Slack, D., Friedler, S. A., Roy, C. D., & Scheidegger, C. (2019, January). Assessing the Local Interpretability of Machine Learning Models. In

NeurIPS Workshop on Human-Centric Machine Learning (HCML).

## Paper Summary

This paper defines a metric to measure the local interpretability^{1} of a model i.e. the number of operations the model
requires to make a prediction. The metric is correlated with local interpretability by measuring the latter with a
human-subject study. Users were asked to predict the output of a model given an input, and predict the output given
counterfactual input i.e., slight modifications of the input. The results show that the operation count metric *is*
correlated with local interpretability, and that decision trees and logistic regression models are more interpretable
than neural networks.

## Methodology

The metric defined to act as a proxy for interpretability is the runtime operation count. Effectively it is the number
of operations a human must carry out in order to simulate a single prediction the model. This was performed by
instrumenting the prediction operation for existing trained models in Python’s
`scikit-learn`

package.

## Experimental Setup

The authors ran a crowdsourced experiment $(N=1000)$ using the Prolific platform.
Participants were asked to predict the output of a model on an input (by replicating the actual calculations involved),
and then asked to predict the output again for a slightly modified version of the input. Three model types were used -
decision trees (DT), logistic regression (LR) and neural networks (NN). They were trained using `scikit-learn`

. Training
details are omitted here.

Users were trained in how to perform these calculations for each model type using a small fill in the blank exercise conducted before the actual prediction task. The models were trained using synthetic data to avoid the effects of domain knowledge on prediction.

The survey was administered through Qualtrics. Hypotheses for the study were pre-registered on the Open Science Framework.

## Results and Analysis

Based on the number of correct responses for the two tasks of simulation and counterfactual prediction for each of the three models, three hypotheses were made regarding local interpretability of the models - $DT > NN$, $DT > LR$, $LR > NN$. p-values and confidence intervals were calculated for each using Fisher’s Exact Test.

*likely*the particular assignment of values to each cell are, with the null hypothesis being independence of all categories. It is a special case of the chi-squared test but works for all sample sizes.

The results show that decision trees are more simulatable and “what-if” locally explainable than logistic regression or neural network models. However, the results did not find evidence for logistic regression to be more locally interpretable than neural network models.

Visual comparison of the plots of accuracy vs. operation count across all three tasks (simulatability, “what-if” local
explainability and local interpretability, the last of which simply measures how many users got both of the previous
tasks right), shows that they *are* correlated. Also graphed was the time taken vs. the operation count, and the
accuracy vs. the time taken.

## Discussion

The authors point out how this metric may be used to assess the interpretability of a model without a user study.

## Impressions

This is a neat paper which attempt to quantitatively validate a hypothesis that most people would lend credence to being true. Operation count is of course a very crude way to approximate interpretability, and is absolutely unfeasible for the mammoth network sizes that power today’s models. The paper also doesn’t attempt to answer questions regarding interpretability for different models with similar operation counts. Is a large decision tree equally as interpretable as a small neural network? The graphs seem to show that the mean operation counts across the three model types are significantly different. I would have liked to see how interpretability is affected when the operation counts for different model types are similar.

I also question whether brute calculation of the model to produce outputs is how most people would try to interpret a
model. One can probably build some intuition from the feature weights and response curves to be able to make accurate
predictions for the smaller model types *without* having to actually simulate the prediction function on paper.
Nevertheless, it is a simple and good enough lower bound to start asking questions about interpretability.

local interpretability assesses whether a single decision is interpretable, as opposed to global interpretability which assesses whether the model as a whole is interpretable ↩︎