Encode–Cluster–Optimize (ECO) Framework

• Encode–Cluster–Optimize is a modular framework that constructs task-relevant encodings, forms structural groupings through clustering, and optimizes objectives based on these representations.
• It applies versatile techniques—including non-parametric filters, GNNs, and soft cluster assignments—to solve complex problems in graph clustering, video encoding, and sensor networks.
• The framework demonstrates practical efficiency and adaptability by integrating contrastive learning, iterative optimization, and tailored loss functions to improve performance and scalability.

Encode–Cluster–Optimize (ECO) denotes a modular decomposition in which a system first constructs a task-relevant encoding, then forms clusters or other intermediate structural groupings, and finally optimizes an objective or policy conditioned on those representations and groupings. In attributed graph clustering (AGC), ECO is formalized as a mapping from $(\mathbf{A}, \mathbf{X})$ to latent embeddings $\mathbf{Z}$, then to soft cluster assignments $\mathbf{P}$, followed by optimization of a total loss; in several other works, the same three-part structure is used as an explicit synthesis or as a natural interpretive lens for video encoding complexity, columnar compression, encoded distributed optimization, large-scale sensor coding, semi-supervised consensus labeling, explainable learning-to-optimize, and clustered shortest-path trees (Liu et al., 21 Mar 2026, Durbha et al., 28 Jun 2026, Jayanth, 2016, Avestimehr et al., 2018, 0809.1330, Acharya et al., 2012, Heaton et al., 2022, Binh et al., 2021).

1. Formalization and conceptual scope

In AGC, ECO is specified on a graph $$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$ with $N = |\mathcal{V}|$, adjacency matrix $\mathbf{A} \in \{0,1\}^{N \times N}$, node features $\mathbf{X} \in \mathbb{R}^{N \times D}$, and target number of clusters $K$. The framework decomposes an AGC method into representation encoding $\mathcal{E}$, cluster projection $\mathcal{C}$, and optimization strategy, with

$\mathbf{Z}$0

and

$\mathbf{Z}$1

Within that formulation, ECO is presented as a modular taxonomy in which every AGC method is expressed as a triple $\mathbf{Z}$2, the modules are orthogonal and composable, and Table 1 classifies more than 70 methods by encoder type, cluster projector, optimization mode, and core objective (Liu et al., 21 Mar 2026).

The same triad acquires wider scope in the other cited works. There, “Encode” may denote a learned visual embedding shaped by compression response, dictionary encoding of a value-ID array, redundancy injection through a random matrix, soft class-probability estimates from source-trained classifiers, or an implicit optimization model whose objective and feasible set encode prior knowledge. “Cluster” may denote K-Means on embeddings, block partitioning of a column array, hierarchical grouping of correlated sensors, a target-data similarity matrix from a cluster ensemble, or a given partition of graph vertices into clusters. “Optimize” may denote contrastive pretraining, block-size selection, projected gradient descent under stragglers, KLD-optimized factor-graph construction, Bregman-divergence consensus labeling, operator splitting, or multifactorial evolutionary search. This suggests that ECO is less a single algorithm than a recurrent design grammar for structured approximation and control across disparate problem classes (Durbha et al., 28 Jun 2026, Jayanth, 2016, Avestimehr et al., 2018, 0809.1330, Acharya et al., 2012, Heaton et al., 2022, Binh et al., 2021).

2. The Encode stage

In the AGC formulation, the encoder $\mathbf{Z}$3 maps graph topology and node attributes to latent node embeddings $\mathbf{Z}$4. The survey groups encoders into non-parametric graph filters and parametric graph encoders. Non-parametric examples include SGC-style smoothing,

$\mathbf{Z}$5

as well as multi-filtering ensembles and subspace-oriented filtering. Parametric examples include message-passing GNNs, graph Transformers with structural bias, heterophily-aware encoders such as ACM-GNN, and scalable mini-batch encoders such as GraphSAGE-style sampling and GraphSAINT. Their stated goals are to fuse topology and attributes, make clusters geometrically separable in latent space, and address homophily, heterophily, noise, and scale (Liu et al., 21 Mar 2026).

