Granular Context Dependency Taxonomy

• Granular context dependency taxonomy is a structured framework that categorizes multi-level contextual features to condition predictions, task complexity, and system adaptation.
• It employs formal structures, discretization, and learned embeddings to rigorously quantify context influence, achieving significant accuracy gains and improved performance benchmarks.
• It informs practical designs in AR, NLP, AGI risk, and software refactoring by enabling adaptive system design, theory-driven classification, and dynamic benchmarking.

Granular context dependency taxonomy refers to a set of rigorous, multi-level frameworks that characterize, model, and operationalize how contextual information conditions or structures entities, predictions, task complexity, system classification, or adaptation in a domain. Such taxonomies are pivotal for systems that must reason about fine-grained state, multi-modal signals, or multi-factor environmental factors, and are foundational in research areas including entity type inference, context-aware adaptation (especially in AR and mobile systems), code dependency refactoring, long-context NLP, AGI risk stratification, and intelligent scientific taxonomy construction. Below, diverse forms and instantiations of granular context dependency taxonomy are synthesized, with an emphasis on their formal structure, mathematical modeling, measurement methodology, design principles, and evaluation protocols.

1. Formal Structures and Contextual Feature Spaces

Granular context dependency taxonomies frequently begin by partitioning the contextual state space into explicit, multi-level, or multi-modal feature hierarchies, often with strict tree structures or compositional context vectors.

• In fine-grained entity type tagging (Gillick et al., 2014), entity types are organized into a strict tree $\mathcal{T}$ rooted at “ALL”, split at level-1 into PERSON, LOCATION, ORGANIZATION, and OTHER, with nested subtypes at levels 2–3. Assigning a fine type $t$ to a mention implies all ancestors via IS-A.
• Intelligent AR systems (Davari et al., 2024) define context as $C = [C_u; C^s; C^{su}]$, with $C_u$ (user profile and state), $C^s$ (setting: real/digital/social, split further into transient/persistent), and $C^{su}$ (user-setting interplay). Each “slot” can be a real-valued, categorical, or structured field.
• Mobile usage context taxonomies (Rahmati et al., 2012) split context sources into sensor (temporal, motion, spatial) and usage (prior usage events) classes, carefully cataloged and mapped to measurable features with associated cost and accuracy.
• In AGI ontologies (Max, 6 Oct 2025), the taxonomy is structured as six normalized technical-institutional axes: actor structure ($x_1$), psychological distance ($x_2$), governance ($x_3$), framing ($x_4$), architecture ($x_5$), and development tempo ($x_6$), each $x_i \in [0,1]$, supporting continuous or thresholded regime classification.
• For long-context NLP, the “Diffusion–Scope” grid (Goldman et al., 2024) provides a two-dimensional formalization indexing tasks by the spread ($D$) and volume ($S$) of required information, with potential subaxes for temporal, referential, hierarchical, and multimodal factors.
• LLM-guided hierarchical taxonomy generation (Zhu et al., 23 Sep 2025) encodes each node as a set of context-conditioned multi-aspect vectors, dynamically constructed via LLMs conditioned on partial taxonomy paths, resulting in adaptive facet creation and aspect-specific document embeddings.

2. Methodological Principles for Granular Context Capture

Granularity in context dependency is achieved by direct quantification, discretization, and/or learned embeddings of contextual features, with systematic protocols for measurement and representation.

• Discretization and binning (equal-width, equal-frequency, k-means, supervised by posteriors) are used to overcome data sparsity and support context quantization (Rahmati et al., 2012). Supervised binning by P(g|bin) clusters increases predictive granularity by up to 15% accuracy over simple bins.
• Composite context vectors are constructed by concatenating feature groups, with empirical feature selection guided by both predictive value and resource/energy cost (as in the SmartContext submodular optimizer (Rahmati et al., 2012)).
• Context-aware hierarchical taxonomy (Zhu et al., 23 Sep 2025) conditions every split (aspect selection, summary embedding, clustering) on the cumulative ancestor path—thus each embedding is context-specifically refined.
• Contextual dependencies may be made explicit as functions (e.g., $$SU.occluded\_entities = f_{occl}(C_u.state.head\_pose, C.real.immediate.depth\_map)$$ (Davari et al., 2024)), or via empirical learned weights in a context-feature scoring or classification function (e.g., $S(x_1,\dots,x_6) = \sum_{i=1}^6 w_i x_i$ (Max, 6 Oct 2025)).

