Hypernetwork Knowledge Graph Embeddings

• Knowledge graphs store facts as entity–relation–entity triples, and link prediction aims to infer missing links in these incomplete graphs.Source: §1 · p. 1
• ConvE applies 2D convolutions to reshaped, concatenated entity and relation embeddings, but the approach is unintuitive for 1D embeddings.Source: §1 · pp. 1–2
• HypER uses a hypernetwork to generate relation-specific 1D filters over subject embeddings, avoiding ConvE’s arbitrary 2D reshaping and enabling more complete entity–relation interaction.The generated filters also support knowledge sharing across relations.Source: §1 · p. 2
• HypER can be understood as a tensor factorization model, with convolution providing sparsity and parameter tying for a trade-off between expressiveness and learnable parameters.Source: §1 · p. 2
• HypER consistently achieves state-of-the-art performance across FB15k-237, WN18RR, FB15k, WN18, and YAGO3-10.Source: §1 · p. 2
• HypER’s convolutional benefit comes from restricting which dimensions interact as explicit regularization, rather than finding higher-dimensional structure in embeddings.Source: §1 · p. 2

• RESCAL uses a bilinear score with a full relation matrix; DistMult diagonalizes that matrix, ComplEx extends DistMult to complex vectors, and TransE models relations as translations.Source: §2 · pp. 2–3
• ConvE convolves reshaped, concatenated subject and relation vectors, then scores objects, offering expressiveness, scalability, and robustness despite unclear motivation for its convolution layout.R-GCNs are a separate convolutional approach that captures graph locality.Source: §2 · p. 3
• Unlike ConvE [3], HypER avoids reshaping and concatenating embeddings, applying hypernetwork-generated relation-specific filters across every subject-embedding dimension for quadratic, non-linear interaction.The passage characterizes this design as more expressive while reducing parameters relative to ConvE.Source: §2 · p. 3
• HypER’s 1D convolution is equivalent to multiplying by a sparse tensor with tied weights, yielding an implicit relation matrix, with ReLU before object interaction distinguishing it from prior factorization models.The convolutional tensor is defined by the relation embedding and hypernetwork; the implicit relation matrices correspond to those in RESCAL, DistMult, and ComplEx.Source: §2 · p. 3

3 Link Prediction

• Link prediction assigns each triple (e1, r, e2) a score, with positive scores indicating true facts and negative scores indicating false ones.The scoring function maps subject and object entities to distributed embedding representations and evaluates them using a relation-specific function.Source: §3 · p. 3
• Applying the logistic sigmoid to a score yields a prediction p = σ(s) ∈ [0, 1] interpretable as the likelihood that the queried fact is true.The score is written as s = φ_r(e1, e2).Source: §3 · p. 4
• Table 1 summarizes scoring functions, relation-parameter dimensionalities, and space complexities for models from the literature and HypER.Its caption describes the scoring functions and complexity terms for state-of-the-art link prediction models.Source: §3 · p. 4, Table 1 · p. 4

4 Hypernetwork Knowledge Graph Embeddings

• HypER maps each relation embedding through a fully connected hypernetwork to generate convolutional filters, restricting embedding-dimension interactions to regularize learning while retaining nonlinear expressiveness.The filter size limits which embedding dimensions interact, allowing nonlinear expressiveness with fewer parameters to help limit overfitting.Source: §4 · p. 4

4.1 Scoring Function and Model Architecture

• The subject embedding is convolved with relation-specific filters, projected to de-dimensional space, passed through ReLU, and scored against every object embedding before sigmoid prediction.The hypernetwork output is reshaped into filters; the projected representation is combined with each object embedding by inner product.Source: §4.1 · p. 5, Figure 1 · p. 5

4.2 Understanding HypER as Tensor Factorization

• HypER can be expressed as tensor multiplication of subject, object, and relation embeddings with a constrained core tensor, connecting it to RESCAL, DistMult, and ComplEx factorization approaches.The convolution corresponds to multiplication by a sparse tensor, and linearity permits a relation-specific matrix interpretation.Source: Figure 2 · p. 5, §4.2 · p. 6

