Odd-One-Out Anomaly Detection

• Odd-one-out anomaly detection is a method that identifies observations deviating from dominant grouping patterns through semantic, perceptual, or structural comparisons.
• It leverages diverse techniques—from cosine similarity in semantic models to a-contrario expectation and latent-space reconstruction—across text, visual, and graph data.
• Empirical results in benchmarks like CIFAR-10 and MNIST show high accuracy, though challenges remain in detecting subtle anomalies and addressing training set contamination.

Odd-one-out anomaly detection denotes a family of detection problems in which an observation is identified as anomalous because it is the element that does not conform to the dominant grouping, neighborhood, or reference structure induced by the rest of the data. Across the literature, the formulation appears in several closely related forms: identifying out-of-distribution examples by their reduced semantic similarity to training data, detecting the odd-looking object in a multi-object scene by comparing it with neighboring objects, finding unknown objects as instances that do not fit any superclass of known categories, marking anomalous tokens in sequences by their “oddballness,” and isolating anomalous graphs or temporal networks as graph-level or time-specific deviations from normal patterns (Rafiee et al., 2020, Bhunia et al., 2024, Yavuz et al., 2024, Graliński et al., 2024, Wang et al., 2024, Kandanaarachchi et al., 2022). The term is therefore both a specific task design and a broader design principle: anomaly is defined relationally, either against a majority grouping or against a learned model of what is semantically, structurally, or temporally coherent (Mohammad, 2021, Salehi et al., 2021).

1. Conceptual foundations

Odd-one-out formulations differ from fixed-criterion anomaly detection by making normality contextual. In the scene-centric formulation, the task is to identify “odd-looking” objects within a scene by comparing them to other objects present, with anomalies defined relative to a scene-specific reference group rather than by a dataset-wide abnormality criterion (Bhunia et al., 2024). In open-world object detection, the same principle appears as identifying unknown objects as instances that do not belong to any of the grouped known classes, using an odd-one-out scoring mechanism over superclasses (Yavuz et al., 2024). In text, odd-one-out detection is instantiated at token level, where anomaly depends not only on low likelihood but on how strange a token is relative to the full probability distribution produced by a LLM (Graliński et al., 2024).

A second conceptual strand defines the odd observation through grouping and expectation. The perception-based formulation draws on the Gestalt School of Psychology and the Helmholtz principle, assuming that anomalies are observations that are unexpected to occur with respect to certain groupings made by the majority of the data (Mohammad, 2021). Under that view, the “odd-one-out” is the observation whose occurrence is sufficiently improbable under an a-contrario model, with the operative criterion being that the expectation of the number of occurrences is less than one (Mohammad, 2021).

The literature also uses the principle outside direct anomaly scoring. “Odd-one-out learning” in self-supervised video representation learning constructs a pretext task in which a network must identify an unrelated or odd element from a set of otherwise related video subsequences; the odd subsequence is formed by wrong temporal order, while the even ones preserve correct order (Fernando et al., 2016). This is not anomaly detection in the deployment sense, but it operationalizes the same relational judgment: one element violates the latent relation shared by the others.

A recurring implication is that odd-one-out detection is especially suited to settings where anomalies are subtle, relative, or previously unspecified. This suggests why the formulation appears in unsupervised OOD detection, open-world recognition, graph-level detection, dynamic networks, and sequence analysis rather than only in classical pointwise outlier detection (Rafiee et al., 2020, Elliott et al., 2019, Kandanaarachchi et al., 2022).

2. Formalizations and scoring principles

Several mathematically distinct scoring rules instantiate the odd-one-out principle.

In semantic-similarity-based detection, SemSAD first finds, for any test example, the semantically closest examples in the training set, where semantic relation is quantified by cosine similarity between feature vectors that leave semantics unchanged under transformations such as geometric transformations, time shifts, and synonymous word substitutions (Rafiee et al., 2020). The cosine similarity is

$$\text{cos\_sim}(h(x_i), h(x_j)) = \frac{h(x_i) \cdot h(x_j)}{\|h(x_i)\| \, \|h(x_j)\|}$$

and the anomaly score may be written as

$$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$

Lower scores indicate lower similarity to the in-distribution set, and a test example is classified as OOD if the semantic similarity to its nearest neighbours is significantly lower than the corresponding similarity for test examples from the in-distribution (Rafiee et al., 2020).

In the perception-based a-contrario model, the expected number of windows of count $n$ under the random model is

$$\mathbb{E}(C_n) = {S \choose n} \times \frac{1}{W^{n-1}}$$

and the key anomaly criterion is to flag an observation as anomalous if $\mathbb{E}(C_n) < 1$ (Mohammad, 2021). After log transformation, the practical criterion becomes

$$-\frac{1}{S}\left( \log {S \choose n} - (n-1) \log W \right) > 0.$$

This formalization makes “unexpectedness” itself the anomaly score.

