.png){kind=icon}
.png){kind=icon}
Return
Dual channel representation-learning with dynamic intent aggregation for session-based recommendation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2024.125273.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Session-based recommendation (SBR) predicts the next item clicked by anonymous users based on the given sessions. Nowadays, numerous SBR models merge global information for representation enhancement, potentially leading to redundancy issues that can diminish recommendation performance. Besides, most methods often only utilize the last clicked item of the session as user's current preference, which makes difficult to accurately capture the user's long-term intent transfer due to the absence of user information. address above issues, we propose the D ual C hannel R epresentation-learning with D ynamic I ntent A ggregation DIA-DCR ) model, which reasonably merging global information for recommendation while considering dynamic user intent. Specifically, we first construct the global graph based on degree-sensitive pruning and use neighbor aggregation to learn the global item representations while reducing redundancy. Then we employ normalization and residual connection to amalgamate item representations from the local channel, obtaining the overall item embedding. To capture user intent, we design a portable Intent Aggregation Module (IAM) aggregate intent and temporal information with item embedding, then intercept the user's last several clicked items for contextual short-term intent enhancement. The IAM can be plugged into the dual-channel GNN structure, effectively learning dynamic intent to enhance prediction accuracy. What is more, we use label smoothing to avoid gradient conflict. Extensive experiments on three real-world datasets illustrate the effectiveness of our proposed method. The source code of our model is available at https://github.com/Sunjr/DIA-DCR.
Keywords:
Session-based recommendation
Dynamic intent extraction
Dual channel learning
Graph neural networks
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W