Feature-Enriched Hierarchical Graphs

• Feature-Enriched Hierarchical Graphs (FEHGs) are graph representations that combine multilevel structural organization with explicit feature enrichment for more effective inference.
• They are applied across domains such as binary reverse engineering, navigation, and graph learning, delivering improvements in precision, recall, and localization accuracy.
• FEHG techniques integrate learned features and structural hierarchies through methods like coarsening, message passing, and latent embedding to optimize performance.

Searching arXiv for papers explicitly using or closely matching “Feature-Enriched Hierarchical Graphs” and key related formulations. Feature-Enriched Hierarchical Graphs (FEHGs) are graph representations in which hierarchical organization and feature enrichment are coupled across multiple levels of abstraction. In the most explicit formulation, introduced in AGNOMIN, a binary is represented as an FEHG $G=(V,E,\Psi)$, where function nodes form a top-level graph and each function node expands into a Feature-Enriched Control Flow Graph $f_i=(BB,E,\psi)$; both function nodes and basic-block nodes carry learned PCode-derived features (Achamyeleh et al., 29 Sep 2025). More broadly, the literature contains a family of closely related constructions in which hierarchy may be over node clusters, graph resolutions, scene layers, hypergraph abstractions, or even relation dimensions, while enrichment may come from explicit attributes, learned latent structure, edge features, priors, or memory states.

1. Terminological scope and canonical formulation

In the narrowest sense, FEHG refers to the binary-analysis representation used by AGNOMIN. There, the top level is a function-call graph over functions $f_i$, and each function node is itself a lower-level graph over basic blocks. Feature enrichment is explicit at both levels: $\nu_i=\text{T5}(\text{PCode}(bb_i))$ for basic blocks and $\mu_i=\text{T5}(\text{PCode}(f_i))$ for whole functions. The representation is therefore a hierarchical graph-of-graphs in which structural relations and learned semantic features are co-defined (Achamyeleh et al., 29 Sep 2025).

In a broader and now common research reading, FEHG denotes any graph formalism that satisfies two conditions. First, it must encode more than one structural level, whether by coarsening, nested graph layers, scene abstractions, or multiresolution organization. Second, it must enrich those levels with feature channels that are meaningful for inference rather than treating hierarchy as purely combinatorial. Under that broader reading, Hierarchical Graph Neural Networks supply a generic multi-level template with graph layers $(L^h,A^h)$, assignment matrices $C^h$, and inter-level mappings $H^{h\to h+1}$ and $H^{h+1\to h}$, while DeepGL supplies an explicit hierarchy of relational feature definitions $\mathcal{F}_1,\mathcal{F}_2,\dots$ that recursively enrich node or edge descriptions (Sobolevsky, 2021, Rossi et al., 2017).

The boundary of the term is not uniform across papers. HMGE, for example, is FEHG-relevant “in spirit” but not canonical in the node-coarsening sense, because its hierarchy is over multiplex graph dimensions rather than over nodes, clusters, or subgraphs. Its enrichment comes from recursively learned latent adjacency operators and progressively refined node embeddings, not from explicit supernodes or engineered multilevel feature channels (Abdous et al., 2023).

2. Axes of hierarchy

One major FEHG axis is node- or cluster-centered hierarchy. In HGNN, level $f_i=(BB,E,\psi)$0 is obtained from level $f_i=(BB,E,\psi)$1 by the coarsening rule $f_i=(BB,E,\psi)$2, and node features at every level are updated by horizontal same-level propagation together with vertical bottom-up and top-down messages. This yields a genuine multi-resolution graph stack rather than a single pooled latent vector (Sobolevsky, 2021). A related but more explicitly feature-centric hierarchy appears in DeepGL, where each deeper layer composes relational operators over prior-layer features, so that $f_i=(BB,E,\psi)$3 defines higher-order structural descriptors (Rossi et al., 2017).

