Multi-Label Few-Shot Learning

• Multi-label few-shot learning is a setting for predicting multiple target labels per instance with limited examples, emphasizing complex label co-occurrence challenges.
• The methodology extends prototypical networks by incorporating metric-based, label-combination, and propagation strategies along with adaptive thresholding.
• Empirical studies show that models leveraging label semantics, hierarchical smoothing, and graph-based reasoning achieve significant gains in F1 and mAP metrics.

Multi-label few-shot learning (ML-FSL) concerns the problem of learning to predict multiple target labels for each input instance when only a small number of examples per label are available. This setting is highly relevant in domains where annotation costs or rare phenomena limit data availability per class, and co-occurrence among labels (e.g., music tags, attributes, entity types) is the norm rather than the exception. Unlike single-label few-shot learning, in ML-FSL, each instance may be associated with an arbitrary subset of the label set, inducing more complex combinatorial structure in inference, training semantics, evaluation, and metric learning.

1. Problem Definition and Formalization

ML-FSL generalizes the classical N-way K-shot episodic meta-learning setup, where episodes are constructed by sampling a subset of N atomic labels and K support instances per label. Each support and query example is associated with a multi-hot label vector $y \subseteq \mathcal{Y}$, $\mathcal{Y}$ being the set of possible labels. Instances can thus belong to arbitrary subsets of the active episode's labels (not just one), and few-shot learning must generalize not only to new labels, but also to new co-occurrence patterns and label combinations.

Explicitly, let $S = \{(x_i, y_i)\}_{i=1}^{NK}$ be the support set, where each $x_i$ is an instance (e.g., image, audio, text) and $y_i \subseteq \mathcal{Y}$ is a subset of active labels. The query set $Q = \{x_q\}$ contains examples whose multi-label assignments must be predicted. The goal is to infer multi-label predictions for novel instances under limited supervision for each atomic label and for the possible label combinations observed in support (Papaioannou et al., 2024, Simon et al., 2021).

This setup induces several challenges beyond those of single-label FSL:

• Non-trivial overlapping of label supports, leading to variable support size per atomic label.
• Arbitrary, often unobserved, combinations of target labels in query instances.
• Difficulty of label cardinality estimation and thresholding for multi-label outputs.
• Exponential growth of the potential label combination space as $|\mathcal{Y}|$ increases.

2. Representative Model Architectures

Metric-Based and Prototypical Extensions

The most widely adapted paradigm is metric learning with class/label prototypes, extended from single-label prototypical networks. For a label $k$, its prototype is typically the mean of embedded support instances possessing $k$: $$p_k = \frac{1}{|S_k|} \sum_{(x_i, y_i) \in S: k \in y_i} f_\phi(x_i)$$ where $\mathcal{Y}$0 is the embedding function, and $\mathcal{Y}$1 is the set of support instances with label $\mathcal{Y}$2 (Liang et al., 2022, Simon et al., 2021, Hu et al., 2021).

Label-Combination Prototypical Networks (LC-Protonets)

A more expressive generalization constructs prototypes not only for atomic labels, but for every unique nonempty subset (combination) of active labels appearing in the support, forming so-called "LC-classes." Each label combination $\mathcal{Y}$3 observed in support has a prototype: $\mathcal{Y}$4 Inference assigns queries to the closest LC-class prototype, with ties resolved in favor of higher cardinality (more labels) (Papaioannou et al., 2024). This offers superior "positive coverage" of the combinatorial multi-label space compared to atomic label prototypes.

Transductive and Propagation Approaches

Label Propagation Networks adopt affinity-based graph inference in the support-query combined set, diffusing label information via a constructed similarity graph (e.g., k-nearest neighbor affinities over embeddings) and performing closed-form harmonic propagation (Simon et al., 2021). This technique enables the rapid spread of label assignments in few labeled, transductive (within-episode) setups, and can be combined with other metric learning approaches.

