Introduction
This research presents a hybrid approach to Knowledge Graph Completion (KGC) that synergizes logical rule mining with graph embeddings. By adopting a transductive setting—where the specific test queries are known during the training phase—the authors augment training data with high-probability triplets derived from relational rules. This targeted augmentation leads to significant accuracy improvements on standard benchmarks.
The Challenge of Knowledge Graph Completion
Knowledge Graphs (KGs), such as Freebase and WordNet, are foundational for many AI applications but are notoriously incomplete. KGC aims to predict missing links between entities. Currently, the field is split between two paradigms:
- Embedding Models: These map entities and relations into high-dimensional vector spaces. While effective at capturing global patterns, they often fail to model explicit logical dependencies.
- Rule-Based Models: These utilize logical sequences (e.g., "If A is born in B, and B is in C, then A is a citizen of C") for prediction. They offer high precision but lack robustness when specific paths are missing from the training data.
Hybrid models like UniKER attempt to bridge this gap but are often computationally intensive due to iterative filtering and retraining. This paper proposes a more efficient hybrid method that leverages transductive data augmentation without discarding potentially useful lower-confidence information.
Methodology: Transductive Data Augmentation
The proposed framework consists of four key stages:
- Rule Mining: Bidirectional random walks are employed to discover relational paths. Unlike traditional methods that prune all but the most accurate rules, this approach retains a broad set (up to 1,000 per relation).
- Transductive Application: The "secret sauce" of this method is its focus. Instead of augmenting the entire graph, rules are applied specifically to the test queries. Triplets suggested by these rules for test queries are added to a candidate set.
- Filtering & Weighting: Candidates are filtered by confidence and quantity. Critically, these new triplets are treated as "soft" facts, weighted by the confidence score of the generating rule.
- Single-Pass Training: A standard embedding model (e.g., RESCAL or TuckER) is trained once on the union of original facts and weighted augmented facts.
Experimental Setup
The method was evaluated on two gold-standard KGC benchmarks: WN18RR and FB15k-237.
- WN18RR: A WordNet subset focusing on English lexical relations.
- FB15k-237: A Freebase subset with redundant inverse relations removed, presenting a more rigorous challenge for path-based reasoning.
Key Insights
- The Transductive Advantage: Focusing on test queries allows the model to "study for the exam," making the data augmentation highly relevant and efficiently targeted.
- Semantic Refinement: Even if a rule-mined triplet is not the exact ground truth, it often involves semantically related entities. This helps the embedding model refine its vector space, positioning the correct answer more accurately during inference.
- Soft Logic Benefits: Including lower-confidence rules with appropriate weighting provides a richer training signal than strict "hard" filtering, which can discard useful relational context.
Results and Analysis
The transductive method was tested using RESCAL and TuckER as base embedding engines.
On the WN18RR dataset, the improvements were substantial. For the TuckER model, the Mean Reciprocal Rank (MRR) increased from 0.464 to 0.508, surpassing the more complex UniKER model (0.485). On FB15k-237, gains were more modest (MRR 0.354 to 0.356), likely due to the dataset's sparser exploitable relational path structure.
Importantly, "Random Augmentation" (adding triplets for arbitrary queries) was found to degrade performance, confirming that the transductive focus is essential for success.
Limitations and Future Work
The effectiveness of this approach is tied to the density of exploitable relational paths within a dataset. While it performs exceptionally well on WN18RR, its impact is lower on more complex graphs like FB15k-237. Furthermore, the transductive requirement—knowing queries in advance—limits its use in real-time, "online" query environments, though it remains highly applicable for batch-processing tasks. Future research will explore advanced noise-reduction techniques and the application of this framework to a broader range of embedding architectures.
Summary updated