Generalized Contextual Group Models

• Generalized contextual groups are mechanisms that encode structured, shared context between isolated observations and global models across various domains.
• They organize context-dependent information via approaches like abstract group profiling, expert weighting, graph propagation, coset geometries, and holonomy in RBMs.
• Applications in content customization, contextual bandits, quantum contextuality, and group-valued Boltzmann machines demonstrate practical trade-offs between local noise and global consistency.

Searching arXiv for the cited works and related uses of the term. Tool call: arxiv_search(query="Generalized Contextual Group OR contextual group", max_results=10) Searching arXiv for exact identifiers: (Dehghani et al., 2016, Qi et al., 2022, Li, 2013, Planat, 2014, Magnot, 5 Sep 2025). “Generalized contextual group” is used in the cited literature for several formally distinct constructions that organize context-dependent structure at a level intermediate between isolated observations and fully global models. In information retrieval, it denotes an abstract group profile extracted from user preference models by separating ubiquitous, individual-specific, and genuinely shared terms; in contextual bandits, it appears as a generalized expert family with prior-weighted updates or as an Arm Group Graph that couples arm groups through learned correlations; in quantum foundations, it denotes coset- and stabilizer-based structures whose failure to align with operator commutation witnesses contextuality; and in group-valued Boltzmann machines, it is defined explicitly as a triple $(G,\phi,\rho)$ whose cycle holonomies yield a contextuality index (Dehghani et al., 2016, Qi et al., 2022, Li, 2013, Planat, 2014, Magnot, 5 Sep 2025).

1. Semantic scope and recurrent structure

In the cited works, the expression does not name a single canonical object. Instead, it refers to a family of constructions in which context is mediated by a group-level representation, and inference depends on separating local variation from structured shared regularities.

Domain	Formal object	Principal mechanism
Content customization	Abstract group LLM $\theta_g$	EM decomposition against background $\theta_c$ and specific $\theta_s$
Contextual bandits	Expert mixture or Arm Group Graph	Multiplicative loss updates, GNN propagation, and UCB exploration
Quantum contextuality	Coset-permutation-stabilizer geometry	Comparison of coset commutators with operator commutators
Group-valued RBMs	Triple $(G,\phi,\rho)$ and cycle holonomies	Gauge-invariant contextuality index from cycle deviations

A recurring pattern across these formulations is the treatment of context as structured rather than merely local. In one case, context is demographic or situational metadata used to form user groups. In another, it is the round-wise arm set and its group graph. In the quantum setting, contexts are commuting sets of observables represented as lines of an incidence geometry. In the RBM setting, contexts become cycles in a discrete bundle, with holonomy measuring global inconsistency. This suggests a common abstraction: a generalized contextual group is a mechanism for encoding which local relations should be treated as coherent, shared, or jointly realizable.

2. Abstract group profiles for content customization

In “Generalized Group Profiling for Content Customization,” the central object is an “abstract” group profile: a latent LLM $\theta_g$ that captures all, and only, the essential features shared by members of a group, while excluding both general/background terms and individual accidental features (Dehghani et al., 2016). The formal setting uses a vocabulary $V$, user preference models $\theta_u$ estimated from liked items, a collection model $\theta_c$ over the corpus, and a specific model $\theta_s$ designed to assign high probability to terms important to exactly one user in a group and not the others. For user $\theta_g$0,

$\theta_g$1

The specificity score is defined by

$\theta_g$2

followed by normalization to obtain $\theta_g$3.

The observed term distribution for each user is modeled as a three-component mixture,

$\theta_g$4

with user-specific mixture weights $\theta_g$5 for $\theta_g$6. Maximum-likelihood estimation is performed with EM. The E-step computes posterior responsibilities

$\theta_g$7

and the M-step updates $\theta_g$8 and the mixture weights. Because $\theta_g$9 and $\theta_c$0 are fixed, EM “implicitly eliminates” general and specific features by assigning their probability mass to those components, leaving $\theta_c$1 to represent the abstract group-level signal.

