ADARec: An Adaptive Data Augmentation Framework for Sequence Recommendation

Guanghu Yuan, Shujie Li, Lei Fang, Chengming Li

Abstract: Recent advancements in sequence recommender systems (SRS) have led to significant performance improvements by harnessing the power of deep learning models. The effectiveness of these models largely hinges on the availability of long training sequences. However, in real-world recommendation settings, the presence of numerous new users and novel scenarios often results in a limited number of interactions between users and items. This scarcity of data makes it difficult for the models to accurately capture user preferences. This challenge, known as the cold-start problem, has been a longstanding and critical obstacle for SRS in delivering accurate recommendations in such scenarios. In this paper, we propose a novel Adaptive Data Augmentation framework for sequence Recommendation (called ADARec). Specifically, we first apply various data augmentation techniques to expand the original sequential data, leveraging the inherent knowledge of pre-trained models to enhance the representation of the original data. Since the augmented data inevitably includes noise, we employ a multi-stage adaptive learning framework to train the augmented data. This framework operates at both coarse-grained and fine-grained levels to amplify the weights of valuable data while diminishing the influence of less valuable data. Extensive experiments conducted on two real-world recommendation datasets demonstrate that our approach achieves competitive results compared to state-of-the-art methods.

External IDs:dblp:conf/smartiot/YuanLFL24