Signed, Non-Dissipative Memory for Graphs Whose Homophily Drifts Over Time

Most graph neural networks, models that update each node by averaging information from its neighbors, quietly assume that linked nodes are alike. This is the homophily assumption. When it holds, repeated averaging helps. When it fails (heterophily: fraud rings, transaction networks, word-adjacency graphs), averaging blurs nodes into their unlike neighbors. Stacking many averaging steps makes every node look the same, which is called over-smoothing. Two separate literatures have made progress here, but on different problems. One handles heterophily and over-smoothing on fixed graphs, using signed messages that can push dissimilar nodes apart. The other handles graphs that change over time, using continuous-time updates built so that information neither fades nor explodes. Neither asks what happens when a stream of timestamped edges changes character midway, for example from mostly-alike endpoints to mostly-unlike ones. Neither asks whether a single node memory can then still serve both node-level and link-level predictions.

We propose to test one specific claim. In a continuous-time event stream, a node-memory update that is (a) non-dissipative in its self-dynamics and (b) signed per event can keep one shared representation useful for both link prediction and node property prediction as edge homophily drifts. Non-dissipative means it neither shrinks nor inflates the state. Signed per event means each arriving edge decides whether to pull its endpoints together or push them apart, based on how similar they are at arrival time. Models that lack either ingredient should fail in different, diagnosable ways. Without the sign, the memory collapses toward agreement and link scoring degrades. Without non-dissipation, long-range information is lost. Training on one task alone should then transfer poorly to the other. The deliverable has three parts: a controlled drifting-homophily stream benchmark, the update rule itself, and a decisive set of ablations and probes. The probes include an honest comparison against the cheap heuristics that already beat deep temporal models on some benchmarks.

Core Claim

Over-smoothing and heterophily are both statements about how similar the representations of linked nodes become. In a static graph that similarity is fixed by the data. In a stream it is a moving target. We hypothesize that the "valley of death" documented for static link prediction also appears over time in streams. That valley is a region where positive and negative node pairs have overlapping feature similarity and simple decoders fail. A model's memory degrades exactly in the time windows whose incoming edges fall into that overlapping regime. Unsigned continuous-time encoders cannot recover there, because every event moves endpoints toward each other. The novelty is not combining "anti-symmetric" with "signed" as a design flourish. It is the testable prediction that the per-event sign is what separates link-and-node co-usability from negative transfer under drift, and that this effect disappears when homophily does not drift.

What Is Known and What Is Missing

Three static works motivate the mechanism. Two Sides of the Same Coin shows that over-smoothing behaves like "pseudo-heterophily": misclassified nodes start acting as unlike neighbors. Its model, GGCN, fixes this with signed, cosine-similarity edge weights and degree corrections. It is transductive, node-classification-only, and static. Link Prediction under Heterophily: A Physics-Inspired GNN (GRAFF-LP) learns attract-or-repel gradient flows for link prediction. It reports 99.34% AUROC on Roman Empire against roughly 56% for GCN, again only on static graphs and only for links. On the Impact of Feature Heterophily on Link Prediction defines feature heterophily through the cosine-similarity distributions of connected versus unconnected pairs. It finds a U-shaped curve with a failing "gated" middle, where non-linear MLP decoders are needed. It explicitly leaves dynamic topologies and depth untouched. We reuse its similarity distributions as our drift measure and its MLP decoder as our default link head.

Two works supply the dynamical-systems backbone. Anti-Symmetric DGN makes layers discretized ordinary differential equations with weights $W - W^\top$, whose purely imaginary eigenvalues keep signals from dissipating across 30+ layers. Long Range Propagation on Continuous-Time Dynamic Graphs (CTAN) carries this to timestamped event streams. It was evaluated on Wikipedia, Reddit and LastFM, which the file describes as homophilic, and it never varies heterophily. Our model is CTAN with one change: a learned, bounded, per-event sign on the neighbor term. CTAN is therefore the main ablation, not a straw man.

Three works constrain the design. Multi-Track Message Passing prevents mixing by routing nodes onto per-class propagation tracks. That costs memory proportional to the number of tracks, depends on pseudo-labels that would shift in a stream, and gives link decoders no single node vector. We avoid this design for those reasons. The dynamic-graph literature notes that precomputed propagation, such as Node Dependent Local Smoothing, is invalidated by every edge change, which argues for an event-local update. Finally, Are we measuring oversmoothing correctly? warns that similarity-based collapse metrics can disagree with functional degradation. We therefore measure collapse functionally, with probes, and report similarity metrics only alongside.

The strongest threat comes from Temporal Graph Benchmark. On tgbn-reddit, a Moving Average heuristic scores NDCG@10 of 0.559 against 0.315 for TGN. Our proposal must clear these heuristics, not just other neural models.

Research Program

The update rule is applied when an event $(u, v, t)$ arrives. Let $h_u$ be node $u$'s memory, $x_u$ its features, and $\Delta t$ the time since $u$'s last event. The update is

$$h_u \leftarrow h_u + \epsilon\, \sigma\!\big((W - W^\top - \gamma I)\, h_u + s_{uv}\, V h_v + U x_u\big), \qquad s_{uv} = \tanh\!\big(a^\top [\,x_u \odot x_v,\ |x_u - x_v|,\ \phi(\Delta t)\,]\big).$$

Here $\epsilon$ is the step size, $\sigma$ a pointwise non-linearity, and $\gamma$ a small damping constant as in CTAN. $V$ and $U$ are learned maps, $\phi$ a time encoding, and $a$ a learned vector. The anti-symmetric part keeps the node's own dynamics non-dissipative. The coefficient $s_{uv} \in (-1, 1)$ lets an event repel its endpoints when they are dissimilar, and the tanh bound is what keeps repulsion from blowing up. A single memory feeds two heads: an MLP link decoder, following Zhu et al., and a node-affinity head.