In video encoding complexity clustering, the encode stage is explicitly redirected away from semantics and toward compression behavior. Compression Echo Contrastive Learning (CECL) uses a ViT-L/16 video transformer initialized from V-JEPA, takes masked clips from original uncompressed videos during pretraining, and outputs global feature vectors that are sensitive to textures, motion, and local structure while being shaped by the response of a video to compression. The supervisory signal is the Compression Echo,

$\mathbf{Z}$6

with masked original and compressed videos defined by

$\mathbf{Z}$7

These features are later frozen and reused as encoding-complexity embeddings rather than semantic descriptors (Durbha et al., 28 Jun 2026).

Outside representation learning in the narrow sense, the cited literature treats encoding as any structured transformation that exposes downstream regularity. In column stores, dictionary encoding produces a sorted global dictionary and a value ID array, which then becomes the substrate for post-dictionary compression (Jayanth, 2016). In encoded distributed optimization, the original least-squares problem is transformed by a Gaussian matrix $\mathbf{Z}$8, yielding $\mathbf{Z}$9 and $\mathbf{P}$0, so that redundancy against stragglers is built directly into the data representation (Avestimehr et al., 2018). In OAC$\mathbf{P}$1, the encoding is the classifier-ensemble output

$\mathbf{P}$2

which serves as a soft class-membership description of target instances (Acharya et al., 2012). In “Learning to Optimize,” inference itself is encoded as

$\mathbf{P}$3

so that prior knowledge is embedded in the objective $\mathbf{P}$4 and feasible set $\mathbf{P}$5 (Heaton et al., 2022). In the clustered shortest-path tree setting, the unified search-space chromosome

$\mathbf{P}$6

encodes one inter-vertex choice per cluster index, providing a bi-level representation aligned with clustered graph structure (Binh et al., 2021).

3. The Cluster stage

In AGC, the cluster projector $\mathbf{P}$7 converts embeddings into soft assignments $\mathbf{P}$8. ECO distinguishes discrete post-hoc projectors and differentiable projectors. The former include K-Means, spectral clustering, and subspace clustering, and are used in decoupled or hybrid regimes because they do not backpropagate into the encoder. The latter include prototype-based soft assignment with the Student-$\mathbf{P}$9 kernel

$$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$0

and softmax-based assignment

$$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$1

which enable end-to-end training (Liu et al., 21 Mar 2026).

CECL contains two distinct levels of clustering. During pretraining, it performs on-the-fly pseudo-clustering of scalar Compression Echo values into $$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$2 groups by Bisecting K-Means within the current minibatch, with $$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$3 tested and $$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$4 as the default. Those pseudo-labels define positive and negative pairs for the contrastive objective, after which the compressed masked frames are discarded. At deployment, the pretrained encoder is frozen, an AttentivePooler is trained with Supervised Contrastive Loss using Meta’s RQ-based complexity labels, and K-Means is applied to the resulting video-level embeddings to produce final complexity clusters matching Meta’s fixed choice of 14 clusters (Durbha et al., 28 Jun 2026).

The database compression and sensor-network literatures broaden the meaning of clustering still further. In column stores, cluster encoding partitions the value-ID array into blocks of size $$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$5, compresses all-equal blocks as a single value, and uses a bit vector to mark compressed blocks. The optimization target is

$$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$6

where $$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$7 is the number of uniform blocks. Indirect encoding also partitions into blocks, but seeks low-entropy blocks in which a local dictionary can replace global IDs; its block size is chosen by minimizing average $$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$8-ary entropy across blocks (Jayanth, 2016). In large-scale sensor networks, hierarchical source-optimized clustering groups sensors into disjoint clusters $$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$9 of bounded size $N = |\mathcal{V}|$0 by minimizing the Kullback-Leibler distance between the true Gaussian joint density and a product-of-clusters approximation (0809.1330). In OAC$N = |\mathcal{V}|$1, clustering is a target-data operation performed by one or more clusterers, aggregated into a co-association matrix

$N = |\mathcal{V}|$2

where $N = |\mathcal{V}|$3 is the fraction of clusterers that place $N = |\mathcal{V}|$4 and $N = |\mathcal{V}|$5 together (Acharya et al., 2012). In CluSPT, the clusters are part of the problem definition rather than learned, and each cluster in any feasible spanning tree has exactly one local root; the paper proves that two local roots in the same cluster would imply a cycle, contradicting tree feasibility (Binh et al., 2021).

