Prospective Graphing: Dynamic Future Methods

• Prospective graphing is a family of techniques that build or forecast graph structures by incorporating current or future state data to reflect temporal interdependencies.
• It spans methods from dynamic graph construction for time series anomaly detection to graph-supported forecasting using regression in embedded spaces.
• Applications include real-time anomaly detection, anticipatory human-object interaction prediction, vulnerability triage via temporal evidence graphs, and survival analysis visualization.

Searching arXiv for papers using or discussing “prospective graphing” and closely related graph-based forecasting/temporal-evidence concepts. arxiv_search(query="^^^^^^^^3all:\3 graphing\"^^^^^^^^3 OR ti:\3^^^^^^^^"prospective graphing\"^^^^^^^^3 OR abs:\3^^^^^^^^"prospective graphing\"", max_results=^^^^3 OR ti:\3all:\3^^^^) Prospective graphing is a heterogeneous forward-looking use of graph-based representation rather than a single standardized method. In current arXiv usage, the phrase and closely related constructions denote several distinct practices: target-time-aware dynamic graph construction for forecasting-based multivariate time series anomaly detection; regression or autoregression over graph-supported states to forecast future graph behavior; graph matching from partial observations to infer forthcoming human-object interactions; decision-time-sliced temporal evidence graphs for vulnerability triage; and visualization of survival surfaces under hypothetical interventions on a continuous variable (&&&^^^^3all:\3^^^^&&&, &&&^^^^3 OR ti:\3^^^^&&&, &&&^^^^3 OR abs:\3^^^^&&&, Alpay et al., 17 Jun 2026, Denz et al., 2022). This suggests that prospective graphing is best treated as a family of techniques in which graph structure is assembled, filtered, or interpreted with explicit reference to future states, target-time relations, or prospective evaluation.

^^^^3 OR ti:\3^^^^. Definitional regimes

The most explicit technical use of the term appears in the PMGC framework for multivariate time series anomaly detection. There, prospective graphing is a graph-construction strategy in which the dynamic inter-series graphs are built from the full sample PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^, including the current or prediction-time observations, while the actual forecasted values are still produced from the historical context only. The paper is explicit that “prospective” does not mean forecasting the graph first; it means constructing the graph with access to the current target-time observations so that the graph reflects the concurrent inter-series relationships at the time being evaluated (&&&^^^^3all:\3^^^^&&&).

A second regime uses graphs prospectively by forecasting future graph states or future graph-supported signals. One line of work formulates graph time series prediction as regression in dissimilarity or kernel spaces, with PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^ predicted from prior graph states and the output usually represented as a point in an embedded pseudo-Euclidean or kernel space rather than as an explicit reconstructed graph (&&&^^^^3 OR ti:\3^^^^&&&). Another line keeps the graph fixed but forecasts vector-valued node states through product graphs that jointly model dependencies across nodes and across feature components (Natali et al., 2020).

A third regime uses graphs to support prospective inference from incomplete or time-restricted evidence. In human-object interaction prediction, a graph built from a partially observed video is matched against graphs of fully observed reference videos to infer the ongoing activity, the next active object, and the time of the forthcoming interaction (&&&^^^^3 OR abs:\3^^^^&&&). In vulnerability triage, exploitability evidence is represented as a temporal evidence graph, but only the portion visible by a fixed decision time is admissible, so the operational graph is a time-filtered subgraph rather than a retrospective all-time archive (Alpay et al., 17 Jun 2026).

A fourth regime is prospective in an interventional or visualization sense. For time-to-event outcomes with a continuous covariate, the survival area plot displays the surface PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3^^^^, estimated by g-computation, so that the viewer can read survival prospectively over follow-up time for each hypothetical value $z$ of the continuous variable (Denz et al., 2022). The common element across these regimes is temporal asymmetry: graph construction or interpretation depends on what is supposed to be known before the future outcome is realized.

^^^^3 OR abs:\3^^^^. Target-time-aware dynamic graph construction in PMGC

In PMGC, prospective graphing is introduced to address a specific weakness of ordinary forecasting-based TSAD. Standard forecasting uses a past window to predict future values and marks large prediction errors as anomalies, but the paper argues that this can fail when the target time contains unpredictable but normal concurrent variations. The difficulty is not only that each sensor’s future value is hard to predict; the cross-sensor dependency structure itself can change at the current step. If the graph is derived only from historical context, the GNN aggregates the wrong neighbor information and can produce inaccurate predictions for non-anomalous behavior (&&&^^^^3all:\3^^^^&&&).

