acceptodds
Under review as a conference paper at ICLR 2027

TG-Any: Train Once, Predict Links on Unseen Temporal Graphs

Abstract

A new temporal graph should not require a new predictor. We introduce TG-Any, a train-once approach to temporal link prediction through a shared interface of 24 streaming behavioral scores. An 18,584-parameter network learns query-dependent, nonnegative fusion on tgbl-wiki; its parameters and source-fitted normalization maps remain fixed on new graphs. A post-hoc pipeline correction and complete matched rerun retain strong transfer relative to source-trained logistic regression and LightGBM: TG-Any reaches 0.1614 mean reciprocal rank on Review and 0.2379 on GoogleLocal, respectively 6.2× logistic regression's MRR and 23.7% above LightGBM. On Coin's official split, its 0.8057 is close to a target-window-trained logistic regression's 0.8056, with an interval containing zero for their difference. Simpler fusion rules qualify these gains: uniform weights reach 0.3133 on Review, and Global is close to TG-Any on GoogleLocal. Global also exceeds query conditioning on the three target validation splits. The shared signal interface supports cross-graph reuse, while greater source accuracy and query conditioning do not guarantee better transfer.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Temporal graphs, temporal link prediction, cross-graph transfer, zero-shot learning, streaming graph heuristics, signal fusion

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_F9KgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors consider the problem of temporal link prediction and focus on the idea of learning temporal patterns once, that is, using one dataset, and then applying the predictor to unseen datasets. The reasoning behind this is that different temporal graphs often follow similar patterns, so that the dynamics of one dataset should enable learning general patterns that can be found in other datasets as well. The authors define a set of 24 features that characterise the activity patterns of node interactions, including aspects such as recency and frequency of interactions. To make scores across the different features comparable, the authors normalise them. To test their approach of learning interaction patterns from one dataset and predicting for others, they fix tgbl-wiki-v2 as the training dataset and predict on 4 datasets (tgbl-review, tgbl-coin, tgbl-flight, and TGB-Seq/GoogleLocal).

While I find the central idea of learning on one dataset to predict on others exciting and see great potential, I had trouble following the authors' discussions. Overall, I felt that it was difficult to understand the purpose of each paragraph: what are the authors saying and why? What are the important take-home messages for the reader? Further, I believe that the evaluation needs to be more extensive: showing that the method works by only focusing on one source dataset seems insufficient to me. Moreover, I am missing some analyses regarding "how to choose a good source dataset?" and "when does the approach work vs. fail?". Overall, I find that the paper requires careful and rather extensive revisions to clarify the take-home messages and show that the method works before it can be considered for publication.

Critical strengths

I believe the biggest strength of this paper is the simplicity of the idea, which, despite, or perhaps exactly because of its simplicity, can have a wide-reaching impact.

Critical weaknesses
  • I had trouble following the authors' discussion and figuring out what the purpose of each paragraph was. I recommend the authors revise their paper for clarity and consider the exact role of each paragraph. What should the reader have learned after each paragraph? What is the take-home message?
  • I do not believe that the experimental evaluation is sufficient. Specifically, I believe that evaluating only one source dataset for training is not enough to make a general statement regarding how well the proposed method works -- essentially, we have no idea what to expect when we use another dataset for training. Moreover, we do not understand the failure modes: when does it not work?
Questions for authors

There are a few things that remained unclear to me while reading the paper:

  1. I was wondering what the motivation for your work is. Did the idea come from a specific application or problem? Does it have to do with saving computational resources?
  2. I encountered some terms that I am not used to in the ML literature. For example, by "control", do you mean "baseline"? What do you mean by "official test" in Tables 1 and 2?
  3. How do you recommend selecting the source dataset? For example, let's assume that I have some datasets and want to select "the best" (or maybe just a good enough) source dataset to predict on the rest of my datasets. How should I do that? Is there some "analytical" way you would recommend?
  4. Connected to the previous question, is there a way to detect "outlier datasets" in the sense that we could say "if trained on dataset X, we should expect that the predictions on dataset Y would be bad, so don't predict on Y"? Perhaps this is actually the same question as the previous one, just put in other words.
  5. The performance shown in your tables does not seem to be particularly convincing. What do you think why that is? Could this be a problem of not having chosen the right source dataset?
Reviewer_CSjcOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper explores a lightweight approach to cross-graph temporal link prediction by combining interpretable behavioral signals with a source-trained fusion network. The shared feature interface allows deployment on new graphs without transferring node identities or fitting model parameters on the target graph.

