Provably Improving Expert Predictions with Prediction Sets

Abstract

Automated decision support systems promise to help human experts solve tasks more efficiently and accurately. However, existing systems typically require experts to understand when to cede agency to the system or when to exercise their own agency. Moreover, if the experts develop a misplaced trust in the system, their performance may worsen. In this work, we lift the above requirement and develop automated decision support systems that, by design, do not require experts to understand when to trust them to provably improve their performance. To this end, we focus on multiclass classification tasks and consider automated decision support systems that, for each data sample, use a classifier to recommend a subset of labels to a human expert. We first show that, by looking at the design of such systems from the perspective of conformal prediction, we can ensure that the probability that the recommended subset of labels contains the true label matches almost exactly a target probability value. Then, we identify the set of target probability values under which the human expert is provably better off predicting a label among those in the recommended subset and develop an efficient practical method to find a near-optimal target probability value. Experiments on synthetic and real data demonstrate that our system can help the experts make more accurate predictions and is robust to the accuracy of the classifier it relies on.

A Note on Information Bias and Efficiency of Composite Likelihood

“…Rastogi et al [2022] and Donahue et al [2022] investigate the optimal ways to aggregate independent algorithmic and human predictions, and under what conditions these aggregations outperform individual predictions. Straitouri et al [2022] proposes an algorithm that selects an optimal set of possible labels and presents it to a human expect for final selection. Xu and Dean [2023] and McLaughlin and Spiess [2023] study theoretically how to adjust the design of decision-aid algorithms to counteract human biases.…”

When Are Two Lists Better than One?: Benefits and Harms in Joint Decision-Making

“…One related area of research is "conformal prediction" where the goal is to optimize the subset that the algorithm presents to the human, such as in (Straitouri et al 2022;Wang, Joachims, and Rodriguez 2022;Angelopoulos et al 2020;Vovk, Gammerman, and Shafer 2005;Babbar, Bhatt, and Weller 2022;Straitouri and Rodriguez 2023). This formulation is structurally similar to ours, but often takes a different approach (e.g.…”

Eliciting and Learning with Soft Labels from Every Annotator

“…which have since been used for several applications, from benchmarking the reliability of foundation models (Tran et al 2022), to informing human-machine teaming (Babbar, Bhatt, and Weller 2022;Straitouri et al 2022), and expanding the empirical understanding of the impact of labels on performance (Wei et al 2022;Schmarje et al 2022). While these labels are very valuable for the community, they are often touted as representing human "label uncertainty" (Tran et al 2022).…”