The first work package is a benchmark. We build synthetic event streams in which node classes and features are fixed. The edge-generation rule moves on a schedule from homophilic, through gated (positive and negative pairs with overlapping cosine similarity), to heterophilic, and optionally back. The schedule is parameterized so we control both drift speed and where in the stream the gated window falls. We also stream static heterophilic graphs (Roman Empire, Amazon Ratings) as event sequences in a fixed arrival order. Real streams from TGB (tgbl-wiki, tgbl-coin, tgbn-reddit) and the GraphPulse blockchain networks used by TRicci are profiled window by window with Zhu et al.'s positive/negative similarity distributions. The point is to learn whether real drift exists at all. If no real dataset shows meaningful drift, we will say the claim is confined to synthetic and constructed streams.

The second work package is analysis. For each model and time window we track three things: link MRR or AUROC, a frozen linear probe's node-class accuracy on the memory, and GRAFF-LP's Gradient Separability score. Together these give a functional collapse curve over time, to be read next to the drift curve. The third work package is the cross-task test. Each encoder is trained on link prediction only, node prediction only, and jointly. The memory is then frozen and the other head is retrained. The gap between the single-task-then-probe score and the joint score is our measure of negative transfer.

Baselines

The naive reference points are TGB's non-learning heuristics. EdgeBank, which predicts that a link recurs if it has been seen, serves for links. Moving Average and Persistence Forecast serve for node affinity. They run in the same streaming protocol with zero training budget. The reported numbers are NDCG@10 0.559 on tgbn-reddit for Moving Average versus 0.315 for TGN (TGB). Beating them on tgbn-reddit is not required, because we expect drift there to be weak. Beating them on drifting streams is required.

The strongest neural link baseline is NAT, which reports MRR 0.749 on tgbl-wiki in TGB. We reproduce it under TGB's negative-sampling protocol. Matching it on tgbl-wiki while exceeding it on drifting streams counts as success. On static heterophilic links the reference is GRAFF-LP's 99.34% AUROC on Roman Empire. On streamed Roman Empire we report how far the streaming setting falls below that ceiling, rather than claiming to beat it. TGN and TGAT are run as standard memory-based and attention-based temporal models. All neural models get the same embedding size, the same MLP link decoder (so decoder choice does not confound results), and the same hyperparameter-search budget measured in trials and GPU-hours.

The decisive ablations isolate the ingredients. CTAN, which is our model with $s_{uv}$ fixed to $+1$, isolates the sign. A signed version with an unconstrained $W$ in place of $W - W^\top$ isolates non-dissipation. A variant whose sign depends only on $\phi(\Delta t)$ and not on features tests whether feature similarity is the operative signal. Success means the full model beats CTAN specifically in gated and heterophilic windows, ties it in homophilic windows, and shows a smaller cross-task transfer gap.

Why not just…?

Why not just use EdgeBank or Moving Average? TGB shows these heuristics can beat deep temporal models, and they are a genuine threat on repetitive, homophilic streams. Heuristics carry no node features, however, so they cannot respond to a change in which kinds of nodes start linking. Our synthetic streams generate new edges after the drift point by rule, not repetition, precisely to separate the two. If the heuristics still match us on the profiled real streams, we will report that the mechanism matters only where drift and novelty co-occur.

Why not just freeze a shallow encoder like SIGN or GraphSAGE and train per-task MLP heads? Zhu et al. report that an MLP decoder gives a 50% boost on heterophilic link tasks and that SIGN beats GCN by up to 44%. Shallow ego-neighbor separation may well suffice on static data. Two things count against it here. A frozen static encoder cannot absorb post-drift structure without the re-computation that the dynamic-graph literature flags as prohibitive. A shallow encoder also cannot carry long-range temporal information, which is CTAN's demonstrated strength. We run exactly this baseline, re-fitted periodically under the same compute budget. If it ties the full model, the sign is not earning its keep and we will say so.

Why not just run GGCN on discrete snapshots? GGCN already supplies signed, degree-corrected messages and stays stable at depth on static graphs, so a snapshot version is the most direct competitor to the sign ingredient. Snapshot binning, though, averages over events within a window. When the gated regime is short, that averaging should blur exactly the transition we study. We run snapshot GGCN at several window widths. If it matches us at fine widths with comparable cost, the continuous-time component is unnecessary and the contribution reduces to the drift benchmark and analysis.

Proposed Experiments and Decisive Checks

The central experiment plots, on the synthetic drifting streams, link MRR and frozen node-probe accuracy over time for every model, with the gated window marked. The hypothesis is falsified if CTAN does not dip in the gated and heterophilic windows, or if the signed variant dips equally. A second check runs the same models on control streams whose homophily is constant. There the sign should give no advantage, so an advantage on controls would indicate the gain comes from extra capacity rather than drift handling. A third check is the cross-task transfer gap: train on links, freeze, probe nodes, and the reverse. We predict the gap is smallest for the full model and largest for unsigned models after heterophilic drift. Finally, we report the learned $s_{uv}$ against true endpoint label agreement. If the sign does not track same-class versus cross-class events, the mechanism is not doing what we claim, even if aggregate scores improve. Real-stream results on TGB and GraphPulse are reported last, scoped by the measured drift on each dataset.