In language-model-based sequence detection, oddballness is defined for a probability distribution $D = \{p_1,\ldots,p_n\}$ by

$$E_D(p_i) = \frac{\sum_j (p_j - p_i)_+}{\sum_j p_j},$$

which simplifies to

$E_D(p_i) = \sum_j (p_j - p_i)_+$

for probability distributions (Graliński et al., 2024). The measure satisfies the axioms listed in the paper, including range, impossibility, most likely outcome, equal probability tokens, monotonicity, and continuity. Its complement is the “probability of a probability,” $T_D(p_i) = 1 - E_D(p_i)$ (Graliński et al., 2024). The essential distinction from likelihood thresholding is that oddballness depends on the shape of the predicted distribution, not only on the observed token probability.

In open-world detection, O1O uses superclass-conditioned recalibration

$$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$0

or, equivalently,

$$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$1

and defines unknownness as

$$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$2

If a query does not fit well into any known class or superclass, $$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$3 is high and the region is labeled as unknown or odd-one-out (Yavuz et al., 2024).

In graph-level OOD and anomaly detection, the unified benchmark reduces both GLAD and GLOD to learning a scoring function $$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$4 such that for a test graph $$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$5,

$$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$6

This abstracts odd-one-out detection as generalized graph-level OOD detection (Wang et al., 2024). In dynamic networks, oddnet instead models a sequence of graphs via features and ARIMA residuals, beginning from

$$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$7

mapping each graph to $$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$8, and detecting anomalies through large residual norms after temporal modeling (Kandanaarachchi et al., 2022).

These formalizations differ in representation space, but they share a structural premise: the odd element is not defined only by absolute rarity; it is defined by failure to align with a learned or inferred relational structure.

3. Methodological families

The major methodological families can be organized by how the reference structure is constructed.

Family	Core mechanism	Representative papers
Similarity and neighbor comparison	Compare an instance to semantically closest training examples or neighboring objects	(Rafiee et al., 2020, Bhunia et al., 2024)
Grouping and expectation	Define anomalies through dominant groupings and a-contrario expectation	(Mohammad, 2021)
Generative and latent-space one-class models	Learn a normal manifold and score deviations by reconstruction or latent compactness	(Chen et al., 2020, Dehghanian et al., 2023)
Sequence and language-model methods	Score anomalous tokens or rules in sequences	(Graliński et al., 2024, Gan et al., 2021)
Graph and network methods	Detect odd graphs, anomalous nodes, or anomalous time points by graph features or unified OOD scores	(Wang et al., 2024, Elliott et al., 2019, Kandanaarachchi et al., 2022)
Self-supervised oddity tasks	Train representations by spotting the odd element in a constructed set	(Fernando et al., 2016)

Similarity-based methods construct the reference set explicitly. SemSAD uses a feature extractor and nearest-neighbour semantic similarity, then a discriminator to classify a test example as OOD when similarity to the nearest neighbours is sufficiently low (Rafiee et al., 2020). The multi-view scene method “Odd-One-Out: Anomaly Detection by Comparing with Neighbors” constructs 3D object-centric models from 2D views, enhances them with geometrically consistent part-aware representations, extracts object-centric crops, and performs anomaly detection through cross-instance comparison with sparse voxel attention (Bhunia et al., 2024). A plausible implication is that these methods treat oddness as a failure of local relational consistency rather than a failure of global density estimation.

Grouping-based methods construct the reference by majority structure. The perception-inspired algorithm assumes a dominant Gestalt law and detects the observation that is unexpected with respect to that grouping, with no parameter tuning and no contamination-ratio assumption (Mohammad, 2021). O1O similarly creates structure by grouping known classes into superclasses so that unknown objects can be identified through their failure to fit any known group (Yavuz et al., 2024).

Generative one-class methods instead learn a compact normal manifold. The One Class Latent Regularized Network uses an adversarial dual autoencoder architecture with a one-class latent regularizer that concentrates normal samples around a latent center $$S(x) = \frac{1}{N} \sum_{j=1}^{N} \text{cos\_sim}(h(x), h(x_j)).$$9, while the auxiliary autoencoder improves training stability (Chen et al., 2020). RCALAD extends cycle-consistent GAN-based anomaly detection through complete cycle consistency, a novel discriminator $n$0, a supplementary distribution $n$1 in input space, and anomaly scores $n$2 and $n$3 that compare real and reconstructed cycles (Dehghanian et al., 2023). In these models, the odd sample is the one that cannot be reconstructed or embedded as a typical normal sample.

