Enhanced ID recommendation method utilizing semantic patterns of interactive behaviors

ID-based recommendation methods in recommender systems utilize unique identifiers of users or items to generate suggestions. However, these methods often encounter challenges such as data sparsity and cold-start problems, especially when using single-domain data. Cross-domain ID-based recommendations can help mitigate cold-start issues by relying on overlapping users or items across different domains. However, cross-domain ID information often lacks overlapped users or items. To address this, latent semantic patterns in behavioral networks across various recommender domains can be leveraged. This method aims to extract user preferences for items from discrete ID data, thereby tackling the limited shared information between these domains.

Based on the study of interaction behaviors, this paper assumes the existence of latent pattern correlations between user-item interactions across different domains. A potential factor connects users across domains, leading some users to exhibit similar interaction behaviors in different contexts. These shared characteristics are referred to as interaction behavior semantic patterns. The proposed pattern-enhanced ID recommendation method enhances ID-based recommendations by leveraging these semantic patterns. In the target domain recommendation task, auxiliary domain information is introduced, and information from both auxiliary and target domains is jointly encoded using a graph neural network. By incorporating interaction behavior semantic patterns, user-item interaction and item description information from the auxiliary domain are transferred to the target domain. This process enhances the semantics of interaction behaviors in ID-based recommendations within the target domain.

This study conducts experiments on nine public datasets. User-item ID interaction data from datasets such as Yelp2018, Amazon-Kindle, Alibaba-iFashion, Amazon-Electronic, Book Crossing, MovieLens10M, MovieLens20M, and MovieLens25M serve as target domain datasets. Meanwhile, item description data from the Citeulike-a dataset is used as the auxiliary domain dataset. There are no overlapping user or item IDs between these domains. Experimental results show that the proposed method outperforms the current state-of-the-art methods, showing improvements in Recall@20 by 3%-30% and in NDCG@20 by 1% to 40%.

This study proposes an ID recommendation method enhanced by interaction behavior semantic patterns based on the assumption of latent pattern correlations in user-item interactions across different domains. By introducing these semantic patterns, this method transfers user-item interaction information and item description information from the auxiliary domain to the target domain, thereby enhancing semantic understanding in ID-based recommendations within the target domain. Experimental results validate the ability of the proposed method to transfer semantic information in the absence of overlapping users and items across domains, yielding better recommendation performance. These findings validate the effectiveness of the proposed assumption and method. Additionally, experiments on ID recommendation tasks in multiple domains show that interaction behavior patterns between similar domains offer better transferability. The closer the auxiliary domain is to the target domain, the more notable the improvement in the target domain's ID recommendation results.