Models Exploiting Label Semantics and Taxonomy

Several models incorporate external label knowledge (e.g., word embeddings, label descriptions, taxonomies, or co-occurrence statistics) to address the sparsity and generalization limitations:

• Attention-guided prototypes with label word vectors or attributes (Yan et al., 2021, Liang et al., 2022);
• Taxonomy-aware, hierarchical smoothing and prototype-alignment (Liang et al., 2022);
• GNN-based routing and message passing to transfer label co-occurrence and semantic structure to new labels (Chen et al., 2020, Lu et al., 2020).

Label Count and Thresholding Modules

Given the cardinality ambiguity of multi-label output, auxiliary modules are often introduced:

• Neural Label Count (NLC) modules predict the number of active labels per query instance, enabling decision thresholding over scores (Simon et al., 2021, Liu et al., 2022, Hu et al., 2021).
• Learnable or nonparametric thresholding functions are also used, for instance, interpolating between meta-learned and data-calibrated thresholds (Hou et al., 2020).

3. Training, Objective Functions, and Inference

ML-FSL models are episodically trained with multi-label support and query structures. The most common loss is binary cross-entropy over either atomic labels or observed label combinations, typically applied independently per label (sigmoid BCE) or per combination (in combination-prototype frameworks). For example, given model outputs $\mathcal{Y}$5, ground-truth $\mathcal{Y}$6, and label set size $\mathcal{Y}$7,

$\mathcal{Y}$8

(Papaioannou et al., 2024, Liu et al., 2022, Simon et al., 2021). Augmented objectives may include MSE between normalized outputs, supervised contrastive losses to drive apart incompatible label embeddings (Liu et al., 2022), or reinforcement-style rewards for learned thresholds.

The assignment of predictions is predominantly threshold-based, with auxiliary count modules or policies calibrating per-example thresholds to determine which label scores are promoted to predicted outputs. Some approaches use ranking of label scores with estimated cardinality (top-$\mathcal{Y}$9) selection.

4. Empirical Results and Benchmarks

A range of benchmarks, including MS-COCO, Open MIC, ImageNet variants, MagnaTagATune, FSD-FS, FewAsp (YelpAspect), TourSG (intent detection), and large-scale LMTC corpora, have been employed.

Models such as LC-Protonets have decisively outperformed both one-vs-rest atomic metric approaches and earlier multi-label prototypical networks, especially on Macro-F1 and Micro-F1 metrics over both base and novel label sets, predominantly in regimes with higher label combination diversity (Papaioannou et al., 2024).

Graph and knowledge-based approaches (KGGR, multi-graph GCNs) have yielded best-in-class results in vision and clinical text, with mAP/nDCG gains of 4–7 points over prior meta-learners and feature augmentation baselines (Chen et al., 2020, Lu et al., 2020). Label-propagation, neural label count, and adaptive thresholding have shown consistent 1–5pp improvements in mAP and improved label-count accuracy across image and fashion attribute tasks (Simon et al., 2021).

End-to-end label relation learning using instance-level graphs with dual support/query losses substantially improves over prior methods for multi-label intent detection, achieving Macro-F1 gains of 10–15pp in the strictest (1-shot) regime (Zhao et al., 9 Oct 2025). Taxonomy-informed label smoothing and prototype alignment produce 5–6pp gains in mAP/F1 over flat prototypical baselines in audio event recognition (Liang et al., 2022).

5. Scalability, Complexity, and Domain Adaptation

The principal computational challenge is the combinatorial scaling of the label combination prototype set: for $S = \{(x_i, y_i)\}_{i=1}^{NK}$0 atomic labels, the worst-case number of nonempty label combinations is $S = \{(x_i, y_i)\}_{i=1}^{NK}$1, which can be intractable for even moderate $S = \{(x_i, y_i)\}_{i=1}^{NK}$2. Empirical results show N=30 yielding $S = \{(x_i, y_i)\}_{i=1}^{NK}$3 combinations per episode, with inference times scaling proportional to $S = \{(x_i, y_i)\}_{i=1}^{NK}$4 (Papaioannou et al., 2024). Prototype deduplication and hierarchical grouping, as well as leveraging strong base-model pretraining, can partially mitigate these costs.

