DMR: disentangled and denoised learning for multi-behavior recommendation DOI Creative Commons
Yijia Zhang, Wanyu Chen, Fei Cai

и другие.

Complex & Intelligent Systems, Год журнала: 2025, Номер 11(2)

Опубликована: Янв. 16, 2025

In recommender systems, leveraging auxiliary behaviors (e.g. view, cart) to enhance the recommendation in target behavior purchase) is crucial for mitigating sparsity issue inherent single-behavior recommendation. This has given rise multi-behavior (MBR). Existing MBR task faces two primary challenges. First, irrelevant that do not align with behavior, can negatively impact prediction accuracy user preference behavior. Second, these methods typically learn coarse-grained preferences, failing model consistency and distinctiveness among multiple at a fine-grained level. To address issues, we propose disentangled denoised (DMR), which employs preferences reflected guide learning of item embeddings behaviors. Specifically, first design graph convolutional network, modeling under view attribute domains. We also contrastive strategy, where by reducing influence noisy data existing Experimental results on real-world datasets show proposal improve performance models effectively, achieves average 3.12% Retailrocket dataset 3.28% Beibei over state-of-the-art baselines. Extensive experiments demonstrate our model's competitive learning.

Язык: Английский

DMR: disentangled and denoised learning for multi-behavior recommendation DOI Creative Commons
Yijia Zhang, Wanyu Chen, Fei Cai

и другие.

Complex & Intelligent Systems, Год журнала: 2025, Номер 11(2)

Опубликована: Янв. 16, 2025

In recommender systems, leveraging auxiliary behaviors (e.g. view, cart) to enhance the recommendation in target behavior purchase) is crucial for mitigating sparsity issue inherent single-behavior recommendation. This has given rise multi-behavior (MBR). Existing MBR task faces two primary challenges. First, irrelevant that do not align with behavior, can negatively impact prediction accuracy user preference behavior. Second, these methods typically learn coarse-grained preferences, failing model consistency and distinctiveness among multiple at a fine-grained level. To address issues, we propose disentangled denoised (DMR), which employs preferences reflected guide learning of item embeddings behaviors. Specifically, first design graph convolutional network, modeling under view attribute domains. We also contrastive strategy, where by reducing influence noisy data existing Experimental results on real-world datasets show proposal improve performance models effectively, achieves average 3.12% Retailrocket dataset 3.28% Beibei over state-of-the-art baselines. Extensive experiments demonstrate our model's competitive learning.

Язык: Английский

Процитировано

0