Heterogeneous representation learning and matching for few-shot relation prediction
- Publisher:
- ELSEVIER SCI LTD
- Publication Type:
- Journal Article
- Citation:
- Pattern Recognition, 2022, 131
- Issue Date:
- 2022-11-01
Closed Access
| Filename | Description | Size | |||
|---|---|---|---|---|---|
| Heterogeneous representation learning and matching for few-shot relation prediction.pdf | Published version | 3.89 MB |
Copyright Clearance Process
- Recently Added
- In Progress
- Closed Access
This item is closed access and not available.
The recent explosive development of knowledge graphs (KGs) in artificial intelligence tasks coupled with incomplete or partial information has triggered considerable research interest in relation prediction. However, many challenges still remain unsolved: (i) the previous relation prediction methods require a significant amount of training instances (i.e., head-tail entity pairs) for every relation, which is infeasible in practical scenarios; and (ii) the representation learning of entities and relations always assumes that all local neighbors and their features contribute equally to the embedding, not sufficiently considering the heterogeneity of the information; and (iii) the state-of-the-art methods usually require a lot of training time, resulting in a high cost in real-world applications. To overcome these challenges, we propose a heterogeneous representation learning and matching approach, Multi-metric Feature Extraction Network (MFEN for short), for few-shot relation prediction in KGs. Our method focuses on knowledge graphs to sufficiently explore the topological structure and node content in graphs. Rather than taking the average of the embeddings of all relational neighbors, a heterogeneity-aware representation learning method is proposed to generate high-expressive embeddings, which capture the heterogenous roles of the relational neighbors of given entity and all of their features via a convolutional encoder. To learn the expressive representations efficiently, a single-layer CNN architecture with multi-scale filters is devised. In addition, multiple heuristic metrics are combined to efficiently improve the accuracy of similarity calculation. The proposed MFEN model is evaluated on two representative benchmark datasets NELL and Wiki. Extensive experiments have demonstrated that our method gets more than 5% accuracy improvement and three times speedup to state-of-the-art models. Code is available on https://github.com/summer-funny/MFEN.
Please use this identifier to cite or link to this item: