Research on cross‐hierarchical graph network recommendation algorithm based on random walk and convolutional neural network
Convolutional network (CNN) has been widely used in processing various graphics and network data analysis tasks. Therefore, facing the recommendation problem of large‐scale heterogeneous interaction, some studies have raised the problem of whether graph convolution neural network (GCN) can optimize...
Uloženo v:
| Vydáno v: | Concurrency and computation |
|---|---|
| Hlavní autoři: | , |
| Médium: | Journal Article |
| Jazyk: | angličtina |
| Vydáno: |
08.11.2021
|
| ISSN: | 1532-0626, 1532-0634 |
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | Convolutional network (CNN) has been widely used in processing various graphics and network data analysis tasks. Therefore, facing the recommendation problem of large‐scale heterogeneous interaction, some studies have raised the problem of whether graph convolution neural network (GCN) can optimize the integration of node features and topology in complex graphs with rich information, but GCN may not be able to adaptively learn some deep related information between topology and node features, Therefore, the ability of GCN in some classification tasks may be seriously hindered. Therefore, this article proposes a heterogeneous network recommendation algorithm based on random walk and convolutional neural network (RW‐CNN), which can not only obtain the deep information of network structure but also aggregate the node representation. The core idea is to use the rule items based on meta‐path similarity to constrain the implicit representation of users and goods and make full use of the rich structural and semantic information of heterogeneous information networks. Then, combined with convolution neural network, the cross‐correlation information between project and user nodes is processed, so as to fully mine the semantic features of nodes. Finally, through the attention mechanism and fully connected network, the global user feature vector and project feature vector of path fusion at different levels are obtained. Finally, the two characteristic matrices are multiplied to obtain the prediction score. |
|---|---|
| ISSN: | 1532-0626 1532-0634 |
| DOI: | 10.1002/cpe.6704 |