acceptodds
Withdrawn from ICLR 2027

Predicting What Happens Next: Next-Event Training and Evaluation for Temporal Knowledge Graph Extrapolation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Temporal Knowledge Graph Extrapolation (TKGE) aims to predict future events from historical events. Existing TKGE methods typically operate in a relation-specified setting, where the relation is given and the task is to predict the missing entity. However, in many real-world applications, only a single subject entity is available, and the goal is to predict both the missing relation and object entity of a future event. We refer to this setting as relation-unspecified and define the task as Subject-conditioned Relation and Object Prediction (SROP). This mismatch motivates our work. To address it, we propose Next-Event eXtrapolation Training (NEXT), a model-agnostic method that addresses SROP by shifting the optimization objective from predicting the missing entity to predicting the missing relation and then the object entity. NEXT applies to existing TKGE models equipped with an entity-history encoder and a relation-conditioned entity decoder. For such models, NEXT adapts them to SROP by adding an MLP-based prediction head that predicts the missing relation, after which the original decoder predicts the object under that relation. NEXT further introduces a new relation-aware training objective that supplements the original entity-prediction loss with a relation prediction loss. Experiments on four datasets show that NEXT raises SROP performance on ICEWS14, ICEWS18, ICEWS0515, and GDELT from 0.08%, 0.08%, 0.20%, and 0.20% to 9.35%, 4.57%, 10.87%, and 1.21%, respectively. An ablation study under identical architectures and parameter count further shows that this improvement is attributable to the change in the optimization objective rather than to any modification of the model architecture or parameters. These experiments demonstrate that NEXT enables existing models to solve SROP without compromising their original entity prediction capability.

Primary area:
learning on time series and dynamical systems
Keywords:
Temporal Knowledge Graphs, Extrapolation, Event Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xZKZOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on temporal knowledge graph forecasting task, which aims to predict future facts based on the historical facts. Specifically, it introduces a new task setting, Subject-conditioned Relation and Object Prediction (SROP), where only the subject entity is given and the model is required to predict both the relation and object entity of a future event. To address this task, the paper proposes NEXT, a model-agnostic framework that extends existing temporal knowledge graph forecasting models. Experimental results on four widely used datasets demonstrate the effectiveness of the proposed method.

The paper studies a practically relevant task setting and proposes a reasonable framework with encouraging experimental results. However, the methodological contribution substantially overlaps with prior work, and the novelty of the task formulation is not sufficiently positioned within the broader knowledge graph completion literature. Considering these limitations, this reviewer leans toward a negative score.

Critical strengths
  1. The paper introduces a more realistic task setting for temporal knowledge graph forecasting.
  2. The proposed framework is reasonable and straightforward.
  3. The paper conducts extensive experiments on four widely used datasets.
Critical weaknesses
  1. The methodological novelty of NEXT is limited.
    The core idea of NEXT is to factorize the joint probability as p(r,o∣s)=pr(r∣s)po(o∣s,r)p(r, o | s) = p_r(r | s) p_o(o | s, r), which substantially overlaps with RE-Net [1]. In particular, RE-Net already incorporates relation prediction and object prediction into an autoregressive framework. Although the paper cites RE-Net, it does not sufficiently discuss the methodological differences or include RE-Net in the experimental comparison.
  2. The novelty of the proposed task setting is not sufficiently positioned with respect to prior work.
    Predicting both the relation and object entity given only the subject entity has already been studied in general knowledge graph completion, such as the instance completion task introduced in RETA [2]. Therefore, the novelty of SROP mainly lies in extending this existing task formulation to temporal knowledge graph forecasting, rather than introducing an entirely new prediction setting. However, the paper does not discuss this line of research in its related work and should narrow its novelty claims accordingly.

[1] Recurrent Event Network: Autoregressive Structure Inferenceover Temporal Knowledge Graphs

[2] RETA: A Schema-Aware, End-to-End Solution for Instance Completion in Knowledge Graphs

Questions for authors

None

Reviewer_G8XUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies temporal knowledge graph prediction when the relation is not given: the model receives only a subject entity and its history and predicts both the relation and the object of a future event. The authors call this task SROP and propose NEXT, which adds a relation prediction head and jointly trains the relation and entity prediction losses. Existing models can then predict the relation first, followed by the object. The paper evaluates this approach on four datasets with four models and examines its impact on the original entity prediction task.

Critical strengths
  1. Predicting future behavior from a subject and its history, without knowing the relation in advance, is closer to the needs of some real-world applications than the standard relation-given setting.

  2. NEXT adds a relation prediction head while reusing the existing encoder and entity decoder, keeping the extra parameter cost small. Its implementation across four models also suggests that it is fairly easy to apply to different backbones.

  3. Besides evaluating relation and event prediction, the paper also checks entity prediction performance when the relation is given.

Critical weaknesses
  1. RE-NET and EvoKG already model relation prediction from a subject and its history, followed by object prediction, and jointly train these components. The claim that this capability has not been addressed before is therefore too strong; the authors should distinguish a new evaluation setting from existing modeling capabilities and clarify what NEXT adds beyond these methods.

  2. The paper explicitly states that the baseline’s relation prediction head receives no training, so its near-random performance is expected. The large gains over this baseline mainly show the benefit of training the relation head, rather than establishing that NEXT outperforms existing methods or other reasonable ways of adapting models to this task.

  3. Each test event is predicted using the history before its occurrence, but the paper does not clearly explain how the prediction target and valid answers are defined when the same subject has multiple events at the same timestamp. It is therefore unclear whether scoring a single predicted pair against each individual event is an appropriate measure of the proposed next-event prediction task.

  4. Once the first step selects the wrong relation, the object prediction step cannot recover from that mistake. Also, the most likely relation does not necessarily lead to the most likely relation–object pair overall.

Questions for authors

Please refer to the weaknesses.

Reviewer_EM8nOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a task setting called SROP (Subject-conditioned Relation and Object Prediction), in which the goal is to predict the relation and object of the next event given only a subject entity ss and its history. The proposed method, NEXT, extends existing temporal knowledge graph embedding (TKGE) models, which consist of an encoder and a relation-conditioned decoder, by adding an MLP-based relation prediction head to decouple the joint distribution.

Critical strengths
  1. The task addresses a practically meaningful problem. Existing TKGE evaluations almost universally assume that the relation is specified in the query.
  2. The method is lightweight and plug-and-play.
Critical weaknesses
  1. RE-NET’s generative framework factorizes the joint event distribution as p(s)⋅p(r∣s)⋅p(o∣s,r)p(s)\cdot p(r\mid s)\cdot p(o\mid s,r) and sequentially samples the relation and object during multi-step inference. This factorization is essentially consistent with that in Eq. (3) of the present paper and should be explicitly discussed.
  2. Are the models listed in the “Backbone” column of Table 2 models whose relation prediction heads have never been trained? Their SROP performance is at near-random levels. Using them as baselines therefore does not provide meaningful evidence of NEXT’s effectiveness.
  3. Unclear evaluation protocol. Section 4.1 states that “for every test event (s, r, o, t) in the subject role, the model receives only the subject s and the history Ht, and outputs the pair (r^,o^)(\hat{r}, \hat{o})”, however, a subject typically participates in multiple events at the same timestamp, while the model produces the same (r^,o^)(\hat{r}, \hat{o}) prediction for all of them.
  4. How many random seeds were used in the experiments?
  5. Eq. equation 2 in L215.
  6. TiRGN Li et al. and LogCL W. Chen et al. in L49-50.
Questions for authors

See weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.