The paper applies this model to TREC 2015 Contextual Suggestion, Batch task, with 207 users, ratings from $\theta_c$2 to $\theta_c$3, and relevant items defined by ratings $\theta_c$4. Candidate items are scored by cross-entropy against either the group model or a smoothed user model. Group-only ranking uses

$\theta_c$5

while the combined model replaces $\theta_c$6 by

$\theta_c$7

Empirically, Mean Average Precision varies strongly with grouping criterion. Group-only MAP is $\theta_c$8 for trip duration, $\theta_c$9 for age, $\theta_s$0 for group type, $\theta_s$1 for season, $\theta_s$2 for gender, and $\theta_s$3 for trip type. The preferences-only personalization baseline is $\theta_s$4. When group profiles are combined with user preferences via the learned mixture weights, MAP rises to $\theta_s$5 for age, $\theta_s$6 for gender, $\theta_s$7 for group type, $\theta_s$8 for trip type, $\theta_s$9 for trip duration, and $(G,\phi,\rho)$0 for season; these improvements over both preferences-only and group-only models are statistically significant by one-tailed $(G,\phi,\rho)$1-test with $(G,\phi,\rho)$2. The granularity analysis on age bins exposes a trade-off between noisy estimation in very small groups and signal dilution in very large groups: $(G,\phi,\rho)$3-year bins yield MAP $(G,\phi,\rho)$4, $(G,\phi,\rho)$5-year bins $(G,\phi,\rho)$6, $(G,\phi,\rho)$7-year bins $(G,\phi,\rho)$8, and $(G,\phi,\rho)$9-year bins $\theta_g$0.

3. Contextual bandits: expert weighting and arm-group graphs

In contextual bandits, one formulation aligned with the generalized contextual group idea is Generalized Thompson Sampling, an expert-learning framework in which a finite set of experts $\theta_g$1 predicts mean rewards $\theta_g$2 and induces greedy policies $\theta_g$3 (Li, 2013). Expert weights are updated multiplicatively using a loss $\theta_g$4:

$\theta_g$5

Action selection mixes expert recommendations with uniform exploration,

$\theta_g$6

With logarithmic loss, $\theta_g$7, and $\theta_g$8, the update coincides with Bayes posterior updating under a Bernoulli likelihood, so the framework reduces to classical Thompson Sampling. The regret analysis is prior-sensitive. Under the stated consistency, informativeness, boundedness, and self-boundedness conditions,

$\theta_g$9

and Bayes regret depends on $V$0. For square loss with $V$1, the bound becomes $V$2; for logarithmic loss it becomes $V$3.

A second, explicitly group-centric formulation is the Arm Group Graph in “Neural Bandit with Arm Group Graph,” where arms are organized into groups and the groups themselves are correlated (Qi et al., 2022). At round $V$4, the learner observes a subset of groups $V$5 and candidate arms $V$6. The generalized contextual group structure is represented by an undirected, fully connected graph $V$7 whose nodes are groups and whose weighted edges are estimated online from kernel mean embeddings of past contexts. With adjacency $V$8 and degree matrix $V$9, the normalized propagation matrix is

$\theta_u$0

Candidate arms are embedded into a group-aware matrix $\theta_u$1, passed through a GNN,

$\theta_u$2

and concatenated with the raw embedding to mitigate over-smoothing:

$\theta_u$3

A fully connected network then produces per-group reward predictions.

Exploration uses a gradient-based upper confidence bound. For candidate arm $\theta_u$4,

$\theta_u$5

with

$\theta_u$6

The analysis decomposes the confidence error into parameter-estimation error $\theta_u$7 and graph-estimation error $\theta_u$8, with $\theta_u$9. Under over-parameterization, AGG-UCB achieves a near-optimal regret bound matching state-of-the-art neural bandit forms up to constants and the graph-estimation term scaling with $\theta_c$0. Empirically, on MovieLens 20M, Yelp, MNIST-Aug, and XRMB, AGG-UCB attains the lowest cumulative regret curves among KMTL-UCB, Kernel-UCB variants, Neural Thompson Sampling, and Neural-UCB baselines. Group-aware neural baselines outperform pooled counterparts, and $\theta_c$1 hop gives the best performance on MovieLens and MNIST-Aug.