Sequence-specific methods separate token-level and pattern-level oddity. Oddballness uses language-model distributions directly for token-level anomaly detection (Graliński et al., 2024). DUOS addresses sequence data through utility-aware outlier sequential rules, incorporating anomalousness and utility into UOSR mining and defining outlier factors over rare high-utility sequential rules (Gan et al., 2021). This suggests two different sequence notions of oddness: local surprisal relative to a predictive distribution, and rare high-utility rule structure relative to the sequence database.

Graph and network methods generalize the same logic to relational data. UB-GOLD unifies graph-level anomaly detection and graph-level OOD detection under generalized graph-level OOD detection (Wang et al., 2024). The financial transaction network method computes 140 node-centric features, including network-comparison, spectral, local-statistics, community, and heavy-path features, then aggregates them via feature sum or random forest to detect previously unspecified anomalies (Elliott et al., 2019). Oddnet models dynamic networks through network features, ARIMA residuals, robust PCA, and the lookout algorithm, thereby making the odd element a time point whose network is unusual relative to its temporal history (Kandanaarachchi et al., 2022).

4. Evaluation settings and empirical patterns

Empirical work on odd-one-out detection is unusually heterogeneous because the task varies by modality and by whether the oddity is semantic, structural, temporal, or scene-relative.

In visual OOD detection, SemSAD reports that it can outperform previous approaches for anomaly, novelty, or out-of-distribution detection in the visual domain by a large margin, and in particular obtains AUROC values close to one for the task of detecting examples from CIFAR-10 as out-of-distribution given CIFAR-100 as in-distribution, without making use of label information (Rafiee et al., 2020). The One Class Latent Regularized Network reports state-of-the-art results on MNIST and CIFAR10 as well as the GTSRB stop signs dataset, with the full model reaching 0.99 on MNIST, 0.97-0.98 on CIFAR-10, and 0.96 on GTSRB in the ablation summary given in the data block (Chen et al., 2020). RCALAD is evaluated on KDDCup99, Arrhythmia, Thyroid, Musk, CIFAR-10, and SVHN, and is reported to set new state-of-the-art F1-scores on several tabular datasets while achieving the highest mean AUROC scores on both CIFAR-10 and SVHN (Dehghanian et al., 2023).

In scene-centric visual odd-one-out detection, the original multi-view benchmark introduces ToysAD-8K and PartsAD-15K. On ToysAD-8K and PartsAD-15K, the method reports 91.78 / 83.21, 89.15 / 81.57, and 86.12 / 79.68 for AUC / Accuracy on seen, unseen, and parts settings respectively, outperforming DETR3D, ImVoxelNet, and COLMAP+PC in the table reproduced in the data block (Bhunia et al., 2024). The later DINO-based efficiency-oriented model reduces the number of parameters by one third and shortens training time by a factor of three compared to the current state-of-the-art, while remaining competitive on Toys and improving substantially on Parts, where it reports 89.72 / 88.81 (Chito et al., 4 Sep 2025). The same work also reports a Multimodal LLM baseline with accuracies 52.2–60.7%, describing current limitations in structured visual reasoning tasks (Chito et al., 4 Sep 2025).

In open-world object detection, O1O evaluates on M-OWOD and S-OWOD and reports significant improvements in unknown recall without compromising known performance. The example given in the data block is M-OWOD Task 1, where O1O achieves 49.3% U-Recall against previous bests of 28.3% and 36% (Yavuz et al., 2024).

In text anomaly detection, oddballness is evaluated on grammatical error detection. On FCE, oddballness improves over raw probability across the reported models, for example from 38.86 to 40.52 Test F0.5 for GPT2-XL, and the max-combination of GPT2-XL and RoBERTa Large reaches 43.15 Test F0.5 (Graliński et al., 2024). Across MultiGED-2023, the paper states that oddballness consistently outperforms the probability and top-K methods with substantial margins (Graliński et al., 2024).

Graph results emphasize variability and non-universality. UB-GOLD covers 35 datasets spanning four practical anomaly and OOD detection scenarios and compares representative GLAD/GLOD methods under AUROC, AUPRC, and FPR95 (Wang et al., 2024). The benchmark reports average AUROC values of 75.86 for SIGNET, 73.14 for OCGTL, 71.05 for GOOD-D, and 71.07 for CVTGAD, with all 2-step methods below 62 (Wang et al., 2024). At the same time, the benchmark states that there is no “universal” best method, that far-OOD samples are much easier than near-OOD samples, and that most methods are sensitive to OOD or anomalous contamination in the training set (Wang et al., 2024).

