What is it about?
In this paper, we propose a general sequential recommendation model with a meta-learning algorithm, which we call Metaoptimized Contrastive Learning for sequential Recommendation (MCLRec). Firstly, auxiliary contrastive learning is chosen to complement the primary task in both the data and model perspectives. A learnable model augmentation method is combined with data augmentation methods in MCLRec for extracting more expressive features. In this way, model augmented views can serve as additional contrastive pairs and be contrasted with data augmented views during training. Additionally, the parameters of the model augmenters could adaptively adjust to different datasets. Secondly, we leverage a meta manner to update the parameters of the augmentation model according to the performance of the encoder. By using such a learning paradigm, the augmentation model could learn discriminative augmented views based on a relatively restricted amount of interactions (e.g., small batch size). Finally, a contrastive regularization term is considered in MCLRec by injecting a margin between the similarities of similar pairs for avoiding feature collapse and generating more informative and discriminative features.
Featured Image
Photo by Muhammad Rizki on Unsplash
Why is it important?
Extensive experiments on different public benchmark datasets demonstrate that MCLRec can significantly outperform the stateof-the-art sequential methods.
Perspectives
We developed a novel contrastive learning-based model called meta-optimized contrastive learning (MCLRec) for sequential recommendation. We took the advantage of data and learnable model augmentation in contrastive learning to create more informative and discriminative features for recommendations. By applying meta-learning, the augmentation model could update its parameters in terms of the encoder’s performance. Extensive experimental results showed that the proposed method outperforms the state-of-the-art contrastive learning based sequential recommendation models. In addition, due to the generalization of our framework, in the future, MCLRec could be applied to many other recommendation models and further improve their performance.
Xiuyuan Qin
Read the Original
This page is a summary of: Meta-optimized Contrastive Learning for Sequential Recommendation, July 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3539618.3591727.
You can read the full text:
Resources
Contributors
The following have contributed to this page
