Dual-Granularity Algorithm: Multi-Scale Processing

• Dual-Granularity Algorithm is a computational approach that processes data at both fine and coarse levels to optimize accuracy and efficiency.
• It is applied across domains such as graphical modeling, metric learning, and contrastive representation, addressing both local details and global structures.
• The paradigm employs adaptive partitioning and dual-loss integration, enabling dynamic control over computational trade-offs and improved performance.

A dual-granularity algorithm is a computational approach that explicitly models or manipulates data, processes, or representations at two distinct levels of granularity—typically “fine” and “coarse”—to improve efficiency, accuracy, robustness, or interpretability in complex systems. This paradigm arises in diverse domains such as probabilistic graphical modeling, metric learning, contrastive representation learning, data sampling, graph analysis, and other applications demanding simultaneous attention to both local and global structures or behaviors.

1. Definitions and Fundamental Principles

Dual-granularity algorithms operate by recognizing that treating all components of a system at a uniform granularity may be computationally suboptimal or hinder expressiveness. Instead, these algorithms adaptively or explicitly organize computation, information exchange, or learning objectives at two (or more) granular scopes, such as:

• Temporal scales: e.g., variable-length intervals in CTBN inference (Saria et al., 2012).
• Spatial regions: e.g., patch-wise and image-level features in dynamic quantization (Wang et al., 22 Sep 2024).
• Representation levels: e.g., sample-based vs. center-based loss in metric learning (Liu et al., 2020); cluster-level and node-level kernels in graph transformers (Huang et al., 9 Oct 2024).
• Information abstraction: e.g., raw token-level prompts for targets and summarized neighborhood text for context in graph-augmented LLMs (Li et al., 29 Jul 2025).

The strategic decoupling or simultaneous fusion of dual granularities enables control over computational trade-offs, selective focus on pertinent phenomena, and more flexible adaptation to heterogeneous or multi-scale data.

2. Dual-Granularity in Inference and Learning Frameworks

Temporal and Structural Dual Granularity: The “Dynamic-EP” algorithm for CTBNs (Saria et al., 2012) exemplifies an inference algorithm where both spatial and temporal granularities are manipulated. Clusters in a cluster graph each cover a subset of variables and an associated (possibly overlapping) time interval, and can subdivide their temporal scope (using information-theoretic criteria) when local dynamics become heterogeneous. This enables piecewise Markov process approximations only where fine granularity is warranted, dramatically reducing computational effort compared to uniform discretization.

Metric Learning: The dual-granularity triplet loss (DGTL) in VT-ReID (Liu et al., 2020) combines a sample-based (fine) branch that focuses on individual instances, with a center-based (coarse) branch that aligns aggregated features from different modalities. The loss is hierarchically arranged:

• Fine: Hard positive/negative mining in feature space (sample discrimination).
• Coarse: Enforces compactness and separation of identity centers across modalities (center-level discrimination).

This arrangement achieves simultaneously strong intra-class and inter-class feature organization.

Contrastive and Representation Learning: Dual-granularity contrastive frameworks (Wang et al., 2023, Mehri et al., 2019) incorporate contrastive objectives both at a coarse (e.g., item/session) and a fine (e.g., factor/latent) level. For instance, factor-level disentanglement via dedicated convolution channels captures latent semantic facets, while item-level augmentations (such as star graph augmentations) address connectivity and sparsity.

3. Granularity Management, Dynamic Adaptation, and Control

Adaptive Granularity Partitioning: A key requirement in dual-granularity algorithms is the ability to adaptively select or partition data/processes at the appropriate granularity. For example, the Dynamic-EP algorithm (Saria et al., 2012) applies a local KL-divergence-based criterion to trigger sub-interval splits within cluster time scopes:

$$C_{KL}^{\mathrm{two-piece}} = D(P_c \mid P_{s_l}) + D(P_c \mid P_{s_r}),$$

and compares against a thresholded improvement over a single-piece approximation, determining the necessity of finer updates.

Explicit Control Signals: In interactive segmentation (Zhao et al., 1 May 2024), granularity-controllable models introduce an input parameter specifying granularity, which is processed via dedicated embeddings and influences the segmentation mask prediction. This allows explicit output control across a continuum from fine (object parts) to coarse (whole objects).

Multiscale Aggregation: Methods in quantization (Wang et al., 22 Sep 2024) and graph coarsening (Xia et al., 24 Jun 2025) dynamically partition or refine representations across scales. Hierarchical splitting (e.g., granular-balls in graph coarsening; patch entropy-based bit allocation in quantization) is governed by data-driven metrics, such that further subdivision is only performed when local quality (connectivity, information density, etc.) justifies it.

