- Rating
- 2
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
For temporal graph link prediction, diagnostic experiments show that query-aware methods outperform query-agnostic baselines mostly on repeated queries. These methods exhibit weakened or reversed performance on first-seen queries. The authors define this phenomenon as Pair-Statistic Support Bias. This bias describes that shortcuts built upon historical pair statistics only work within historical support ranges. The proposed StatE integrates time-decay smoothed node memory, leverages relational extrapolation encoding to model witness-target dependency and adopts lightweight gMLP aggregation equipped with attention pooling. Experiments across ten datasets and eleven baselines achieve up to 17.93% relative MRR improvement. However, the paper suffers from poor writing, which substantially hinders readability.
Critical strengths
This is a distinctive perspective. The authors partition test queries first into repeated and first-seen categories, and further split them across direct, co-occurrence-only, and none support regimes as shown in Figures 1, 3, and 5. This evaluation clearly demonstrates that aggregate MRR conflates pattern exploitation and relational generalization. The statistic-removal study in Observation II adopts an interventional, rather than observational, design. Removing pair statistics reduces MRR for repeated queries and improves MRR for first-seen queries. This research indicates that statistical shortcuts not only cease to function outside their support range but can also introduce active bias during optimization.
Critical weaknesses
-
The evaluation relies on non-standard datasets and metrics instead of the official TGB benchmark, limiting the credibility of empirical results. Moreover, the proposed method constitutes largely combinatorial innovation assembled from existing methods. The related work does not discuss a prominent line of static graph link prediction research exemplified by NCN, BUDDY and SEAL, which replaces hand-crafted heuristic statistics with learnable pairwise representations. This omission obscures the close conceptual parallel between this prior paradigm and the authors’ method, substantially weakening the justification for its originality.
-
Poor writing clarity and insufficient conceptual grounding render the paper difficult to follow. Core concepts including Pair-Statistic Support Bias, relational extrapolation, pattern exploitation, statistic dependence, co-occurrence, repeat frequency, and identity matching are all introduced without explicit definitions prior to their use. Moreover, it remains unclear whether the dichotomy between query-aware and query-agnostic methods has been adopted in prior literature. If this taxonomy is not widely recognized within the field, the paper provides an overly concise discussion on the performance advantages of query-aware methods.
Questions for authors
-
What distinguishes the pairwise target-dependent representation from the line of static graph link prediction methods that replace hand-crafted heuristic statistics with learnable pairwise representations, such as NCN, BUDDY, and SEAL?
-
Is the taxonomy of query-aware and query-agnostic methods adopted in prior literature?
-
The ablation study presents an inconsistency with the core claim of the paper. Table 3 shows that StatE+Co-occ outperforms vanilla StatE across all three datasets (UCI: 79.92 > 79.89, MOOC: 68.44 > 66.06, YouTube: 68.38 > 65.37). StatE+Stats also achieves better performance on UCI and MOOC. In other words, the best-performing proposed variant still incorporates explicit pairwise statistics. This directly contradicts the "beyond pair statistics" claim in the title and abstract. The authors rationalize this observation by stating that co-occurrence retains complementary information. However, if StatE+Stats is indeed the optimal model, the narrative of going beyond pair statistics becomes questionable. We suggest the authors either acknowledge this inconsistency and revise their claim to "learning relations as the primary component with explicit statistics as an optional supplement", or explain why the combined variant is not adopted as the main model.