A second axis is hierarchy over explicit scene or system strata. In the navigation framework built on Dynamic Scene Graphs, the environment is represented with Place, Object, Room, and Action nodes, while the full underlying DSG includes five layers from metric-semantic mesh up to building level. In S-Graphs+, the hierarchy is a four-layer factor graph comprising keyframes, walls, rooms, and floors, optimized jointly for localization and mapping (Ravichandran et al., 2021, Bavle et al., 2022). These are FEHGs in a metric-semantic sense: higher levels are not merely summaries but active structural entities coupled to lower-level estimation.

A third axis is relation- or channel-centered hierarchy. HMGE starts from a multiplex graph $f_i=(BB,E,\psi)$4 with shared nodes and features, then recursively combines adjacency matrices according to

$f_i=(BB,E,\psi)$5

so that higher-level dimensions become learned compositions of lower-level relation types (Abdous et al., 2023). This suggests that FEHGs need not always be built by node coarsening; they can also be built by semantic reorganization of relational channels.

3. Mechanisms of feature enrichment

Feature enrichment in FEHGs ranges from learned semantic encoders to explicit geometric or statistical feature construction. In AGNOMIN, enrichment is learned from Ghidra PCode using a pre-trained T5 model at both the basic-block and function levels. The lower-level GCN over each FECFG produces block embeddings $f_i=(BB,E,\psi)$6, those are summed into $f_i=(BB,E,\psi)$7, and a higher-level GAT over the function-call graph produces $f_i=(BB,E,\psi)$8; concatenating $f_i=(BB,E,\psi)$9 yields the final function embedding $f_i$0 (Achamyeleh et al., 29 Sep 2025). The FEHG is therefore feature-enriched both before and after hierarchical message passing.

DeepGL represents a contrasting enrichment regime: features are explicit, symbolic, and interpretable rather than latent neural states. Base features can include graphlet orbit counts, degree-derived statistics, egonet features, and attributes; deeper features are formed by neighborhood operators such as mean, sum, max, product-like operators, and diffusion transforms. The framework outputs both the feature matrix and the feature definitions, so enrichment is not only numerical but also semantic and transferable across graphs (Rossi et al., 2017).

In embodied and spatial FEHGs, enrichment is frequently hand-designed but structurally coupled. The DSG-based navigation model assigns each node a 10-dimensional vector containing robot-relative position, object bounding-box dimensions when applicable, node category and semantic class, a binary visited/history flag, and ESDF-based occupancy or clearance. The graph itself accumulates over time, so memory is encoded both structurally and in node state (Ravichandran et al., 2021). S-Graphs+ similarly enriches nodes with plane parameters, room centers, floor centers, class tags such as vertical or horizontal wall, and free-space-derived support relations; its edges are optimization factors carrying geometric consistency rather than simple symbolic predicates (Bavle et al., 2022).

Other FEHG variants emphasize enrichment beyond node attributes alone. The molecular sparse hierarchical model updates both node features and edge features, then defines effective superedges during pooling so that coarse graphs remain chemically meaningful (Bal et al., 2019). The multiresolution feature- and prior-based classification framework enriches graphs through feature-aware edge weights, prior-based sample-to-class edges, classifier reliabilities, and Tukey-based implicit boundary detection (Fallah, 2023). HGFE enriches CNN feature maps by turning local windows and pooled supernodes into two graph levels and modulating propagation through an adaptive low-/high-frequency gate, so that local detail and global context are fused back into the visual feature tensor (Zhao et al., 15 Aug 2025).

4. Learning and inference paradigms

FEHGs do not prescribe a single optimization regime. One line uses hierarchical message passing with latent grouping. GraphHAM assigns each node a layer-specific mixed membership $f_i$1, samples a manifested group $f_i$2 via Gumbel-Softmax, and combines individual-level attention $f_i$3 with group-level attention $f_i$4 in the update

$f_i$5

Its hierarchy is regularized by must-link and cannot-link constraints across layers, so the FEHG is learned as a latent multi-granularity grouping structure rather than being pre-specified (Lin et al., 2021).