A common misconception is that the cluster stage must always be an unsupervised partition of data points. The collected work shows otherwise: it can be a differentiable projector, a block partition, a hierarchy over sensors, a similarity graph, or a fixed combinatorial decomposition. The common role is not the specific algorithmic form, but the introduction of an intermediate structure that constrains or simplifies the optimization stage.

4. The Optimize stage

In the AGC survey, optimization strategy is defined as the combination of representation objectives, clustering objectives, and their coordination. Representation objectives include adjacency reconstruction, feature reconstruction, mutual-information or InfoNCE-style contrastive losses, and decorrelation losses such as Barlow Twins. Clustering objectives include semantic embedding-space objectives such as DEC-style KL self-training and structural objectives such as modularity, normalized cut, map equation, structural entropy, and SBM-derived likelihood. ECO organizes these methods into three coordination patterns: decoupled, joint, and hybrid. Decoupled methods train the encoder with $N = |\mathcal{V}|$6 only and apply discrete clustering afterward; joint methods minimize $N = |\mathcal{V}|$7 end-to-end; hybrid methods feed cluster assignments or pseudo-labels back into the encoder through iterative or confidence-gated procedures (Liu et al., 21 Mar 2026).

In CECL, optimization begins with an InfoNCE-like contrastive loss in which positives are clips whose Compression Echo values fall in the same pseudo-cluster and negatives are clips in different pseudo-clusters. It then proceeds to offline complexity-model training with a frozen encoder and AttentivePooler, and finally to offline cluster-specific rate–quality modeling. For each complexity cluster, the method fits mean RQ curves and convex hulls, derives crossover bitrates and bitrate ladders, and at deployment assigns a new video to a cluster through a single forward pass plus AttentivePooler before applying the precomputed cluster-specific ladder instead of full per-title RQ exploration (Durbha et al., 28 Jun 2026).

In the other domains, the optimize stage takes markedly different but structurally homologous forms. In column stores, optimization means both selecting an encoding scheme through heuristics based on cardinality, sortedness, sparsity, repetition, and clusterability, and selecting block sizes by explicit objectives for cluster and indirect encoding (Jayanth, 2016). In encoded distributed optimization, it means running projected gradient descent on encoded data with straggler tolerance. The encoded update uses non-straggler rows $N = |\mathcal{V}|$8,

$N = |\mathcal{V}|$9

and the paper derives a fundamental load–straggler trade-off summarized by

$\mathbf{A} \in \{0,1\}^{N \times N}$0

where $\mathbf{A} \in \{0,1\}^{N \times N}$1 is the minimal computational load and $\mathbf{A} \in \{0,1\}^{N \times N}$2 is the straggler toleration parameter (Avestimehr et al., 2018). In sensor networks, optimization means KLD-optimized factor-graph approximation of the joint PMF together with sum-product decoding, yielding total decoding complexity $\mathbf{A} \in \{0,1\}^{N \times N}$3 for loopy graphs and $\mathbf{A} \in \{0,1\}^{N \times N}$4 for cycle-free graphs, linear in the number of sensors for fixed $\mathbf{A} \in \{0,1\}^{N \times N}$5 and $\mathbf{A} \in \{0,1\}^{N \times N}$6 (0809.1330). In OAC$\mathbf{A} \in \{0,1\}^{N \times N}$7, the optimize stage is the alternating minimization of

$\mathbf{A} \in \{0,1\}^{N \times N}$8

with updates derived from Bregman-centroid identities and Legendre duality; the paper proves convergence to the unique global minimizer and a $\mathbf{A} \in \{0,1\}^{N \times N}$9-linear rate under its stated conditions (Acharya et al., 2012). In learning-to-optimize, operator splitting itself is the optimize stage: proximal gradient, linearized ADMM, and Davis–Yin splitting are used to solve structured implicit models, while interpretable certificates derived from property values classify outputs as pass, warning, or fail (Heaton et al., 2022). In CluSPT, optimization is a multifactorial evolutionary search over local-root assignments, combined with Dijkstra-based exact intra-cluster shortest-path trees and approximate inter-cluster construction, with a repair procedure that guarantees validity of decoded solutions (Binh et al., 2021).

