Dynamic Exploratory Graph Analysis (DynEGA)

• DynEGA is a dynamic graph analysis framework that systematically examines evolving community structures over ordered dimensions like embedding depth and time.
• It uses a multi-objective optimization approach by balancing normalized mutual information and total entropy fit index to select the optimal embedding depth.
• The method also clusters timestamped edge events into coherent 'conversations,' enabling enhanced visualization and machine learning applications.

Dynamic Exploratory Graph Analysis (DynEGA) denotes a graph-analytic workflow for ordered data in which network structure is examined as the ordering variable changes rather than being treated as fixed. In the embedding-based formulation, DynEGA “systematically traverse[s] embedding coordinates, treating the dimension index as a pseudo-temporal ordering analogous to intensive longitudinal trajectories,” so that increasing embedding depth corresponds to walking through an embedding landscape and monitoring how community structure evolves (Golino, 14 Jan 2026). In a dynamic-graph workflow, timestamped edges are clustered into temporally and topologically coherent “conversations,” and these induced subgraphs become the atomic units for downstream exploration, visualization, and machine learning (Ostroski et al., 2023). Taken together, these formulations present DynEGA as an ordered-graph framework in which the central problem is not only estimating communities, but determining how graph organization changes across a meaningful axis such as time, edge chronology, or embedding depth.

1. Core conceptualization

DynEGA is motivated by a rejection of the static view of representation. In the LLM-embedding setting, the “traditional view” is that an embedding is a static $D$-dimensional vector $x \in \mathbb{R}^D$ used “as-is” in downstream analyses. The “landscape view” instead treats each coordinate index $j = 1 \ldots D$ as a pseudo-time dimension, so that each item embedding $e_i = (e_{i,1}, \ldots, e_{i,D})$ becomes a trajectory through a semantic manifold as the coordinate index increases (Golino, 14 Jan 2026).

This reframing is operational rather than merely metaphorical. By increasing the embedding depth $d$, that is, by subsetting the first $d$ coordinates, DynEGA evaluates a sequence of network estimates and community partitions. The resulting object is an optimization landscape over depth rather than a single graph. The reported findings state that “embedding landscapes” are “non-uniform semantic spaces requiring principled optimization rather than default full-vector usage” (Golino, 14 Jan 2026).

A related ordered-graph perspective appears in the dynamic-edge setting. There, the basic data are timestamped relational events between vertices, and the organizing problem is to group records “based on how and when they are occurring.” The summary integrated with DynEGA describes these groups as “higher-order events,” specifically edge-clusters or “conversations,” that can be visualized, compared, and embedded for downstream analyses (Ostroski et al., 2023). This suggests that DynEGA is best understood as a family of workflows in which graph structure is explored across an ordered axis rather than estimated once at a fixed representation.

2. Embedding-landscape formulation

In the embedding application, the input is a full embedding matrix $E \in \mathbb{R}^{N \times D}$, true dimension labels $L_i \in \{1 \ldots F\}$ when available for evaluation, a depth grid $d_1 < d_2 < \cdots < d_m$, and weights $w_{\mathrm{NMI}}$ and $x \in \mathbb{R}^D$0 (Golino, 14 Jan 2026). For each embedding depth $x \in \mathbb{R}^D$1, the workflow subsets the embeddings to $x \in \mathbb{R}^D$2. Optional state-space reconstruction and derivative estimation can then be performed via GLLA, using time-delay embedding and the transformation $x \in \mathbb{R}^D$3.

Network estimation proceeds on either raw correlations of $x \in \mathbb{R}^D$4 or on derivatives $x \in \mathbb{R}^D$5. The specific estimator reported is a partial-correlation network via the Triangulated Maximally Filtered Graph (TMFG), followed by community detection with the Walktrap algorithm, yielding a partition $x \in \mathbb{R}^D$6 into communities (Golino, 14 Jan 2026). The pipeline therefore combines depth-wise representation truncation, network estimation, and partition comparison.

The optimization target is not fixed a priori. For each depth, DynEGA computes both normalized mutual information and the Total Entropy Fit Index, then normalizes TEFI to $x \in \mathbb{R}^D$7 and forms the composite score

$x \in \mathbb{R}^D$8

The selected depth is

$x \in \mathbb{R}^D$9

In the reported implementation, $j = 1 \ldots D$0 and $j = 1 \ldots D$1, reflecting a stated prioritization of structural recovery while still rewarding entropy-coherence (Golino, 14 Jan 2026).

The embedding formulation is therefore dynamic in a specific technical sense: the coordinate index replaces clock-time, and the researcher observes the trajectory of estimated graph structure over increasing depth. The principal methodological claim is that useful structural information is not uniformly distributed across coordinates.