Approaches that incorporate label hierarchies, statistical co-occurrence, or graph-based propagation (as in knowledge-graph methods) provide scalable inductive biases and substantial gains in both few-shot and zero-shot generalization, especially where label ontologies or external descriptions are available (Chen et al., 2020, Lu et al., 2020, Chalkidis et al., 2020).

The domain adaptation problem—cross-dataset and cross-domain transfer—has been addressed by episodes that vary both the support-query label overlap and data augmentation strategies, as in the GenCDML-FSL protocol (Aimen et al., 2023). Episodic inclusion of "finetune" sets matching target label spaces and simulating domain shifts during meta-training yields systematic improvements in both accuracy and expected calibration error, demonstrating that robust meta-training strategies are as critical as architectural choices.

6. Key Insights, Model Selection, and Open Questions

Several insights have emerged:

• Explicit modeling of label combinations (e.g., LC-Protonets), hierarchy, or relations yields richer representation and improved coverage of the label space at the cost of increased complexity (Papaioannou et al., 2024, Liang et al., 2022).
• Strong pretraining on large-scale, single-label or multi-label corpora, followed by episodic fine-tuning or even zero-shot transfer via prototype recombination, is effective and sometimes sufficient for moderate $S = \{(x_i, y_i)\}_{i=1}^{NK}$5 (Papaioannou et al., 2024, Yan et al., 2021).
• Adaptive or label-count-based thresholding is critical, as multi-label FSL models are highly sensitive to the choice of decision threshold, especially when label prevalence varies (Simon et al., 2021, Hou et al., 2020, Liu et al., 2022).
• Incorporation of external label semantics, either via text descriptions, embeddings, or label graphs, is a consistently beneficial inductive bias, particularly in zero/few-shot settings and for rare or novel label types (Chen et al., 2020, Yan et al., 2021, Lu et al., 2020, Chalkidis et al., 2020).

Open directions include handling extreme-scale label spaces with highly imbalanced distributions, optimizing inference for high-cardinality multi-labels, learning with hierarchical label dependencies beyond parent-child pairs, and developing continual and incremental update protocols for ML-FSL with non-stationary label or domain distributions.

7. Summary Table: Model Classes and Key Features

Class	Key Idea	Sample Works
Atomic Prototypical Nets	Prototype per atomic label, one-vs-rest BCE	(Simon et al., 2021, Liang et al., 2022)
Label-Combination ProtoNets	Prototype for every observed label subset	(Papaioannou et al., 2024)
Label Propagation	Transductive label diffusion across embeddings	(Simon et al., 2021)
Label Semantics/Trees	Word vectors, taxonomy-aware smoothing	(Yan et al., 2021, Liang et al., 2022)
Graph-based (KGGR, KAMG)	GNNs over label co-occurrence/taxonomy/similarity	(Chen et al., 2020, Lu et al., 2020)
Instance Relation Networks	End-to-end graph on support/query, label propagation	(Zhao et al., 9 Oct 2025)
Threshold/Count Modules	Neural or nonparametric label cardinality estimates	(Simon et al., 2021, Liu et al., 2022, Hou et al., 2020)

Further details about architectural specifics, loss terms, episode structuring, evaluation metrics, and transfer results can be found in the referenced works. These approaches collectively constitute the current landscape of multi-label few-shot learning research, with continued innovation at the intersections of metric learning, meta-learning, graph-based reasoning, and incorporation of structured label knowledge (Papaioannou et al., 2024, Simon et al., 2021, Chen et al., 2020, Liang et al., 2022, Zhao et al., 9 Oct 2025, Lu et al., 2020, Chalkidis et al., 2020).