A second line uses self-supervised pretraining. Hyper-FM first enriches text-derived vertex features through Hierarchical High-Order Neighbor Guided Vertex Knowledge Embedding, then constructs a hierarchical multi-hypergraph with sampled vertices, cluster-level virtual vertices, and domain-level bond vertices. Pretraining uses

$f_i$6

where the paper describes $f_i$7 as HyperGCL-style structural contrastive learning and $f_i$8 as GraphMAE-style masked feature reconstruction (Feng et al., 3 Mar 2025). The result is a foundation-model-style FEHG pipeline in which feature enrichment precedes hierarchical structural pretraining.

A third line uses unsupervised graph-level objectives over learned hierarchies. HMGE maximizes mutual information between node embeddings and a graph summary vector after recursively composing relation dimensions, training by corrupted-feature contrast in the style of Deep Graph Infomax (Abdous et al., 2023). In multimodal emotion recognition, HFGCN performs a two-stage hierarchy—modality graph within each utterance, then utterance graph across the conversation—and trains with a multitask loss over emotion, valence, and arousal after relational graph convolution and pooling (Tang et al., 2021). In robotics, S-Graphs+ uses weighted nonlinear least-squares factor graph optimization rather than representation-learning losses, summing odometric, pose-plane, room-wall, floor-room, and loop-closure costs over the full hierarchy (Bavle et al., 2022). These cases show that FEHGs can be optimized by contrastive, reconstructive, discriminative, reinforcement-learning, or geometric-estimation objectives depending on domain.

5. Representative domains and empirical roles

The most literal FEHG application in the surveyed literature is binary reverse engineering. AGNOMIN reports improvements of up to $f_i$9 in precision and $\nu_i=\text{T5}(\text{PCode}(bb_i))$0 in recall across its testing dataset, and on unseen architectures it achieves $\nu_i=\text{T5}(\text{PCode}(bb_i))$1 higher recall than the closest baseline. The empirical role of the FEHG there is architecture-agnostic semantic alignment: CFG structure, FCG context, and PCode-derived features are combined into function embeddings that support multi-label name prediction and cross-architecture function matching (Achamyeleh et al., 29 Sep 2025).

In navigation and SLAM, FEHGs act as world models with explicit memory and semantic abstraction. The DSG-based navigation system uses accumulated hierarchical scene graphs plus an Action layer to drive reinforcement learning, achieving $\nu_i=\text{T5}(\text{PCode}(bb_i))$2 targets found in object search versus $\nu_i=\text{T5}(\text{PCode}(bb_i))$3 for RGB-D + semantics and $\nu_i=\text{T5}(\text{PCode}(bb_i))$4 for ESDF baselines, while its “No Memory” ablation drops target finding to $\nu_i=\text{T5}(\text{PCode}(bb_i))$5 (Ravichandran et al., 2021). S-Graphs+ uses its four-layer semantic-metric factor graph to improve localization and mapping, outperforming the second-best method by $\nu_i=\text{T5}(\text{PCode}(bb_i))$6 on average over its datasets (Bavle et al., 2022). In both cases, hierarchy is operational rather than decorative: it supplies constraints that improve long-horizon reasoning or metric accuracy.

In graph representation learning, FEHG-like mechanisms are especially useful when flat aggregation is structurally inadequate. HMGE shows that as multiplex dimensionality increases from 3 to 41, it still reaches about $\nu_i=\text{T5}(\text{PCode}(bb_i))$7 classification accuracy while competing methods fall below $\nu_i=\text{T5}(\text{PCode}(bb_i))$8, which the paper attributes to latent compositional structures emerging across dimensions (Abdous et al., 2023). Hyper-FM reports that its multi-domain hierarchical hypergraph pretraining outperforms baselines by approximately $\nu_i=\text{T5}(\text{PCode}(bb_i))$9, emphasizing the value of joint feature enrichment and structural hierarchy across domains (Feng et al., 3 Mar 2025).

