acceptodds
Under review as a conference paper at ICLR 2027

CTDG-THOR: Query-Side Second-Order Temporal History Routing for Continuous-Time Dynamic Graphs

Abstract

Events in a continuous-time dynamic graph (CTDG) arrive as a stream of timestamps and messages. Most CTDG encoders build a node representation from that node’s own events, so a neighbor’s trajectory reaches a link score only when the model routes it across nodes. We present CTDG-THOR, which isolates candidate-independent routing of second-order temporal history from the query side. The model combines a TGN-style memory, history-conditioned attention, and pair-level aggregation of first-hop contexts for the query’s neighbors. It expands second-order context only for the query, so the negative candidate set does not multiply path computation. We formalize an incident-only endpoint and a one-hop endpoint, and prove a witness separation: exponentially many event streams that the incident-only endpoint cannot distinguish become separable once second-order history is routed. Under matched seeds and a shared backbone, CTDG-THOR improves test MRR by +0.0653 on tgbl-wiki and +0.0368 on tgbl-enron. Fixed-checkpoint interventions also indicate that the trained predictor uses this route. Under the fixed-negative TGB-v1 protocol, it reaches test MRR 0.6416 ± 0.0123, 0.2478 ± 0.0255, and 0.1223 ± 0.0084 on these three datasets. The second-order route adds one scalar and a mean aggregation over the one-hop control and keeps peak memory below 90 MB. The comparisons show that where cross-node history is routed affects prediction in memory-based CTDGs.

Primary area:
learning on time series and dynamical systems
Keywords:
Continuous-time dynamic graphs, temporal link prediction, second-order routing, history routing, TGB

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_efw7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents CTDG-THOR, a memory-based CTDG model with second-hop neighbor aggregation. It begins by motivating the need for neighbor context in learning node embeddings for a CTDG and proposes a framework to implement it. The authors then analyze its performance against a model without second-hop neighbor aggregations using the TGBv1 protocol. Furthermore, the paper provides theoretical analysis comparing the framework with and without the proposed aggregation.

Critical strengths

The proposed work presents an efficient CTDGs model based on second-hop neighbor aggregation with TGN-like memory module, improving upon the overall architecture with multi-hop neighbor context.

Critical weaknesses

W1. Clarity and Presentation

The manuscript is currently challenging to follow due to several advanced terms—such as history-conditioned attention, second-order temporal history routing, and separation witness—being introduced and heavily used without clear prior definitions. Additionally, some terms require clarification, such as "Direction-A context" on line 276. There are also minor notation inconsistencies, such as the subscript for the weight matrix W\textbf W appearing as kk in Eq. 6 and vv in Eq. 9.Furthermore, the proofs would benefit from more detailed steps and a cleaner structural layout, as formatting parts of the proofs as individual subsections currently disrupts the flow of the theoretical argument.


W2. Message Definition and Theoretical Alignment

The paper adopts a message-passing block inspired by TGN, but it omits a key component. In standard TGN formulations, the message for an event (u,v,t,e)(u,v,t,e) is defined as msg(su(t−),sv(t−),Δt,e)\text{msg}(s_u(t^-), s_v(t^-),\Delta t, e), incorporating the state of the destination node svs_v. The CTDG-THOR framework omits this destination state term, which appears to limit the expressiveness of the resulting memory module. Consequently, the subsequent theoretical analysis builds upon this modified (and more limited) formulation. While removing svs_v yields a valid memory module variant, it departs from the standard TGN memory design that CTDG-THOR claims to follow. We encourage the authors to clarify how Theorems A and B generalize to or provide insight into standard CTDG memory modules that incorporate destination states.


W3. Evaluation and Baseline Consistency

The experiments rely on TGBv1 without explicit justification for omitting the more recent TGBv2 benchmark standard. Moreover, there are noticeable discrepancies between the reported results and the official TGB leaderboard figures. For instance, the authors report an MRR of 0.2931±0.01030.2931 \pm 0.0103 for tgb-wiki-v1, whereas leaderboard benchmarks or related figures differ significantly. Additionally, the finding that a "TGAT-style" model outperforms DyGFormer contrasts sharply with public TGB leaderboard trends (where DyGFormer generally outperforms TGAT). We recommend that the authors double-check their baseline implementations, hyperparameter tunings, and evaluation scripts to ensure these gaps can be addressed.


