In standard supervised learning, we assume that our training datasets contain ground-truth labels that accurately represent the entire population. However, when machine learning models are deployed in consequential real-world domains—such as credit lending, hiring, pretrial bail, or medical triage—this assumption fundamentally breaks down.
The breakdown occurs because the model's own decisions dictate whether or not a ground-truth label can even be observed. If a bank approves a loan application, it eventually observes whether the applicant repays or defaults. But if the bank rejects the application, the applicant's true repayment capability remains completely hidden. This phenomenon is known as the selective labels problem, censored feedback, or one-sided feedback.
The Selection Trap and the Blind Spot
To understand why one-sided feedback degrades algorithmic performance, consider a model trained to predict loan repayment. The training data is inherently restricted to individuals who successfully passed a historical decision gate. As noted by Yahav Bechavod et al. in Equal Opportunity in Online Classification with Partial Feedback, this partial feedback loop prevents the system from learning the true labels of the rejected population, creating a persistent information gap.
This missing counterfactual data introduces severe selection bias. If the historical selection process was influenced by human prejudice or flawed heuristics, the training sample becomes systematically skewed, as formalized by Nathan Kallus et al. in Residual Unfairness in Fair Machine Learning from Prejudiced Data.
The danger intensifies when models are iteratively retrained on their own history. In Algorithmic Censoring in Dynamic Learning Systems, Jennifer Chien et al. demonstrate that dynamic learning systems can trap entire subgroups in a self-reinforcing cycle of exclusion. Because the model never observes its false negatives—the individuals it rejected who would have actually succeeded—it remains blind to its own errors.
Worse yet, standard evaluation metrics fail to flag this decline. Because accuracy and fairness metrics are typically calculated only on the observed (approved) subset, the model's performance appears deceptively high while its true accuracy across the broader population plummets.
Shifting from Prediction to Decision via Exploration
How can an algorithmic system break out of this blind spot? The core insight is that we must abandon the traditional paradigm of "learning to predict" and instead focus on "learning to decide."
As argued by Niki Kilbertus et al. in Fair Decisions Despite Imperfect Predictions, deterministic decision thresholds perpetuate historical biases by starving the model of new information. To overcome this, systems must deploy stochastic exploring policies. By occasionally approving individuals who fall just below the current acceptance threshold, the system intentionally incurs a short-term risk to gather the counterfactual data it needs to maximize long-term utility and fairness.
However, unconstrained exploration is economically and socially costly; a bank cannot randomly hand out millions of dollars in high-risk loans purely for data collection. To solve this, Yifan Yang et al. introduce a framework called Adaptive Data Debiasing through Bounded Exploration. Their approach uses controlled deviations from the current optimal policy. By strictly bounding how far the algorithm explores, it recovers unbiased estimates of the population characteristics over time while keeping the immediate costs of bad decisions within acceptable limits.
Structural and Causal Workarounds
Beyond pure exploration, researchers have developed structural and causal mechanisms to infer the missing labels or lower the cost of data acquisition.
Instrumental Variables
When unobserved factors influence both the historical decisions and the final outcomes, standard statistical corrections fail. In Learning with Selectively Labeled Data from Multiple Decision-makers, Jian Chen proposes leveraging the natural variation across multiple human experts (such as different judges or loan officers) as an Instrumental Variable (IV).
Because different experts possess varying degrees of leniency, individuals with identical underlying profiles may be rejected by a strict decision-maker but accepted by a lenient one. This variation acts as a natural experiment, allowing the learning algorithm to mathematically correct for selection bias without requiring explicit counterfactual labels for every rejected individual.
Intermediate Actions
Another practical mechanism involves expanding the decision space beyond a binary "yes" or "no." Yang Liu et al. explore this in Adaptive Bounded Exploration and Intermediate Actions for Data Debiasing by introducing intermediate actions, such as micro-loans or paid internships.
[Applicant Profile]
│
├─► Full Approval ──► High Risk / High-Quality Feedback
│
├─► Intermediate ──► Low Risk / Noisy, Cheap Feedback
│ (e.g., Micro-loan)
│
└─► Full Rejection ──► Zero Risk / No Feedback (Censored)
These intermediate steps serve as a low-cost, low-risk compromise. While they do not provide the pristine feedback of a full approval, they yield noisy, affordable signals that reveal a candidate's true potential, effectively lowering the financial barrier to breaking the selective labels loop.