4.3 Training Procedure

• Training uses 1-N scoring with Adam and binary cross-entropy, with label smoothing to regularize predictions; the authors leave filtering both subject and object embeddings to future work.1-N scoring evaluates all candidate objects for each subject-relation pair; the cited prior work reports 3x train-time and 300x test-time speedups over 1-1 scoring.Source: §4.3 · p. 6

4.4 Number of Parameters

• HypER uses 4.3M parameters versus ConvE’s 5.1M on FB15k-237, because it directly transforms relations into convolutional filters.Table 2 compares the parameter counts; the prose reports the values and attributes the reduction to HypER’s direct relation-to-filter transformation.Source: §4.4 · p. 6, Table 2 · p. 7

5.1 Datasets

• Experiments cover FB15k, WN18, FB15k-237, WN18RR, and YAGO3-10; FB15k-237 and WN18RR remove inverse relations found in their source datasets.The datasets are subsets of Freebase, WordNet, and YAGO3, respectively.Source: §5.1 · p. 7

5.2 Experimental Setup

• HypER uses 200-dimensional embeddings and 1-N scoring; hyperparameters are selected by validation MRR, and test triples are evaluated with filtered ranking across five reported metrics.The metrics are MR, MRR, hits@10, hits@3, and hits@1.Source: §5.2 · pp. 7–8

5.3 Results

• HypER outperforms other models on nearly all reported metrics across five datasets, with exceptions for WN18 MRR and mean rank on four datasets.The authors note that mean rank is sensitive to outliers and interpret the pattern as HypER ranking many true triples in the top 10 while making larger errors elsewhere.Source: §5.3 · p. 8, Table 4 · p. 9, Table 5 · p. 9
• Reducing ConvE to 16 feature maps to match HypER’s parameter count did not improve ConvE, indicating HypER’s gains are not merely due to fewer parameters.The comparison tests whether HypER benefits from implicit regularization through a reduced parameter count.Source: §5.3 · pp. 9–10
• On WN18RR, HypER remains the best-performing model after DistMult, ComplEx, and ConvE are retuned with 200-dimensional embeddings and 1-N scoring.The comparison reports means and standard deviations across five random-seed runs.Source: §5.3 · pp. 8–9, Table 7 · p. 10
• Removing HypER’s hypernetwork reduces performance on WN18RR and FB15k-237, supporting relation-level knowledge sharing through multi-task learning.Without the hypernetwork, relation embeddings directly correspond to convolutional filters.Source: Table 8 · p. 10, §5.3 · p. 10
• Label smoothing improves WN18RR prediction scores but harms FB15k scores, so the authors recommend evaluating it separately for each dataset.They leave its general utility for future work.Source: §5.3 · p. 10
• Reshaping entity embeddings for 2D convolution provides no benefit, while 1 × 6 and 1 × 9 filters perform comparably and smaller or larger filters fare worse.The authors suggest convolution may limit which embedding dimensions interact, helping avoid overfitting.Source: §5.3 · pp. 8, 10, Table 9 · p. 11

6 Conclusion

• HypER generates relation-specific convolutional filters for subject embeddings, sharing information across relation vectors and creating non-linear entity–relation interactions.The shared information enables multi-task learning across relations; the authors describe HypER as the first link prediction model to create such non-linear interactions through convolution.Source: §6 · p. 11
• HypER achieves state-of-the-art results across almost all metrics on multiple link prediction datasets.The conclusion also describes it as fast, robust to overfitting, and relatively low in parameter count.Source: §6 · p. 11
• HypER gains no benefit from 2D over 1D filters, is closely related to tensor factorization, and uses convolution for a trade-off between expressiveness and parameter count.The conclusion says the 2D result dispels the suggestion that entity embeddings contain 2D structure implied by ConvE.Source: §6 · p. 11
• Future work could apply convolutional filters to both subject and object embeddings, analyze label smoothing, and explore convolutional feature-map interpretability.The authors suggest interpretability analysis could provide insight and potentially improve the model.Source: §6 · p. 11