Is EERM a universal standard for explainable AI?

No. It names a particular research principle and related implementations.

Does a lower penalty prove better understanding?

Not by itself. It improves the chosen proxy under its assumptions; human understanding needs suitable independent evaluation.

Can the weighting change which model is chosen?

Yes. As the worked example shows, emphasizing the explainability penalty can favor a different model and change the prediction-loss tradeoff.

AI Glossary / Governance, ethics and safety / Intermediate

Explainable empirical risk minimization

Definition

Explainable empirical risk minimization, or EERM, is a research approach that adds a user-related explainability objective to ordinary empirical risk minimization. In the formulation developed by Zhang and colleagues, training balances prediction loss against an information-theoretic measure of how uncertain a model's predictions remain given a signal representing a particular user's understanding. It is a specific proposed method, not a general synonym for every explainable model.

What it means

Ordinary ERM asks which allowed model best fits training examples under a chosen loss. EERM adds another concern: whether the resulting predictions are understandable relative to a specified user signal. Two users may have different background knowledge, so the same model need not be equally predictable or explainable to both under this formulation. The approach makes that user dependence explicit. It does not establish that predictability from a user signal captures every important aspect of explanation, such as causal understanding, useful recourse, or the ability to challenge an incorrect decision.

How it works

The research formulation uses examples containing features, target labels, and a user-related signal. It estimates conditional entropy of the model's predictions given that signal and adds a weighted version of the estimate to average training loss. Lower conditional entropy means less remaining uncertainty under the specified representation. A nonnegative weighting parameter controls how much the training objective penalizes that uncertainty. The paper also presents constrained formulations and implementations for particular model families. Estimation assumptions and the quality of the user signal matter: optimizing a proxy for subjective explainability is not the same as directly proving that a person understands the model.

Three practical examples

A research team studies whether a prediction rule can remain accurate while being easier to anticipate from a domain specialist's recorded judgments.
A prototype compares candidate models using both held-out prediction quality and an explainability-related penalty defined for a particular user group.
A learning study investigates how the selected model changes when the signal used to represent user knowledge changes. These are plausible applications of the research idea, not claims that EERM is a standard deployed feature in commercial products.

Think of it this way

Think of choosing an instructional example that is both correct and connected to what a learner already knows. The learner's existing understanding affects which explanation is accessible. The analogy is limited because EERM uses a mathematical proxy involving predictions and user signals; it does not literally assess all the reasoning a learner can articulate. It helps explain the motivation for a user-dependent objective while preserving the distinction between the proxy and actual comprehension.

Worked example

Suppose two candidate models have illustrative empirical losses of 0.10 and 0.14. Their estimated explainability penalties are 0.50 and 0.10 respectively, with lower penalty preferred. Use the simplified objective loss plus lambda times penalty. At lambda = 0.05, Model A scores 0.10 + 0.05×0.50 = 0.125, while Model B scores 0.14 + 0.05×0.10 = 0.145, so A is preferred. At lambda = 0.20, A scores 0.20 and B scores 0.16, so B is preferred. The example demonstrates a tradeoff in an objective, not a measured improvement in human understanding. The penalty values are hypothetical and their scales must be interpreted consistently; lambda cannot be chosen sensibly without knowing the loss and estimator. Both candidates would still require independent prediction and user evaluation.

Illustrative case study

Fictional case: Rowan Research develops a prototype adviser for experienced equipment planners. It collects task judgments as one possible user signal and investigates an EERM-style objective. A review finds that the signal represents experienced planners better than newcomers. The team therefore limits its claim, evaluates the two audiences separately, and tests whether users can recognize model errors rather than merely predict outputs. This invented case illustrates the importance of validating the user representation, not evidence that the method guarantees explainability or fairness.

Common misconceptions and mistakes

EERM does not simply add a natural-language explanation after an unchanged model makes a prediction. The explainability-related term influences model selection or training itself. It is also not the same as minimizing ordinary prediction error with any regularizer: the regularizer's intended role is specifically tied to the chosen explainability measure. A constant predictor may be easy to anticipate while being useless, which shows why the prediction objective still matters. Matching a user's expectations can preserve mistaken beliefs, so user agreement must not be confused with correctness or ethical acceptability.

Business use

Treat EERM as a specialized research option requiring careful methodological review. Clarify who the intended users are, how their knowledge is represented, and whether collecting that signal is appropriate. Evaluate the tradeoff using held-out data and real user tasks, not only the optimized objective. Report prediction loss, the proxy measure, and comprehension evidence separately. Examine whether different groups of users are represented adequately and whether the chosen signal changes over time. A business should not buy an explainability claim merely because a training objective contains an explanation-related penalty.

Check your understanding

A candidate has a lower EERM objective because it has a smaller user-related penalty, but it makes more prediction errors. Has it unconditionally improved?

Answer: No. It is preferred under the chosen weighted objective. Whether the tradeoff is useful requires separate evaluation of accuracy, user understanding, and the intended task.

Related terms

Learn next

Read the original formulation and identify the features, labels, user signal, loss, and entropy estimator. Explain what each represents in a proposed application. Before optimization, design a user test that could reveal whether the proxy fails to capture meaningful understanding.

Sources

Zhang and colleagues, Explainable empirical risk minimization, original open-access article: https://link.springer.com/article/10.1007/s00521-023-09269-3 Aalto Dictionary of Machine Learning, EERM entry: https://aaltodictionaryofml.github.io/ADictML_Main.pdf

Previous: Explainable AI Next: Explained variance score