3. Fit indices and optimization logic

The two fit indices in the reported DynEGA framework are TEFI and NMI, and they are explicitly treated as competing objectives. Given a weighted adjacency or partial-correlation matrix $j = 1 \ldots D$2, the density matrix is defined as

$j = 1 \ldots D$3

Von Neumann entropy is then

$j = 1 \ldots D$4

where $j = 1 \ldots D$5 are the eigenvalues of $j = 1 \ldots D$6. For a partition into $j = 1 \ldots D$7 communities with submatrices $j = 1 \ldots D$8, TEFI is given by

$j = 1 \ldots D$9

Its interpretation in the reported study is that lower TEFI, that is, more negative TEFI, indicates greater within-community coherence and less overall entropy (Golino, 14 Jan 2026).

NMI is defined between a true partition $e_i = (e_{i,1}, \ldots, e_{i,D})$0 and an estimated partition $e_i = (e_{i,1}, \ldots, e_{i,D})$1 by first computing the mutual information

$e_i = (e_{i,1}, \ldots, e_{i,D})$2

with entropies $e_i = (e_{i,1}, \ldots, e_{i,D})$3 and $e_i = (e_{i,1}, \ldots, e_{i,D})$4, and then

$e_i = (e_{i,1}, \ldots, e_{i,D})$5

Its range is reported as $e_i = (e_{i,1}, \ldots, e_{i,D})$6, with $e_i = (e_{i,1}, \ldots, e_{i,D})$7 indicating no agreement and $e_i = (e_{i,1}, \ldots, e_{i,D})$8 perfect recovery (Golino, 14 Jan 2026).

The central empirical observation is that TEFI and NMI “lead to competing optimization trajectories across the embedding landscape.” TEFI achieves minima at deep embedding ranges, specifically $e_i = (e_{i,1}, \ldots, e_{i,D})$9–$d$0 dimensions, where entropy-based organization is maximal but structural accuracy degrades, whereas NMI peaks at shallow depths where dimensional recovery is strongest but entropy-based fit remains suboptimal (Golino, 14 Jan 2026). The paper therefore states that “single-metric optimization produces structurally incoherent solutions,” and uses the weighted composite criterion to identify depth regions that jointly balance accuracy and organization.

A common misconception in this setting is that all embedding coordinates contribute comparably to dimensional recovery. The reported optimization landscape directly contradicts that assumption. A plausible implication is that DynEGA functions as a regularization strategy over representation depth, not merely as a graph-estimation procedure.

4. Monte Carlo evidence and depth-selection behavior

The reported Monte Carlo study constructs a pool of 200 AI-generated items for grandiose narcissism, equally across 5 dimensions, and embeds them using OpenAI’s text-embedding-3-small model with $d$1 (Golino, 14 Jan 2026). The number of items per dimension is varied as $d$2, yielding total sample size $d$3. Embedding depths are sampled as $d$4, that is, in increments of 5, and the procedure is repeated over Monte Carlo draws, with “500 iterations for vector-field aggregation.”

For each $d$5, the workflow estimates a TMFG + Walktrap network on the first $d$6 dimensions and computes $d$7 and $d$8. The resulting outcome curves exhibit systematic differences by metric. NMI “peaks very shallow (≈3–50 dims), then declines,” and its optimal depth is “largely insensitive to pool size once $d$9.” TEFI, by contrast, “achieves deep minima in large ranges (≈900–1 200 dims), with periodic troughs every ≈300 dims,” and its optimum moves deeper for larger $d$0 (Golino, 14 Jan 2026).

The composite optimum $d$1 lies in an intermediate region and scales approximately with item-pool size. The reported ranges are:

• small pools ($d$2): $d$3–$d$4;
• moderate pools ($d$5–$d$6): $d$7–$d$8;
• larger pools ($d$9): $E \in \mathbb{R}^{N \times D}$0–$E \in \mathbb{R}^{N \times D}$1, with plateauing thereafter.

The study further reports that composite-optimized DynEGA outperforms vanilla EGA on full embedding vectors, with the largest gains for $E \in \mathbb{R}^{N \times D}$2 (Golino, 14 Jan 2026). Within the boundaries of the reported simulation, the substantive significance is clear: optimal embedding depth scales with item pool size, and neither shallow nor deep regions alone are universally preferable.

5. Dynamic-graph workflow and higher-order events

A second DynEGA-oriented formulation arises from timestamped relational data modeled as a directed dynamic multigraph $E \in \mathbb{R}^{N \times D}$3, where each edge is an event

$E \in \mathbb{R}^{N \times D}$4

The construction allows multiple edges $E \in \mathbb{R}^{N \times D}$5 and assumes no self-loops (Ostroski et al., 2023).