W4. Limited Novelty

The primary novelty of the work lies in assuming a weak or limited memory module and proposing a way to circumvent this issue. However, existing models like CTAN and TGN (which incorporate state update and memory modules) do not restrict updates solely to the event node itself. Given these standard memory update practices, the paper appears to over-claim its novelty. Furthermore, the idea of incorporating multi-hop neighbors is already part of TGN's node embedding generation module, which uses recursive neighbor sampling and attention weights to collect multi-hop context for a query.

Questions for authors

The above-mentioned weaknesses and

Q1 Failure for LRD?

Is the context strictly limited to second-hop neighbors? If so, wouldn't this framework struggle in long-range dependency setups presented in CTAN, where relevant information resides 10-hops or more away (assuming the memory message is derived using Equation 4)?

Reviewer_m7JLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

To better use the cross-node history, the authors propose CTDG-THOR, which improve the training quality by isolating candidate-independent routing of second-order temporal history from the query side. Experiments show that CTDG-THOR improves the model quality.

Critical strengths

S1. Experiments demonstrate the effectiveness of the method.

S2. The authors provide some theoretical analysis.

Critical weaknesses

W1. The presentation of this work needs needs significant improvement. For example, in the introduction, the key challenges of using cross-node history are not clearly explained. The discussion of the trade-offs is also somewhat generic. Ideally, the introduction should provide a more detailed and quantitative experimental analysis to support these claims.

W2. The model-level novelty is not clearly established. The proposed model largely follows a straightforward TGN + attention design. It remains unclear whether the authors’ ideas are effective compared to more recent T-GNN architectures, such as TIDFormer [1].

Moreover, it would be helpful to clarify what types of models would better align with the proposed routing method.

[1] Tidformer: Exploiting temporal and interactive dynamics makes a great dynamic graph transformer. KDD 25

W3. The theoretical analysis in Section 5 needs to be improved. It is difficult to follow how different theorems relate to one another. It is also unclear how each theorem supports the validity and effectiveness of the proposed method.

Questions for authors

Q1. What are the key challenges of using cross-node history? (W1)

Q2. Could the authors clearly specify the model-level novelty of the proposed method? (W2)

Q3. Can the authors clarify the logical relationship among the theorems in Section 5? (W3)

Reviewer_4QqpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

CTDG-THOR extends a TGN-style memory model with a second-order context on the query node. The query endpoint averages its neighbors first-hop attention context, this context is then included in a adaptive fusion gate. Only the query receives this attention context, so it is computed computed once per query and reused across all negative candidates. On the TGB-v1 comparisons this additional second-order context provides significant gains.

Critical strengths
  • There is a clear improvement using candidate-independent query-side second-order temporal history routing. Most improvements are recorded as significant with p-tests.
  • The added cost is shown to be small. It adds only 0.8s per UCI epoch against the memory only backbone method.

This provides a critical strength to this method as we have significant improvements on link-prediction with cheap access to it.

Critical weaknesses
  • Why is v2 not used? "We use v1 protocol because it provide a stable, matched-seed setting for controlled route comparisons" seems a bit vague as a reason to not compare against baselines on updated datasets.
  • Nitpicked baselines?
    • There doesn't seem to be an explanation for why some models were chosen over others given this leaderboard of v2 we can clearly see better performing models. Why have they not been compared against? Is this method more efficient? Are we trying to showcase only against memory-based methods? There are some memory-free models that are benchmarked.
    • The related work section discusses models that are never evaluated. CAW, NAT, TCL, etc which are discussed as using routes directly are not used as baselines. Again it does not seem clear as to why some are chosen to be bench marked while others (which clearly do very well on the v2 leader board) are ignored.
Questions for authors

Why is v2 not used? The v2 ships with fixed negative files. Why can v2 not be used for "stable matched seed comparisons"?

Why have other models not been included in the baselines? Are we making an efficiency argument or comparing with only memory-based methods?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.