The Tug of War: Accuracy and Interpretability in Machine Learning Models

Image of a man strugging which way to go — Image

Agenda:

What is model prediction accuracy?
What is model interpretability?
Why care about model interpretability?
A trade-off between accuracy and interpretability
Factors affecting model accuracy and interpretability

What is model prediction accuracy?

The model prediction accuracy is the measure of how well the predictions made by trained model match with the ground-truth values.

What is model interpretability?

The interpretability of the model as the name suggests refers to how well the decision process of trained model is understandable to humans.

Why care about model interpretability?

Model prediction accuracy is crucial, but not always the top priority. Sometimes, we may need to sacrifice some accuracy to make the model more interpretable. Why would we do this?

The importance of accuracy or interpretability depends on the project’s objective. If the goal is to achieve the highest possible accuracy, we might sacrifice interpretability. However, if the project focuses more on understanding the model’s decisions, we would prioritize interpretability over accuracy.

The goal of obtaining an interpretable model is to identify and avoid hidden biases in the data. If a model is trained on biased data, it will produce biased results.

For example, we don’t want a loan approval model that approves applicants from certain locations while rejecting those from other locations. This would indicate a bias in the model, which we aim to avoid.

Let’s take some examples of each case.

Accuracy priority

In cases like cancer prediction, automated trading systems, etc., model accuracy is very important. It doen’t matter whether the trained machine learning model is interpretable or not as long model is giving high accuracy.

Interpretability Priority

In situations such as loan approval process, hiring candidate shortlisting, targeted advertising, etc., model interpretability is very important. We need to know how the output is predicted along with some decent accuracy.

A trade-off between accuracy and interpretability

In every project, we need to balance model accuracy and interpretability. Neural network algorithms function like a black box; just as you can’t see inside a black box, you can’t interpret how these models make predictions. As we delve deeper into deep learning, models become more accurate but also increasingly difficult, if not impossible, to interpret.

On the other hand, many machine learning algorithms, such as linear regression, logistic regression, decision trees, and k-nearest neighbors, are quite interpretable. These can be thought of as “see-through box” algorithms in layman’s terms. However, these algorithms often don’t achieve high accuracy with very complex data.

Factors Affecting Model Accuracy and Interpretability

Model Complexity

Generally, the accuracy of a model increases with its complexity. More complex models tend to make more accurate predictions. However, this complexity reduces interpretability, making it harder to understand how inputs are processed into predictions.

Data Quantity and Quality

The more data you have, the harder it is to interpret patterns within it. The same applies to data diversity; higher diversity makes it more challenging to grasp the data’s true nature. On the other hand, a large quantity and high diversity of data lead to highly accurate models.

Summary

In short, the model accuracy and interpretability both are very important. However, we would want the trade-off between the two based on your project objective.