The associated increment-weighted line graph $E \in \mathbb{R}^{N \times D}$6 uses original edges as nodes. A directed link $E \in \mathbb{R}^{N \times D}$7 exists if and only if there is tip-to-tail connectivity $E \in \mathbb{R}^{N \times D}$8 and temporal causality $E \in \mathbb{R}^{N \times D}$9, with weight

$L_i \in \{1 \ldots F\}$0

Smaller $L_i \in \{1 \ldots F\}$1 therefore implies stronger temporal coupling. Because the full line graph can be infeasible to construct, the method defines a sufficient subgraph or line-graph skeleton

$L_i \in \{1 \ldots F\}$2

with

$L_i \in \{1 \ldots F\}$3

and the summary states that this skeleton “exactly preserves all weight-thresholded connectivity in $L_i \in \{1 \ldots F\}$4” (Ostroski et al., 2023).

Edge clustering then proceeds by a hierarchical agglomerative scheme. The steps are: construct the skeleton in parallel, undirect $L_i \in \{1 \ldots F\}$5, compute its global MST, sort MST edges by ascending weight, build a single-linkage dendrogram via Union–Find, condense all trees smaller than the minimum cluster size $L_i \in \{1 \ldots F\}$6, and select clusters by maximizing “cluster stability” as in HDBSCAN. In the HDBSCAN-inspired formulation, the threshold variable is

$L_i \in \{1 \ldots F\}$7

and the stability of a cluster $L_i \in \{1 \ldots F\}$8 is

$L_i \in \{1 \ldots F\}$9

The output is a partition of edges into “conversations” plus a noise set (Ostroski et al., 2023).

Within the DynEGA workflow, each conversation $d_1 < d_2 < \cdots < d_m$0 induces a small dynamic subgraph $d_1 < d_2 < \cdots < d_m$1, and these subgraphs become the atomic units for downstream analysis. The reported uses include timeline-linked views, interactive brushing by cluster, force-directed sketches of $d_1 < d_2 < \cdots < d_m$2, and graph-level features $d_1 < d_2 < \cdots < d_m$3 such as clustering coefficient and entropy of inter-arrival times, which can then be embedded with UMAP or t-SNE and used to classify “anomalous” versus “routine” conversations (Ostroski et al., 2023).

6. Practical recommendations, limitations, and extensions

Several practical recommendations recur across the reported DynEGA settings. In the embedding application, embedding depth should be treated “as a tunable hyperparameter, not a fixed $d_1 < d_2 < \cdots < d_m$4,” and researchers are advised to perform a landscape search across depths, use a multi-objective criterion, normalize TEFI before combination with NMI, and sample depths in sufficiently fine increments, such as 5–10 dimensions, in order to capture TEFI oscillations (Golino, 14 Jan 2026). The same source recommends robust network estimation, notes that optimal depth profiles are model-specific, and states that embedding-based structures at $d_1 < d_2 < \cdots < d_m$5 should be validated against empirical networks once response data are collected.

The dynamic-graph workflow has its own parameter and modeling constraints. The minimum cluster size $d_1 < d_2 < \cdots < d_m$6 trades noise against fragmentation and is recommended to be chosen by an elbow-plot on number of clusters versus $d_1 < d_2 < \cdots < d_m$7. The number of dendrogram thresholds $d_1 < d_2 < \cdots < d_m$8 trades resolution against runtime, and non-uniform spacing is recommended, with more bins at small $d_1 < d_2 < \cdots < d_m$9. The method also “assumes ‘causal’ tip-to-tail is appropriate”; if interactions are broadcast-style, one may include tail-to-tip links or relax the tip-to-tail rule with a small negative-weight tolerance $w_{\mathrm{NMI}}$0 (Ostroski et al., 2023).

Scalability is an explicit concern rather than a peripheral implementation detail. Skeleton construction is reported as $w_{\mathrm{NMI}}$1 work and $w_{\mathrm{NMI}}$2 size, with overall wall-clock time $w_{\mathrm{NMI}}$3 for the skeleton and $w_{\mathrm{NMI}}$4 for the coarse dendrogram. The empirical example on 7.2 B Reddit edges reports a skeleton with 12.8 B links and timings of approximately 9 s for skeleton construction and approximately 220 s for the dendrogram on 256 nodes (Ostroski et al., 2023). Very high-degree vertices, however, can still dominate skeleton size; the stated recommendations are pre-filtering known bots or applying degree-based down-sampling.

Extensions reported in the dynamic-graph summary include generalized weights,

$w_{\mathrm{NMI}}$5

dynamic hypergraphs with events $w_{\mathrm{NMI}}$6, and integration of unstructured data on vertices and edges (Ostroski et al., 2023). A plausible implication is that DynEGA is adaptable whenever a meaningful ordering variable and a graph-construction rule can be specified. The reported literature, however, simultaneously underscores that the resulting optimization landscape is domain-specific, metric-sensitive, and parameter-dependent; DynEGA does not eliminate model selection, but relocates it to the level of the ordered representation itself.