Dynamic-network and network-structure evaluations likewise stress operational rather than merely aggregate performance. The financial transaction network method reports that in all cases, the top 2 percent of flagged anomalies contained on average over 90 percent of the planted anomalies (Elliott et al., 2019). Oddnet is reported to outperform LAD and Tensorsplat particularly in dynamic settings where the baseline changes over time (Kandanaarachchi et al., 2022). In saliency research, by contrast, “Do Saliency Models Detect Odd-One-Out Targets?” concludes that nearly all saliency algorithms do not adequately respond to singleton targets in synthetic and natural images, and that additional training data does not lead to a significant improvement in the ability of current CNN-based saliency models to find odd-one-out targets (Kotseruba et al., 2020). This negative result is notable because it shows that strong fixation prediction does not imply competence on canonical pop-out oddity detection.

5. Relations to anomaly detection, OOD detection, novelty detection, and open-world recognition

The odd-one-out perspective cuts across several research domains that are often studied separately. The unified survey states that anomaly detection, novelty detection, one-class learning, open set recognition, and out-of-distribution detection share similar and shared concepts but have often been investigated independently, creating research barriers (Salehi et al., 2021). Within that taxonomy, odd-one-out anomaly detection is not always a formal separate term; instead, it is frequently instantiated as a self-supervised or relational mechanism for highlighting samples that stand out relative to prevailing data characteristics or transformations (Salehi et al., 2021).

This cross-domain role is visible in the benchmark literature. UB-GOLD explicitly argues that unsupervised graph-level anomaly detection and unsupervised graph-level OOD detection indeed share the same objective and should be unified under generalized graph-level OOD detection (Wang et al., 2024). A similar unification appears implicitly in SemSAD, whose abstract treats classifying samples as in-distribution or out-of-distribution as a challenging problem of anomaly detection and a strong test of generalisation power (Rafiee et al., 2020). O1O frames unknown-object detection in open-world settings as an odd-one-out problem over grouped known classes (Yavuz et al., 2024). These examples show that odd-one-out reasoning often converts open-world uncertainty into a structured comparison problem.

At the same time, odd-one-out detection is not reducible to any single neighboring field. In one-class generative models, the reference is a learned normal manifold rather than a co-occurring peer group (Chen et al., 2020, Dehghanian et al., 2023). In dynamic networks, the reference is temporal history rather than contemporaneous neighbors (Kandanaarachchi et al., 2022). In data-depth methods, the reference is the center-outward ordering of the multivariate data cloud, and lower depth values correspond to higher abnormality (Mozharovskyi et al., 2022). This suggests that “odd-one-out” is best understood as a relational criterion that can be realized through peers, prototypes, groups, or histories.

6. Limitations, misconceptions, and future directions

A common misconception is that odd-one-out detection is equivalent to selecting the lowest-likelihood or lowest-confidence sample. The oddballness formulation explicitly argues against this simplification: some low-probability tokens are not anomalous, while some mid-probability tokens can be truly out of place depending on the distribution’s structure (Graliński et al., 2024). SemSAD similarly emphasizes semantic similarity rather than model confidence or reconstruction fidelity (Rafiee et al., 2020). In visual saliency, the failure of many models on singleton targets further indicates that generic saliency or fixation prediction is not the same as odd-one-out competence (Kotseruba et al., 2020).

Another misconception is that a universal odd-one-out detector already exists. UB-GOLD states that no “universal” graph anomaly/OOD detector exists, that methods can have “cold spots,” and that near-OOD and contaminated training sets remain challenging (Wang et al., 2024). The DINO-based efficient visual odd-one-out model reaches competitive performance while reducing parameters and training time, but the accompanying MLLM baseline still performs poorly, underscoring limitations in structured visual reasoning (Chito et al., 4 Sep 2025). In open-world detection, O1O improves unknown recall without compromising known performance, but its mechanism depends on better partitioning of the feature space with superclasses and on pseudo-label quality from geometric cues (Yavuz et al., 2024).

Across the literature, several recurrent open problems are explicitly identified. These include robustness to subtle and diverse OOD or anomaly patterns, mechanisms for handling contaminated or noisy training sets, and structure-aware solutions that do not overfit to specific attributes or domains in graph data (Wang et al., 2024). The saliency results argue that architectural changes may be required to capture grouping, feature binding, and object individuation for odd-one-out targets (Kotseruba et al., 2020). The survey highlights the need for better self-supervised oddity tasks, explainability, robustness, and generalization to real unseen anomalies or distribution shifts (Salehi et al., 2021). The efficient multi-view visual work likewise points to future methods that further leverage large pretrained models while enhancing context-aware and efficient relational modules (Chito et al., 4 Sep 2025).

A plausible overall implication is that odd-one-out anomaly detection is most powerful when anomaly is inherently relational: an object differs from its peers, a token is strange relative to a predictive distribution, a graph is out-of-family relative to normal graphs, or a time point violates its own temporal trajectory. The strongest current methods therefore tend to combine expressive representations with explicit comparison, grouping, or normality modeling rather than relying on a single absolute score.