Operationally, each sample $X^t$ is split into a context window

$C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$

and a prediction window

$y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$

The crucial asymmetry is that $C^t$ is used for node features and forecasting, whereas the full sample $X^t$ is used to generate the dynamic graphs. Prospective graphing therefore constructs a set of short-term, instance-wise dynamic adjacency matrices

$$\{A^{t,i}\}_{i=1}^k,\qquad A^{t,i}\in\mathbb{R}^{N\times N}.$$

A two-layer fully connected network first extracts a representation PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3all:\3^^^^, splits it into PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3 OR ti:\3^^^^ with PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3 OR abs:\3^^^^, and then applies a graph generator PRESERVED_PLACEHOLDER_^^^^3 OR ti:\33^^^^ so that PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^4. The graph generator uses pairwise cosine similarity followed by ReLU,

PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^5

so each graph encodes pairwise channel similarity derived from the sample representation.

The forecasting stage then uses context-derived node features under these target-time-informed graphs: PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^6 The graph convolution follows MixHop-style propagation,

PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^7

The paper emphasizes that this does not turn the model into a reconstruction method and does not directly predict from the current values. Current values influence graph topology or weights, not the node feature content being forecast from. That distinction is central to the PMGC meaning of prospective graphing.

3. Static structure, graph cohesion, and anomaly scoring

Prospective graphing in PMGC is coupled to a second structural device: a learnable long-term static graph

PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^8

where PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^9 contains learnable node embeddings. The dynamic graphs are sample-specific and target-aware, but they are regularized by a graph cohesion loss so that they do not degenerate into arbitrary sample noise. The paper first introduces a simplified loss that sums PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3all:\3^^^^ and notes that this would collapse all dynamic graphs onto the static graph. It then uses an improved cohesion objective based on

PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3 OR ti:\3^^^^

so that each dynamic graph is encouraged to align with PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3 OR abs:\3^^^^ while remaining dissimilar to the other dynamic graphs (&&&^^^^3all:\3^^^^&&&).

The theory in the paper is about this cohesion term rather than prospective graphing directly, but it is structurally important. The simple loss yields the trivial optimum PRESERVED_PLACEHOLDER_^^^^3 OR abs:\33^^^^, whereas the improved objective is shown not to be minimized either when all dynamic graphs equal the static graph or when all dynamic graphs are identical to one another. The intended balance is therefore adaptivity to current conditions, stability with respect to long-term structure, and diversity across the PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3^^^^4 dynamic graphs.

Training combines prediction and graph regularization: PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3^^^^5 After training on normal data, anomaly scores are derived from normalized forecast errors,

PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3^^^^6

with PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3^^^^7 the median and PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3^^^^8 the IQR of the error distribution for channel PRESERVED_PLACEHOLDER_^^^^3 OR abs:\3^^^^9.

Empirically, the paper reports that removing prospective graphing causes a significant performance drop, which it interprets as evidence that real-time information in graph construction is important because it allows the graphs to reflect current interactions among time series during normal operation. Across SWaT, WADI, MSL, SMAP, and HAI, PMGC reports the best average performance among baselines; the reported examples are SWaT $z$^^^^3all:\3^^^^, ROC $z$^^^^3 OR ti:\3^^^^; WADI $z$^^^^3 OR abs:\3^^^^, ROC $z$3; SMAP $z$4, ROC $z$5; and HAI PR $z$6, ROC $z$7. The paper also notes that performance peaks around prediction window $z$8, suggesting that incorporating more future steps better utilizes dynamic graphs that reflect the current inter-series relationship.

4. Forecasting future graph states and graph-supported processes

One major adjacent meaning of prospective graphing is whole-graph or graph-supported forecasting. In "Time Series Prediction for Graphs in Kernel and Dissimilarity Spaces" (&&&^^^^3 OR ti:\3^^^^&&&), the object of prediction is the next graph state as a whole. A time-varying graph is written

$z$9

with temporal subgraphs $X^t$^^^^3all:\3^^^^, and the forecasting task is to learn a mapping

$X^t$^^^^3 OR ti:\3^^^^

The method uses graph dissimilarities or graph kernels so that regressors such as ^^^^3 OR ti:\3^^^^-nearest neighbor, kernel regression, Gaussian process regression, and rBCM can operate on graphs. The important caveat is that the output of kernel regression or Gaussian process regression is generally not an explicit future graph; it is a point in an embedded pseudo-Euclidean or kernel space. This makes the method suitable for downstream comparison, clustering, or classification of anticipated future graph configurations, but not a complete solution to graph pre-image reconstruction. On synthetic graph-evolution models, kernel regression is often sufficient, whereas on the real intelligent-tutoring datasets Gaussian process regression improves the prediction error.

