Temel Kavramlar
User intents and behavior distributions are disentangled effectively through Intent-aware Recommendation via Disentangled Graph Contrastive Learning (IDCL), improving recommendation performance and interpretability.
Özet
IDCL introduces a novel approach to disentangle user intents and infer behavior distributions. By utilizing a graph neural network, the model learns interpretable intents and behavior distributions simultaneously. The proposed method involves modeling user behavior data as a user-item-concept graph, designing a GNN-based behavior disentangling module, and implementing intent-wise contrastive learning to enhance intent disentangling. Additionally, coding rate reduction regularization is introduced to ensure behaviors of different intents are orthogonal. Experimental results demonstrate the effectiveness of IDCL in improving recommendation performance across various datasets.
İstatistikler
Extensive experiments demonstrate substantial improvement in Recall@20, Recall@50, Recall@100, and NDCG@100.
For ML-100k dataset: Recall@20 - 0.3235, Recall@50 - 0.4450, Recall@100 - 0.5554, NDCG@100 - 0.3378.
For ML-1M dataset: Recall@20 - 0.3160, Recall@50 - 0.4888, Recall@100 - 0.6268, NDCG@100 - 0.4302.
For MtBusiness dataset: Recall@20 - 0.2973.
Alıntılar
"We propose the Intent-aware Recommendation via Disentangled Graph Contrastive Learning (IDCL) to simultaneously learn interpretable user intents and behavior distributions over them."
"Extensive experiments demonstrate the effectiveness of IDCL in terms of substantial improvement and interpretability."