5. Representative instantiations across domains

The following examples show how the same three-part pattern recurs with different mathematical objects and operational targets.

Domain	Encode	Cluster–Optimize
Attributed graph clustering	Graph filters, GNNs, Transformers, heterophily-aware encoders	K-Means, spectral, subspace, prototypes, or softmax; decoupled, joint, or hybrid losses
Video encoding complexity	ViT-L/16 features shaped by Compression Echo	Bisecting K-Means pseudo-labels; K-Means complexity clusters; per-cluster RQ curves and ladders
Columnar databases	Dictionary encoding into value IDs	Block partitioning; block-size optimization; heuristic encoding selection
Distributed least squares	Gaussian encoding $\mathbf{X} \in \mathbb{R}^{N \times D}$0	Worker-row layout under stragglers; encoded PGD with load–tolerance trade-offs
Sensor networks	Scalar quantization plus distortion-optimized index assignments	KLD-based hierarchical clustering; KLD-optimized factor graph; sum-product decoding
Semi-supervised transfer	Classifier-ensemble soft labels $\mathbf{X} \in \mathbb{R}^{N \times D}$1	Cluster-ensemble similarity matrix; Bregman-divergence consensus labeling
Clustered shortest-path trees	Unified chromosome over inter-vertices	Given vertex clusters; repair plus Dijkstra-assisted evolutionary search
Explainable learning-to-optimize	Objectives and feasible sets encode priors	Structural grouping is implicit; operator splitting and certificates govern trust

In video encoding, CECL provides one of the clearest operational ECO instantiations. On LAVIB 720p labels, CECL reached ARI $\mathbf{X} \in \mathbb{R}^{N \times D}$2, NMI $\mathbf{X} \in \mathbb{R}^{N \times D}$3, and FMI $\mathbf{X} \in \mathbb{R}^{N \times D}$4, the best among the reported baselines. For real-world BD metrics against a fixed ladder, CECL achieved BD-Rate $\mathbf{X} \in \mathbb{R}^{N \times D}$5 on Inter4K and $\mathbf{X} \in \mathbb{R}^{N \times D}$6 on YouTube-UGC, compared with V-JEPA at $\mathbf{X} \in \mathbb{R}^{N \times D}$7 and $\mathbf{X} \in \mathbb{R}^{N \times D}$8; the oracle upper bound using Meta’s ground-truth cluster labels was $\mathbf{X} \in \mathbb{R}^{N \times D}$9 mean BD-Rate on Inter4K and $K$0 on YouTube-UGC (Durbha et al., 28 Jun 2026). The paper also notes that CECL adds only about $K$1 extra pretraining cost beyond V-JEPA.

In sensor coding, the Encode–Cluster–Optimize interpretation corresponds to low-complexity distributed quantization, KLD-based source clustering, and factor-graph decoding. For $K$2 randomly placed sensors with strong correlation $K$3, the scheme’s “IR” mode improved output SNR over decoder-only operation from $K$4 dB to $K$5 dB at $K$6 bit and from $K$7 dB to $K$8 dB at $K$9 bits. In the quadratic Gaussian CEO setting with $\mathcal{E}$0 and $\mathcal{E}$1, IR raised SNR from $\mathcal{E}$2 dB to $\mathcal{E}$3 dB at $\mathcal{E}$4 bit and from $\mathcal{E}$5 dB to $\mathcal{E}$6 dB at $\mathcal{E}$7 bits (0809.1330).

In multifactorial optimization for CluSPT, the bi-level encoding and cluster-aware evaluation produce large empirical gains. The proposed K-MFEA outperformed HB-RGA on 138 of 139 instances, outperformed G-MFEA on 108 of 139, was equal on 29, and lost only on 2 large instances. It was also consistently faster than both baselines; on large instances, HB-RGA could be $\mathcal{E}$8 slower than K-MFEA, and disabling the paper’s speed-up mechanism increased running time by $\mathcal{E}$9–$\mathcal{C}$0 on large types and $\mathcal{C}$1–$\mathcal{C}$2 on Types 3 and 4 (Binh et al., 2021).