3. Taxonomy Instantiation in Domains

Software Dependency Cycles

• The granular context dependency taxonomy for two-class dependency cycles (Feng et al., 2023) identifies five untangling/refactoring patterns (Remove Unused Code, Move Between Classes, Move to Third Class, Shorten Call Chain, Leverage Built-in Feature), governed by a 76-dimensional vector parameterizing internal structure (24 binary features) and external neighbor context (52 integer features).
• Three “cycle shift” anti-patterns (to parent, to child, to unrelated third) are recognized as counterintuitive but empirically prevalent, underscoring the multidimensional context determining cycle refactoring.

Patterns, Context Features, and Triggers in Dependency Refactoring

Pattern	Internal Traits	Neighbor Context
Remove Unused/Deprecated Code	Calls, imports, uses but unused	None
Move Between Two Classes	Cohesive calls/use	Minimal
Move to Third Class	Shared utility calls/extends	Existing mediator
Shorten Call Chain	Trivial call delegation	Third entity visible
Leverage Built-In Feature	Framework inheritance	None, use core API

Mobile and AR Systems

• Context sources are categorized and evaluated for predictive power and energy cost; context dependencies are often handled via classifier combination (e.g., Naïve Bayes, Max, Mean) and Laplace smoothing (Rahmati et al., 2012).
• Dynamic adaptation in AR relies on explicit context taxonomies. Features drive learned or rule-based inferences for interface adaptation, with optimization over adaptation conflicts (Davari et al., 2024).

NLP and Scientific Literature

• Fine-grained entity typing in NLP utilizes context-sensitive tree-based taxonomies, with local context (sentence/document) restricting label admissibility (Gillick et al., 2014).
• Long-context NLP tasks are classified along orthogonal axes (Diffusion and Scope), with a formal task grid identifying setting difficulty and required annotation protocols (Goldman et al., 2024).
• Context-aware taxonomy generation for scientific corpora employs multi-aspect, hierarchical splits, with dynamic clustering informed by LLM-generated aspect lists and context-specific facet generation at each branch (Zhu et al., 23 Sep 2025).

4. Mathematical Formalization and Algorithmic Principles

Taxonomies often adopt explicit formal models for context representation and dependency measurement:

• In context-aware mobile prediction, the estimation accuracy under MAP is defined as $Acc_1 = E_x \left[ \max_i P(g = g_i | x) \right]$, with generalization to $Acc_m$ for $m$-guess return sets (Rahmati et al., 2012).
• The AR context taxonomy encodes the entire state as a sparse or dense record, $C \in \mathbb{R}^n$, enabling both rule-based and ML inference for adaptation optimization, e.g.,

$$A^* = \operatorname{arg\,max}_{A\subseteq\text{possible adaptations}} \sum_{i\in A} \text{impact}_i - \lambda \cdot \text{conflict penalty}(A)$$

(Davari et al., 2024).

• AGI regime classification is mapped as $f(\mathbf{x}) = \text{High-AGI}$ if $S(\mathbf{x}) \ge \tau$, $S(\mathbf{x}) = \sum_i w_i x_i$ for normalized feature axes; dependencies among axes are modeled via directional influence (e.g., $x_1 \rightarrow x_3, x_4$) and political risk variables (Max, 6 Oct 2025).
• Long-context difficulty is modeled as
  • Scope: $S = |R|$ or $S = \sum_{r\in R} \ell_r$
  • Diffusion: $D = \frac{1}{|R|-1} \sum_{i < j} |i - j|$, $D_{\text{norm}} = D/N$
  • (Goldman et al., 2024).
• In LLM-based taxonomy generation, paper embeddings are contextually generated along $m$ node-specific aspects, with clustering (GMM, EM objective), and assignment maximized over aspect–cluster tuples. At each node, aspect generation and summary embedding are conditioned on the current path and corpus fragment (Zhu et al., 23 Sep 2025).