Taken together, these two bandit formulations show two distinct ways of instantiating a contextual group. One uses a weighted population of experts and makes prior mass explicit in the regret. The other treats groups as graph nodes and propagates information along learned inter-group correlations.

4. Coset geometries and quantum contextuality

In “Geometry of contextuality from Grothendieck’s coset space,” generalized contextual group structure is realized through subgroups $\theta_c$2 of the two-generator free group, right cosets of $\theta_c$3, and the induced permutation group $\theta_c$4 acting on the coset space (Planat, 2014). The paper works with $\theta_c$5 and, in the constructions discussed there, restricts to $\theta_c$6. A subgroup of index $\theta_c$7 yields $\theta_c$8 right cosets $\theta_c$9, and the action of $\theta_s$0 and $\theta_s$1 on cosets defines a dessin d’enfant encoded by $\theta_s$2 with $\theta_s$3. Points of a finite point-line geometry are labeled by cosets, while lines are obtained from orbits of two-point stabilizer subgroups.

The contextuality criterion is formulated by comparing group-theoretic commutation with operator commutation. The group commutator is

$\theta_s$4

while the observable commutator is

$\theta_s$5

A geometry is contextual whenever there does not exist a dessin for which the group commutator precisely corresponds to the operator commutator on all lines of the geometry. Equivalently, if a line of mutually commuting observables cannot be labeled by cosets satisfying the required commutation law, the geometry is contextual in this framework.

The small-index examples illustrate both non-contextual and contextual behavior. At index $\theta_s$6, two explicit permutation representations stabilize the octahedron: one has contextual triangles, while the other is non-contextual. At index $\theta_s$7, a non-contextual dessin stabilizes the Fano plane. Contextuality first appears, in the sense emphasized by the paper, at index $\theta_s$8 in Mermin’s square and at index $\theta_s$9 in Mermin’s pentagram. In Mermin’s square, the contextual embedding has a defective right-hand column, paralleling the standard parity proof with commuting two-qubit Pauli observables. The geometrical contextuality measure is

$\theta_g$00

where $\theta_g$01 is the number of lines and $\theta_g$02 the number of lines whose cosets commute. For Mermin’s square, $\theta_g$03, $\theta_g$04, so $\theta_g$05. In the four-qubit geometry $\theta_g$06, only $\theta_g$07 out of $\theta_g$08 lines have commuting coset labels, giving $\theta_g$09. For generalized polygons the reported values are even larger, including $\theta_g$10 for $\theta_g$11, $\theta_g$12 for $\theta_g$13, $\theta_g$14 for $\theta_g$15, and $\theta_g$16 for the dual of $\theta_g$17.

Within this framework, the “group” organizing contextuality is not only the ambient free-group quotient but also the permutation group of the dessin and its two-point stabilizers. The cited constructions therefore treat contextuality as a structural failure of these groups to enforce abelian behavior on every line-orbit of the stabilized geometry.

5. Group-valued Boltzmann machines, holonomy, and discrete bundles

In “Contextuality, Holonomy and Discrete Fiber Bundles in Group-Valued Boltzmann Machines,” the notion is formalized directly: a generalized contextual group is a triple $\theta_g$18, where $\theta_g$19 is a possibly non-abelian group carried by RBM edges, $\theta_g$20 is a conjugation-invariant deviation with $\theta_g$21, and $\theta_g$22 is an optional representation acting on node feature spaces (Magnot, 5 Sep 2025). The RBM graph is bipartite with visible nodes $\theta_g$23, hidden nodes $\theta_g$24, and edge weights $\theta_g$25.

Two modeling choices are given. In the representation-based version, with $\theta_g$26 and vector states $\theta_g$27, the energy is

$\theta_g$28

In the coherence-based version, a section $\theta_g$29 is penalized through

$\theta_g$30

with typical choices such as $\theta_g$31 or $\theta_g$32 near the identity.