A related but distinct formulation appears in "Forecasting Multi-Dimensional Processes over Graphs" (Natali et al., 2020). There the graph topology is fixed, and what is forecasted is a vector-valued node state. The method introduces a feature graph $X^t$^^^^3 OR abs:\3^^^^ and combines it with the node graph $X^t$3 through a product graph $X^t$4, for the Cartesian product

$X^t$5

The resulting PG-VAR model is

$X^t$6

and the generalized version uses Kronecker-structured terms $X^t$7. The parameter count is independent of the number of nodes $X^t$8 and feature dimension $X^t$9, and for Cartesian product graphs the complexity is stated as

$C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$^^^^3all:\3^^^^

The paper is explicit that this is forecasting future graph-supported vector states rather than forecasting topology itself. Its proof-of-concept experiment uses a walking dog mesh with $C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$^^^^3 OR ti:\3^^^^, $C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$^^^^3 OR abs:\3^^^^, and $C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$3, and finds that PG-VAR is at least comparable to scalar G-VAR in a low-data setting and improves when more training samples are available.

5. Anticipatory graph matching and temporal evidence graphs

In human-object interaction prediction, prospective graphing takes the form of graph-based anticipation from partial observation. "Graphing the Future" (&&&^^^^3 OR abs:\3^^^^&&&) represents a video prefix as a fully connected, undirected, attributed graph $C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$4 whose nodes are upper-body joints and scene objects. Edge weights combine motion and semantic dissimilarity,

$C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$5

with semantic dissimilarity

$C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$6

from WordNet/Wu-Palmer similarity, and motion dissimilarity

$C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$7

A partially observed test graph $C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$8 is matched against complete reference graphs $C^t=\{X_{\cdot i}\}_{i=(t-w+1)}^{(t-p)}$9 by Bipartite Graph Edit Distance, using node and edge insertion, deletion, and substitution costs and the Kuhn-Munkres algorithm. The method predicts the ongoing activity, the class of the next active object, and the time at which that object will be used. On CAD-^^^^3 OR ti:\3 OR abs:\3all:\3^^^^, the reported single-NAO accuracy rises from $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$^^^^3all:\3^^^^ at $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$^^^^3 OR ti:\3^^^^ s and $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$^^^^3 OR abs:\3^^^^ s before the next action to $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$3 at $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$4 s; the normalized time-to-NAO error falls from $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$5 to $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$6; and multiple-NAO accuracy rises from $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$7 at $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$8 observation to $y^t=\{X_{\cdot i}\}_{i=(t-p+1)}^t.$9 at $C^t$^^^^3all:\3^^^^. The paper therefore treats the graph not as a static descriptor but as the mechanism through which a partially observed present is matched to a plausible future continuation.

A stricter decision-theoretic use appears in "Compute-Budgeted Exploitability Evidence Graphs for Prospective Vulnerability Triage" (Alpay et al., 17 Jun 2026). There the temporal evidence graph is

$C^t$^^^^3 OR ti:\3^^^^

with CVE nodes, public evidence document nodes, and CWE/product nodes. The core prospective primitive is the per-CVE decision time

$C^t$^^^^3 OR abs:\3^^^^

which induces the admissible evidence set

$C^t$3

with undated evidence excluded. Evidence selection is compute-budgeted: $C^t$4 Each score is paired with an auditable certificate containing the CVE identifier, risk, rank, decision time, budget, and for each selected evidence item its source layer, provenance string, timestamp, score, and a leakage flag. On $C^t$5 prospective CVEs, budgeted evidence selection raises leakage-safe prospective recall@5^^^^3all:\3^^^^ from $C^t$6 for a severity-only baseline to $C^t$7; two evidence documents per CVE capture most of the value; a strong cross-encoder reranker lowers prospective recall to $C^t$8; and a naive random split with unfiltered evidence inflates apparent prospective recall by $C^t$9 and EPSS-high recall by $X^t$^^^^3all:\3^^^^. Here prospective graphing is not richer message passing but time-respecting evidence accumulation under a fixed decision boundary.