Critical strengths

S1. The paper reports informative comparisons with uniform and global-weight fusion, including results unfavorable to the proposed query conditioning. It also explicitly distinguishes corrected post-hoc results from the original experiments, making the limitations of the evidence visible.

Critical weaknesses

W1. The central architectural contribution lacks convincing transfer benefits. Query-dependent fusion improves source performance but underperforms a fixed global mixture on all three target validation sets. W2. The feature interface has important limitations for cross-graph transfer. The source-fitted quantile transformation maps channels 12–15 to constants on every target graph because these channels are constant on the bipartite source graph. Consequently, potentially informative target-side structural differences are discarded. W3. Fixed temporal windows and decay constants are applied in unrescaled dataset-specific timestamp units, without a verified common physical unit. The paper does not sufficiently establish robustness to these structural and temporal differences. W4. The presentation obscures the final method and conclusions. The manuscript repeatedly interleaves historical experiments, corrected results, evaluation gates, and implementation audit records. Meanwhile, essential feature definitions and normalization behavior are deferred to the appendix. A clearer account of the final algorithm, evaluation protocol, and supported claims would substantially improve readability.

Questions for authors

See weaknesses.

Reviewer_2To8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose "TG-Any" an approach to temporal link prediction where the model is trained "once" on a source graph and that model is used for temporal link prediction on multiple "target" graphs without having to retrain the model on each of the targets. The main motivation is that patterns in the source graph such as recency, popularity, repeat interaction, neighborhood statistics etc remain meaningful for transfer downstream. So, the authors adopt a feed forward neural network that utilizes the distribution of the 24 hand crafted features across candidates for a query to predict a score for the candidates, listwise InfoNCE is the objective used for training the single model. The authors evaluate this approach by choosing tgbl-wiki as the singular source graph and test on a variety of downstream target graphs like review, googleLocal, Coin and Flight. The baselines chosen to be alternative ways to utilize the proposed 24 features like for e.g., logistic regression , LightGBM. They also choose CrossDyG which is neural cross graph predictor which uses graph history as opposed to hand crafted features. Results do show the benefits of TG-Any in terms of improved MRR on target graphs compared to baselines.

Critical strengths
  1. Clear and easy to understand, the hand crafted features make sense and the simplicity of the featurization also makes it a good candidate for a diagnostic tool and good baseline for future methods even for methods that train bespoke models for the target as it can help with dataset insights for temporal link prediction applications and practitioners.
Critical weaknesses
  1. The authors choose tbgl-wiki as the only training graph for their approach "TG-Any", while they do show generalization benefits on downstream target "eval" graphs like review, googleLocal, Coin and Flight, IMO the core argument of TG-Any which is that training on general properties of source graph transfer over to downstream targets are not robustly shown in experiments. Ideally, would like to experimentation with multiple runs on "diverse" sources (in addition to tgbl-wiki), kind of like source / training graph sensitivity of TG-Any.

  2. The experiment design is not fully consistent to show the value of TG-Any compared to baselines, for eg. CrossDyG is only evaluated on wiki -> review. Why are other targets not evaluated, even though the authors choose to report on targets like flights, coin and others for some other baselines in Table 1. In Table 3, other targets like Coin and Flight are chosen for ablations on top of TG-Any, why is GoogleLocal not included there ? In table 24 in the appendix CrossDyG is only evaluated against "review" target and why are the same benchmarks used for table 1 not chosen ?

  3. Neural baselines chosen are extremely limited, important well-known works from literature such as CAWN, TGRank, CrossLink, DyGFormer are natural choices for inductive temporal link predictions and can be applied to this setting of TG-Any and should be chosen as baselines, this omission hurts / weakens the opening argument of the proposed paper.

Questions for authors
  1. Please address the questions raised about the experiment setup and in consistent sets of targets graphs chosen to evaluate baselines and ablations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.