In vision and multimodal learning, FEHG-like structures often serve as compact alternatives to flat fusion or quadratic global attention. HFGCN shows that using both intra-utterance and inter-utterance graphs yields an average F1 of $\mu_i=\text{T5}(\text{PCode}(f_i))$0 on IEMOCAP 4-class, compared with $\mu_i=\text{T5}(\text{PCode}(f_i))$1 using only the first stage and $\mu_i=\text{T5}(\text{PCode}(f_i))$2 using only the second stage (Tang et al., 2021). HGFE improves CIFAR-100 Top-1 from $\mu_i=\text{T5}(\text{PCode}(f_i))$3 to $\mu_i=\text{T5}(\text{PCode}(f_i))$4 and improves VisDrone $\mu_i=\text{T5}(\text{PCode}(f_i))$5 from $\mu_i=\text{T5}(\text{PCode}(f_i))$6 to $\mu_i=\text{T5}(\text{PCode}(f_i))$7, with ablations indicating that the hierarchical combination of intra-window and inter-window graphs is more effective than either component alone (Zhao et al., 15 Aug 2025).

6. Boundaries, misconceptions, and open technical issues

A common misconception is that FEHGs necessarily imply node pooling into supernodes. The literature is more heterogeneous. HGNN, edge-aware molecular pooling, S-Graphs+, and DSG-style navigation all instantiate explicit multilevel objects or strata, but HMGE is hierarchical over relation dimensions, not nodes, and is therefore better described as FEHG-adjacent than canonical (Sobolevsky, 2021, Bal et al., 2019, Bavle et al., 2022, Abdous et al., 2023). Conversely, a hierarchical graph is not automatically feature-enriched: the hierarchy must carry informative features, priors, memory, or learned latent channels that affect inference.

A second misconception is that enrichment must be purely neural or purely attribute-based. DeepGL shows a non-neural, interpretable, transferable FEHG-style construction in which enrichment is explicit relational feature synthesis (Rossi et al., 2017). S-Graphs+ shows a geometric-semantic version in which enrichment is carried by plane parameters, room centers, and factor residuals rather than embeddings (Bavle et al., 2022). Hyper-FM and AGNOMIN show learned-feature variants based on BERT and T5, respectively (Feng et al., 3 Mar 2025, Achamyeleh et al., 29 Sep 2025). FEHGs are therefore better understood as a design pattern than as a fixed architecture class.

Open technical issues recur across the literature. Some methods use shallow or externally induced hierarchies: HiGen provides a strong structural scaffold for hierarchical graph generation, but it assumes graphs do not have node attributes and obtains its hierarchy from partitioning such as Louvain rather than learning it jointly (Karami, 2023). HG2NP introduces a two-phase hierarchy for heterogeneous graph generation with node feature pools, but its hierarchy is only two-stage, edge types are not explicitly generated, and experiments are capped at graphs with at most 200 nodes (Ghosh et al., 2024). In robotics, the DSG-based navigation policy is sensitive to node position noise and omits dense mesh-level reasoning for efficiency (Ravichandran et al., 2021). In HMGE, scalability still depends explicitly on the number of dimensions, including an $\mu_i=\text{T5}(\text{PCode}(f_i))$8 time term and a $\mu_i=\text{T5}(\text{PCode}(f_i))$9 memory term (Abdous et al., 2023). At the opposite extreme, the algebraic theory of graph lineages supplies a rigorous multiscale backbone—graded graphs, bipartite inter-level structure, prolongation maps, skeletal products, and thickening—but does not itself provide learned feature processing, so it is best viewed as a hierarchy-and-composition substrate onto which FEHG learning mechanisms could be added (Mjolsness et al., 31 Jul 2025).

Taken together, these distinctions indicate that FEHGs are not a single standardized model family but a convergent research theme: represent multiple graph scales or semantic strata explicitly, enrich those strata with features that matter for downstream inference, and allow information to move across levels in a controlled way. The exact realization may be symbolic, neural, geometric, probabilistic, or algebraic, but the defining technical idea is the same: hierarchy and feature enrichment are jointly constitutive of the graph representation rather than post hoc additions.