In weakly supervised learning, Partial Label Learning (PLL)—sometimes called superset label learning—offers a pragmatic way to collect training data. Instead of forcing an annotator to pinpoint a single exact label for an instance, PLL allows them to provide a set of candidate labels. For example, an image of a rare bird might be labeled with the candidate set {"falcon", "hawk", "osprey"}. The classical formulation of this problem relies on a foundational assumption: the true label is guaranteed to be hidden somewhere within that candidate set.
However, real-world data collection is messy. Annotators make mistakes, automated labeling pipelines omit categories, and out-of-distribution samples creep into datasets. When this happens, the true label may be missing from the candidate set entirely.
Early theoretical work by Li-Ping Liu 0001 et al. in Learnability of the Superset Label Learning Problem warned that a lack of robustness to such noisy supervision could cause learning systems to fail. More recently, Wei Wang demonstrated in Realistic Evaluation of Deep Partial-Label Learning Algorithms that standard PLL frameworks are highly fragile in practical environments because they cannot distinguish between instances where the true label is merely hidden and instances where it is completely absent.
To solve this, a new paradigm has emerged: Noisy Partial Label Learning (NPLL). This direction focuses on relaxing the completeness assumption, designing mechanisms that allow models to look outside the provided candidate sets to find the ground truth.
The Failure Mode of Forced Disambiguation
To understand why standard PLL fails when the ground truth is missing, we must look at how these models learn. Traditional PLL algorithms perform disambiguation: they try to estimate which label inside the candidate set $Y$ is the true label $y^*$. They typically optimize the model to maximize the probability mass distributed across the candidate set while minimizing the probability of labels outside it.
If an image of a dog is mistakenly given the candidate set {"cat", "fox"}, a standard PLL model is forced to choose between "cat" and "fox". It has no mathematical mechanism to say "neither." As a result, the model updates its weights to associate dog-like features with cats or foxes, actively corrupting its own representation space.
To break out of this trap, algorithms must find a way to trust the model's own internal representations over the flawed supervision of the candidate set.
Shifting Trust via Dynamic Label Importance
One core mechanism for surviving missing ground truth is dynamic trust adjustment. Early in training, a deep neural network tends to learn simple, generalized patterns before overfitting to noise. We can leverage this "early learning" phenomenon to correct the candidate sets.
Instead of treating the candidate set as an absolute boundary, the Adjusting Label Importance Mechanism (ALIM) introduced by Mingyu Xu et al. in ALIM: Adjusting Label Importance Mechanism for Noisy Partial Label Learning dynamically balances trust between the initial candidate labels and the model's own predictions. If the model becomes highly confident in a label that sits outside the candidate set, ALIM allows the model to override the annotator's supervision.
Standard PLL: [Candidate Set: A, B] ---> Force model to pick A or B
Noisy PLL: [Candidate Set: A, B] <--- Dynamic Balance ---> [Model Prediction: C]
Similarly, the Iterative Refinement Network (IRNet) developed by Zheng Lian et al. in IRNet: Iterative Refinement Network for Noisy Partial Label Learning treats candidate sets as a soft prior. IRNet iteratively detects instances where the model's predictions strongly conflict with the candidate set, flags them as noisy, and rewrites the target distribution to introduce the missing ground-truth label.
Softening Targets with Label Smoothing
When the ground truth might be missing, forcing a model to make hard, confident predictions on a candidate set is dangerous. A simple yet highly effective countermeasure is to alter the optimization landscape using Label Smoothing (LS).
In Does Label Smoothing Help Deep Partial Label Learning?, Xiuwen Gong et al. proved theoretically that label smoothing provides a robust risk bound against missing labels. By distributing a small amount of probability mass to all non-candidate classes, label smoothing prevents the model from penalizing the true label too harshly if it happens to be excluded from the candidate set.
This insight is operationalized in frameworks like PALS (Pseudo-labelling meets Label Smoothing for Noisy Partial Label Learning) by Darshana Saravanan et al. PALS pairs label smoothing with a pseudo-labeling strategy. The smoothing keeps the model's representations flexible and prevents overconfidence, while the pseudo-labeler gradually shifts probability mass toward the true class, whether it was originally in the candidate set or not.
Reconstructing Labels via Neighborhood Consistency
If an instance's own candidate set is completely corrupted and lacks the true label, where does the clean signal come from? The answer lies in the geometry of the feature space, governed by the manifold assumption: similar instances should share similar labels.
Even if a specific dog image is labeled {"cat", "fox"}, its visual features will still place it near other dog images in the embedding space. If those neighboring images have candidate sets like {"dog", "wolf"} or {"dog", "coyote"}, we can deduce the true label through context.
Target Instance (Corrupted): [Features] ---> Candidate Set: {"cat", "fox"}
Neighbor 1 (Clean): [Features] ---> Candidate Set: {"dog", "wolf"}
Neighbor 2 (Clean): [Features] ---> Candidate Set: {"dog", "coyote"}
|
v
Inferred True Label: "dog"
Xijia Tang et al. formalize this intuition in Neighbor-aware Label Refinement: Enhancing Unreliable Instance-Dependent Partial Labels. Their framework uses a neighborhood-based voting mechanism to iteratively prune incorrect candidates and inject missing labels back into corrupted sets. By aggregating information across local feature neighborhoods, the model corrects individual annotation errors before they can warp the global decision boundaries.
Explicit Out-of-Candidate Modeling
Rather than treating missing labels purely as an error to be smoothed away, some researchers argue that the possibility of an incomplete candidate set should be explicitly integrated into the mathematical formulation of the problem.
In Partial-label Learning with Mixed Closed-set and Open-set Out-of-candidate Examples, Lei Feng et al. partition the problem into distinct "Out-of-Candidate" (OOC) tracking mechanisms. They explicitly model two risks:
- Closed-set OOC: The true label is a known class, but it was accidentally left out of the candidate set.
- Open-set OOC: The instance belongs to an entirely unknown, out-of-distribution class that isn't part of the label space at all.
By explicitly calculating the probability that an instance belongs to an OOC state, their framework prevents the model from forcing square pegs into round holes.
Alternatively, probabilistic graphical approaches like PARD (Partial Multi-Label Learning with Probabilistic Graphical Disambiguation) by Jun-Yi Hang et al. treat the true labels as latent variables within a generative model. Instead of assuming a deterministic mapping from the true label to the candidate set, PARD models the noisy generation process of the candidate sets themselves, allowing the mathematical framework to naturally absorb and correct for missing ground truth.
Key Foundational Work
Learnability of the Superset Label Learning Problem (2014) — contributed 2 concept elements to this direction.
Label Ranking under Ambiguous Supervision for Learning Semantic Correspondences (2010) — contributed 1 concept elements to this direction.