Source: §2.1 · pp. 3–4
2.2 ALIME
- ALIME pre-samples Gaussian data and uses a denoising autoencoder trained on standardized, noise-corrupted training data to obtain latent representations.The autoencoder is trained to reconstruct uncorrupted inputs using standard L2 loss; pre-sampling also reduces sampling operations when explaining instances.Source: §2.2 · p. 5
- ALIME computes latent-space distances, discards points beyond a threshold, and exponentially weights selected points before fitting the local linear model.The paper presents the autoencoder as a weighting function that supports locality and is intended to improve local fidelity.Source: Figure 1 · p. 5, §2.2 · p. 6
3 Experiments and Results
Experiments on three healthcare datasets compare ALIME with LIME for local fidelity and stability. ALIME performs better on both measured dimensions across the datasets.
3 Experiments and Results
- The experiments use Breast Cancer, Hepatitis Patients, and Liver Patients datasets with a one-hidden-layer neural network and a 70–30 training-test split.The datasets contain 699, 155, and 583 patient observations, respectively.Source: §3 · p. 7
- ALIME clearly outperforms LIME on local-surrogate R2 and test-set mean MSE across all three datasets.The figures vary the number of points used to train the local surrogate in log scale.Source: §3 · p. 8, Figure 3 · p. 8
- ALIME outperforms LIME on mean coefficient standard deviation and standard-deviation-to-mean stability measures across the three datasets.The metrics average across features over 10 iterations; absolute coefficient values are used when computing means and standard deviations.Source: §3 · p. 9, Figure 4 · p. 9
4 Conclusion
The paper proposes an autoencoder-weighted modification to LIME for explaining tabular predictions and reports empirical improvements in stability and local fidelity, while noting that more analysis is needed.
4 Conclusion
- The proposed approach modifies LIME by using an autoencoder as the weighting function for explanations of tabular predictions.Source: §4 · p. 9
- Experiments report improved stability and local fidelity, but the authors say a more thorough analysis is needed to substantiate the improvements.Source: §4 · p. 9
- Future work is to add theoretical analysis and exhaustive empirical evaluation spanning different data types.Source: §4 · p. 9