Learning representation of heterogeneous temporal graphs for recommendation
Abstract Heterogeneous temporal graphs are important abstractions for organising data in recommender systems, for which an effective representation learning method is presented in this paper. Specifically, an attention‐based two‐stage aggregation technique is adopted to aggregate the message passed...
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wiley
2021-10-01
|
| Series: | Electronics Letters |
| Online Access: | https://doi.org/10.1049/ell2.12265 |
| Summary: | Abstract Heterogeneous temporal graphs are important abstractions for organising data in recommender systems, for which an effective representation learning method is presented in this paper. Specifically, an attention‐based two‐stage aggregation technique is adopted to aggregate the message passed over each edge. The aggregation stage first involves temporal aggregation for each group of neighbouring nodes connected by the same type of edges. Then a further aggregation is performed among the aggregated results of each edge type. The method can capture the temporal evolution patterns of heterogeneous temporal graphs in continuous time‐space and model the graphs' heterogeneity without requiring predefined meta paths. Experimental results on public recommendation datasets demonstrate that the recommendation algorithm based on the representation learning method outperforms the state‐of‐the‐art baselines. |
|---|---|
| ISSN: | 0013-5194 1350-911X |