The key geometric object is holonomy. For an oriented cycle

$\theta_g$33

the holonomy is

$\theta_g$34

Gauge transformations attach $\theta_g$35 to each node and transform edges by

$\theta_g$36

so the holonomy transforms by conjugation at the basepoint:

$\theta_g$37

A cycle is flat if $\theta_g$38. Given a family of cycles $\theta_g$39, the contextuality index is

$\theta_g$40

or its normalized variant. In the near-identity regime, $\theta_g$41, so the index acts as an aggregate curvature magnitude.

The paper states a discrete flatness/coherence proposition: there exists a global section $\theta_g$42 such that

$\theta_g$43

if and only if $\theta_g$44 for every cycle $\theta_g$45 in the graph. This gives a direct equivalence between trivial holonomy and global trivializability, aligning the framework with sheaf-theoretic “no global section $\theta_g$46 contextuality” language.

Training can include curvature regularization:

$\theta_g$47

For matrix groups, the paper derives cycle-wise gradients by factorizing the holonomy as $\theta_g$48. With $\theta_g$49, the ambient gradient is

$\theta_g$50

and for $\theta_g$51 it is projected to the tangent space via a skew-Hermitian component before the update $\theta_g$52.

The examples make the construction concrete. For a $\theta_g$53 $\theta_g$54-cycle, the reported holonomy is approximately

$\theta_g$55

with Frobenius deviation $\theta_g$56. For an $\theta_g$57 triangle built from Pauli rotations with angles $\theta_g$58, the gauge-invariant scalar phase observable is numerically $\theta_g$59 rad. In an infinite-dimensional $\theta_g$60 example, $\theta_g$61 yields a typical value $\theta_g$62 for $\theta_g$63.

6. Cross-cutting issues, limitations, and directions

Across these formulations, the generalized contextual group serves as a device for isolating structure that should persist across contexts while suppressing contamination from uninformative or inconsistent local variation. In group profiling, this appears as the separation of general terms, accidental features, and shared group essence. In bandits, it appears as either a prior-weighted ensemble of experts or a graph of correlated arm groups. In the quantum and RBM settings, it appears as a distinction between globally consistent and holonomy- or commutator-obstructed local assignments. This suggests that the common mathematical theme is not a single algebraic definition but a family of context-sensitive coherence constraints.

The limitations are correspondingly heterogeneous. The content-customization model assumes a three-component mixture and fixes $\theta_g$64 and $\theta_g$65 heuristically rather than learning them jointly; computing $\theta_g$66 naively costs $\theta_g$67, and the static $\theta_g$68 may need online updating as groups or contexts change. It also notes that grouping criteria such as demographics and travel context may introduce or reinforce biases, and that group granularity must balance under-customization against noisy estimation in small groups (Dehghani et al., 2016). In generalized Thompson sampling, the analysis assumes realizability and bounded or self-bounded losses, and the resulting regret guarantees scale as $\theta_g$69 rather than the optimal problem-independent $\theta_g$70 rate (Li, 2013). In AGG-UCB, performance depends on accessible and stable group labels, stationarity of group context distributions for kernel mean embeddings, over-parameterization and small-step assumptions, and the cost of maintaining dense confidence matrices $\theta_g$71 (Qi et al., 2022). In group-valued RBMs, the main issues are nonconvexity, gauge identifiability, and numerical instability of logarithms for noncompact or non-normal matrices (Magnot, 5 Sep 2025).

The proposed extensions remain domain-specific but structurally related. For content customization, the cited directions include jointly learning richer specific components, combining multiple grouping criteria in hierarchical or multi-view models, and extending beyond text to images, check-ins, and multi-modal data. For bandits, the proposed directions include dynamic group discovery, time-varying Arm Group Graphs, Bayesian variants, alternative UCB constructions, sparse AGG, and low-rank approximations to $\theta_g$72. For group-valued RBMs, the framework opens curvature-aware regularization and topological regularization. A plausible implication is that future uses of the term will continue to revolve around the same design problem: how to represent context-dependent regularity at a group level without collapsing either into purely individual noise or into overly coarse global averaging.