5. Empirical Taxonomies and Performance Benchmarks

A critical characteristic is the operationalization of taxonomic granularity via benchmarking and empirical quantification.

• In mobile usage, empirical context-dependency is ranked by predictive gain: GPS and cell-ID drive high gain (>+20%), time/day and accelerometer medium, prior usage typically low, and context combination via Naive Bayes yields up to +42% accuracy for app launches (Rahmati et al., 2012).
• Scientific taxonomy generation with LLMs is evaluated on TaxoBench-CS, with 156 computer science trees from arXiv survey papers (mean tree depth 3.1, mean internal nodes 24.8) (Zhu et al., 23 Sep 2025). Metrics include NMI, ARI, Purity, CEDS, HSR, and node ratio, as well as human evaluation of coverage, relevance, structure, validity, and adequacy.
• In long-context NLP, tasks are plotted over the Diffusion–Scope plane, with book summarization occupying the “high–high” quadrant. Recommendations are made to develop parametric benchmarks with dialable $S$ and $D$ and to report accuracy heatmaps on the $S$–$D$ grid (Goldman et al., 2024).

6. Domain-General Methodologies and Extension Principles

Granular context dependency taxonomies can be generalized beyond their original application:

• The AGI high/low taxonomy (Max, 6 Oct 2025) offers a template for any technical or policy domain: (i) identify context axes (technical, institutional, discursive), (ii) theoretically anchor each, (iii) normalize to $x_i \in [0,1]$, (iv) specify a weighted classification function, (v) map dependencies, (vi) embed dynamic “risk vectors” as additional context, and (vii) validate via case studies and event monitoring.
• The five-step pipeline in mobile context (data collection → discretization → posterior estimation → classifier combination → energy-aware selection) (Rahmati et al., 2012) and adaptive multi-aspect expansion in scientific taxonomy (Zhu et al., 23 Sep 2025) are directly portable with modifications to annotation, clustering, and adaptation layers.

7. Practical Design, Tooling, and Guidance

By systematizing context constructs, granular taxonomies enable more effective tool design, refactoring assistance, adaptive interfaces, and research benchmarks.

• In software engineering, mapping internal and neighbor context features to refactoring pattern labels enables MLP-based or rule-based refactoring advisors to recommend or warn about specific cycle-breaking operations (Feng et al., 2023).
• In mobile and AR systems, dynamic selection of context sources balances energy and predictive accuracy through greedy submodular optimization (Rahmati et al., 2012), supporting “SmartContext” automations.
• For scientific corpus organization, context-aware multi-aspect clustering leads to significant gains in NMI (+8.5), ARI (+2.9), and Purity (+4.6) over previous LLM and non-LLM baselines, and the approach is robust to retrieval noise and over-fragmentation (Zhu et al., 23 Sep 2025).

---

A granular context dependency taxonomy is thus characterized by: (1) explicit cataloging of context features and structures, (2) rigorous measurement and modeling methodologies, (3) operational mapping from context to prediction, adaptation, or classification, and (4) empirical demonstration of granularity-dependent effects and performance. Such taxonomies are modular, extensible, and critical for advancing context-aware adaptive systems, rigorous benchmarking, and theory-driven classification protocols across both technical and sociotechnical domains (Rahmati et al., 2012, Gillick et al., 2014, Feng et al., 2023, Goldman et al., 2024, Davari et al., 2024, Zhu et al., 23 Sep 2025, Max, 6 Oct 2025).