In semi-supervised and transfer learning, OAC$\mathcal{C}$3 casts classifier outputs as encodings, cluster-ensemble structure as soft constraints, and consensus labeling as a convex Bregman optimization problem. Under matched base models and data, it achieved higher accuracy than BGCM, MCLA, and HBGF on all 11 reported classification tasks, and it also outperformed S$\mathcal{C}$4VM, BGCM, and the pure classifier ensemble on most of six semi-supervised benchmarks (Acharya et al., 2012). In explainable learning-to-optimize, the same structural logic appears as prior-aware implicit modeling plus certificate-based post-conditions; on sparse-angle CT reconstruction, the L2O model achieved the best average PSNR and SSIM and the lowest failure rates on the reported certificates, while in crypto arbitrage it achieved $\mathcal{C}$5 trade execution with executed trades always profitable, whereas the analytic method had $\mathcal{C}$6 execution and always failed the risk certificate (Heaton et al., 2022).

By contrast, the columnar database paper is chiefly analytical and algorithmic: it provides explicit block-size objectives, recurrences, pseudocode, and heuristics, but does not provide detailed numerical evaluation comparisons (Jayanth, 2016). That contrast is itself informative: some ECO instantiations are empirical end-to-end systems, while others are principled decomposition-and-selection frameworks whose main contribution lies in objective design and decision logic.

6. Evaluation, misconceptions, limitations, and future directions

The strongest explicit critique of ECO-style work appears in the AGC survey’s evaluation section. It identifies dataset myopia, scalability neglect, and a metric paradox in which an unsupervised task is evaluated only with label-dependent metrics. In response, it advocates a holistic standard comprising supervised semantic alignment when labels exist, unsupervised structural integrity through metrics such as modularity and conductance, and efficiency and scalability profiling through peak GPU memory, total training plus clustering time, inference latency, and failure modes such as out-of-memory errors (Liu et al., 21 Mar 2026).

A second recurring issue is that intermediate clustering quality does not necessarily coincide with downstream operational quality. CECL reports that it always outperforms V-JEPA in BD metrics, even where clustering metrics are not always best, which directly cautions against treating proxy clustering scores as sufficient operational criteria (Durbha et al., 28 Jun 2026). This suggests that, in ECO systems, the optimize stage should be judged on the actual terminal objective—rate–distortion, query efficiency, convergence rate, reconstruction trustworthiness, or tree cost—rather than only on the quality of the cluster representation.

The surveyed literature also exposes domain-specific limits. CECL pretraining is tied to a specific codec, preset, and CRF set; generalization to other codecs or presets is described as plausible but not guaranteed, and the paper works at video level rather than explicit per-shot or per-segment clustering (Durbha et al., 28 Jun 2026). The columnar compression heuristics depend on tunable thresholds $\mathcal{C}$7, $\mathcal{C}$8, and $\mathcal{C}$9, and the size formulas often neglect bit-vector overhead as negligible (Jayanth, 2016). The encoded optimization analysis is specific to least squares with Gaussian encoders, while randomized DCT encoders are only shown empirically to behave similarly (Avestimehr et al., 2018). OAC$\mathbf{Z}$00 relies on the assumption that similar target instances are more likely to share class labels; when cluster structure and class structure diverge, the clustering term can hurt (Acharya et al., 2012). The learning-to-optimize framework works best when domain experts can express approximate models, constraints, and priors in optimization form (Heaton et al., 2022). In CluSPT, the paper identifies graph size as a possible influential factor, with the advantage of the bi-level encoding becoming more pronounced on larger instances (Binh et al., 2021).

Future directions are most explicitly charted in AGC. The survey prioritizes heterophily-robust and noise-aware encoders, better projectors supporting unknown $\mathbf{Z}$01 and interpretable prototypes, scalable joint optimization, unsupervised model selection, and stronger theory on cluster recoverability under varying homophily, feature signal-to-noise ratio, and sparsity (Liu et al., 21 Mar 2026). Read alongside the other papers, a plausible implication is that future ECO systems will increasingly couple domain-specific encodings with cluster structures chosen for operational rather than merely geometric fidelity, and will evaluate them by end-task efficiency, robustness, and trust rather than by representation quality alone.