4. Theoretical Foundations and Analytical Mechanisms

Information-Theoretic Criteria: Theoretical justifications for controlling granularity often rely on measuring approximation error or structural information loss (e.g., using KL divergence (Saria et al., 2012), or spectral distance in graph coarsening (Xia et al., 24 Jun 2025)) and setting thresholds for when to refine or merge.

Dual-Kernel or Dual-Loss Mechanisms: In dual-granularity attention architectures (Huang et al., 9 Oct 2024), attention scores are computed via combinations (either tensor products or convex linear combinations) of node-level and cluster-level kernels:

$$\text{N2C-Attn-L}(X)_i = \frac{\sum_{j} A_{i,j}^p \sum_t C_{t,j} (\alpha \kappa_C(Q_i, K_j) + \beta \kappa_N(q_i, k_t)) v_t}{\sum_{j} A_{i,j}^p \sum_t C_{t,j} (\alpha \kappa_C(Q_i, K_j) + \beta \kappa_N(q_i, k_t))}.$$

This allows simultaneous and weighted integration of multiple granularities in attention or message passing.

Optimization over Granularity Levels: Retrieval-Augmented Generation tasks (Zhong et al., 1 Jun 2024) utilize routers that are trained (with soft-label BCE loss) to select or mix retrieval chunk sizes dynamically, optimizing over multiple granularities per query.

5. Empirical Performance and Computational Benefits

Empirical evaluations across application domains consistently demonstrate that dual-granularity algorithms provide strong trade-offs between computational efficiency and predictive accuracy:

• Inference in dynamic graphical models (Saria et al., 2012): Comparable accuracy to uniformly fine discretizations with an order-of-magnitude computational saving.
• VT-ReID (Liu et al., 2020): Substantial improvements in rank-1 accuracy and mean average precision over single-granularity ReID baselines.
• Session-based recommendation (Wang et al., 2023): Higher precision and mean reciprocal rank than GNN or CL baselines; joint item/factor-level CL addresses sparsity and subtlety.
• Graph coarsening (Xia et al., 24 Jun 2025): Orders-of-magnitude speedup with improved or preserved classification and spectral fidelity over prior methods.
• Super-resolution with dynamic quantization (Wang et al., 22 Sep 2024): Lower bit usage, competitive PSNR/SSIM, and substantial reduction in BitOPs versus state-of-the-art quantization, with explicit patch- and layer-level control.

6. Broader Implications and Applications

The dual-granularity paradigm is naturally suited to domains where different system components, attributes, or data instances demand varying degrees of representation or processing:

• Healthcare: Medication recommendation leveraging both entity-level causality and molecular structure (Liang et al., 1 Mar 2024).
• Industrial diagnostics: Class-incremental few-shot learning with class-agnostic and class-specific parallel branches (Yang et al., 16 Aug 2025).
• Interactive annotation: Granularity-controllable segmentation enhancing both efficiency and annotation precision (Zhao et al., 1 May 2024).
• Fraud detection and explainable AI: Bi-level prompting for LLMs to preserve fine detail where needed while compressing context for scalability (Li et al., 29 Jul 2025).
• Graph learning: Simultaneous modeling of global structural patterns and localized node/cluster information in hierarchical attention and message-passing (Huang et al., 9 Oct 2024).
• Optimization and search: Adaptive coverage of the solution space at coarse scale with local refinement in promising subregions (Xia et al., 2023).

Applications extend across domains where scale, heterogeneity, or cost-sensitive modeling necessitates strategic allocation of computational and representational resources.

7. Limitations and Parameter Tuning

Dual-granularity models typically introduce additional hyperparameters—such as KL-divergence thresholds, confidence weights, adaptive mask generators, or selection criteria for boundary-aware instances—which must be carefully calibrated to balance trade-offs in accuracy and efficiency. Practical deployments may require data-driven or domain-specific strategies for setting these parameters. In some cases, the challenge of selecting optimal thresholds or fusion schemes across dual scales remains an open area for further research, particularly in highly heterogeneous or evolving systems.

---

The dual-granularity algorithmic paradigm has emerged as a powerful approach for reconciling fine-level expressivity with coarse-level computational or representational efficiency. By recognizing, modeling, and dynamically controlling processes at multiple granularities, these algorithms enable advances in computational tractability, predictive accuracy, robustness to noise and sparsity, and interpretability across a wide range of AI and data science tasks.