6. Prospective visualization and graph construction from text

Prospective graphing also includes forward-looking visualization of counterfactual or standardized processes. "Visualizing the (Causal) Effect of a Continuous Variable on a Time-To-Event Outcome" (Denz et al., 2022) addresses the case where a continuous variable $X^t$^^^^3 OR ti:\3^^^^ affects survival over time. The target estimand is

$X^t$^^^^3 OR abs:\3^^^^

and the g-computation estimator is

$X^t$3

The survival area plot uses time on the $X^t$4-axis, survival probability on the $X^t$5-axis, and color to encode $X^t$6, so that each “infinitely thin” slice of the area corresponds to one specific estimate of $X^t$7. The paper presents continuous and binned color versions and related alternatives such as contour plots, heatmaps, value-specific survival curves, landmark survival plots, quantile plots, and RMST plots. It is explicit that causal interpretation requires no interference, consistency, conditional exchangeability, positivity, and correct model specification. In the getABI example, contour plots with B-splines suggest a nonlinear relationship in which survival improves as ABI increases up to about $X^t$8, then appears to decrease again for ABI values above roughly $X^t$9.

A different extension of graphing appears in social history. "Granularity in Action: Graphing sources for social history" (&&&^^^^3 OR ti:\3^^^^8&&&) proposes a pipeline for turning historical sources into structured data organized around action as the basic and constitutive unit of analysis. The GRAM framework uses three node types—Actions, Actors, and Places—and core edge types such as NEXT, TAKES_PLACE, HAS_ORIGIN, HAS_DESTINATION, HAS_AGENT, HAS_PATIENT, and HAS_INSTRUMENT. The pipeline proceeds from OCR or HTR and source segmentation, through relevance classification with SetFit, to verb detection with HERMOD, and finally to LLM-based extraction orchestrated with DSPy. The first extraction prompt produces fields such as Subject, Title, Verb, Object, Infinitive, and Snippet; a second prompt extracts location. The output is a skeletal graphing or “auto-GRAMS” representation intended to integrate with close readings and extensive manual graphing. Although this work is not future-oriented in the PMGC sense, it broadens graphing into a modular data-production process that supports later comparative and exploratory analysis.

7. Conceptual boundaries and related but distinct uses of graphing

The heterogeneity of the literature produces several recurrent misconceptions. Prospective graphing does not uniformly mean future graph prediction. In PMGC, it means constructing dynamic graphs with access to current target-time observations while forecasting still depends on historical context, and the paper is explicit that the method does not become a reconstruction model and does not directly feed current values as node features into the predictor (&&&^^^^3all:\3^^^^&&&). In PG-VAR, the graph is fixed and only the node states are forecasted (Natali et al., 2020). In kernel- or dissimilarity-based graph time series prediction, many methods output an embedded point rather than an explicit graph (&&&^^^^3 OR ti:\3^^^^&&&). In survival visualization, prospective interpretation can be causal only under standard identifiability assumptions (Denz et al., 2022). In vulnerability triage, the central issue is not graph expressiveness alone but leakage-safe temporal admissibility (Alpay et al., 17 Jun 2026).

The term is also bounded by other uses of graphing that are structurally important but technically different. In topology, Koytcheff’s graphing map

$$\{A^{t,i}\}_{i=1}^k,\qquad A^{t,i}\in\mathbb{R}^{N\times N}.$$^^^^3all:\3^^^^

turns a family of embeddings into a single embedding in one higher source and target dimension. There graphing is a geometric operation that increases both source and target dimensions by one, and it functions as the central device for transporting homotopy information between spaces of long links and knots (&&&^^^^3 OR abs:\3^^^^4&&&). In graphene defect classification, graphing and grafting means cutting out a patch of graphene and replacing it with another patch with the same number of dangling bonds; in the dual triangular lattice the ring size along the boundary is determined by

$$\{A^{t,i}\}_{i=1}^k,\qquad A^{t,i}\in\mathbb{R}^{N\times N}.$$^^^^3 OR ti:\3^^^^

and alternating $$\{A^{t,i}\}_{i=1}^k,\qquad A^{t,i}\in\mathbb{R}^{N\times N}.$$^^^^3 OR abs:\3^^^^ produces a closed loop of alternating 5- and 7-membered rings (&&&^^^^3 OR abs:\3^^^^5&&&). These are genuine graphing operations, but they are not target-time-aware predictive graphs.

A plausible implication is that prospective graphing is best understood not as a single algorithmic family but as an organizing perspective. Across anomaly detection, graph forecasting, anticipation, vulnerability triage, survival analysis, and structured source extraction, the graph is made prospective by one of three mechanisms: it is built with respect to a target-time state, sliced by a decision-time visibility constraint, or used to encode a future-oriented surface or continuation. What varies is whether the graph represents current inter-series relations, an evolving graph state, retrieved analogies from partial observations, temporally admissible evidence, or structured action descriptions. The unifying feature is temporal directionality in graph construction or interpretation, not a shared model class.