Perspectives on incorporating expert feedback into model updates

Abstract

Machine learning (ML) practitioners are increasingly tasked with developing models that are aligned with non-technical experts' values and goals. However, there has been insufficient consideration of how practitioners should translate domain expertise into ML updates. In this review, we consider how to capture interactions between practitioners and experts systematically. We devise a taxonomy to match expert feedback types with practitioner updates. A practitioner may receive feedback from an expert at the observation or domain level and then convert this feedback into updates to the dataset, loss function, or parameter space. We review existing work from ML and human-computer interaction to describe this feedback-update taxonomy and highlight the insufficient consideration given to incorporating feedback from non-technical experts. We end with a set of open questions that naturally arise from our proposed taxonomy and subsequent survey.

Figure 1

To incorporate an expert’s preferences to improve a model, practitioners must turn non-technical, domain expert preferences into usable model updates In this work, we propose a feedback-update taxonomy that focuses on the ways in which expert feedback can be translated into model updates. Our taxonomy has two axes, expert feedback and model updates, forming six categories of feedback-update interaction. We map existing work onto each category and use our taxonomy to motivate improving the interaction between practitioners and domain experts.

Figure 2

The practitioner generates data, shown in translucent circles, to remove reliance on one feature, after retraining, per the expert’s domain-level feedback The updated dataset is used to learn a new model.

Figure 3

The expert specifies that model behavior should be similar in each color block The practitioner converts this feedback into a regularizer, which is used in the new loss function and obtains an updated model.

Figure 4

For certain types of models, the expert may be able to directly edit the model parameters Here, the expert changes the coefficients of the linear model to obtain their desired model.

Figure 5

The stakeholder labels new points, which are denoted by the purple circles These points are then added to the dataset and used to obtain an updated model.

Figure 6

The expert specifies that the cost of mislabeling a yellow point is higher than the cost of mislabeling a blue point Retraining with a loss function that incorporates the cost of error yields a cost-sensitive model.

Figure 7

The expert adds a new feature to the dataset This addition allows the data points to become linearly separable, which leads to a more accurate model.

Figure 8

Beyond the feedback-update taxonomy introduced in mapping prior work to our taxonomy, we present open questions in the broader pipeline that a practitioner follows to solicit and incorporate feedback from stakeholders

Figure 9

We illustrate four different combinations of feedback type and prompt styles that a practitioner might ask from an expert (A) The practitioner presents a model, and the expert specifies the addition of the striped point with a purple border to the blue class. The induced change by this feedback yields a new model. (B) The practitioner presents the expert with a forced choice to provide a label for the point denoted with a question mark. Depending on the selected label, the induced classifier may be option A or option B. (C) The practitioner presents the expert with a forced choice of two properties: enforce max-margin between two classes (option A) or avoid reliance on the vertical feature (option B). (D) The expert specifies a property that the model should not use the vertical feature to make predictions, inducing the same classifier (option B) in (B) and (C).