基于自监督学习和数据回放的新闻推荐模型增量学习方法  

Incremental News Recommendation Method Based on Self-supervised Learning and Data Replay

作　　者：林威 LIN Wei(School of Computer Science,South China Normal University,Guangzhou 510631,China)

机构地区：[1]华南师范大学计算机学院,广东广州510631

出　　处：《计算机与现代化》2023年第12期1-6,共6页Computer and Modernization

基　　金：国家自然科学基金资助项目(62172166,61772366);广东省基础与应用基础研究基金资助项目(2022A1515011380);上海市自然科学基金资助项目(17ZR1445900)。

摘　　要：个性化的新闻推荐技术对于缓解信息过载和提高用户体验具有重要意义。新闻推荐模型通常基于固定数据集进行迭代训练,然而在现实场景下,新闻推荐模型需要不断学习以适应新的用户和新闻。为此,增量学习被提出用于帮助新闻推荐模型进行增量更新。新闻推荐模型增量学习的主要挑战为灾难性遗忘问题,即模型会忘记它之前学习过的用户偏好。鉴于此,本文提出基于自监督学习和数据回放(Self-supervised Learning and Data Replay)的新闻推荐模型增量学习方法SSL-DR。SSL-DR首先在新闻推荐任务中加入自监督学习任务来获取用户的稳定偏好,有效减轻灾难性遗忘问题。为了巩固已学知识,SSL-DR进一步通过基于用户对于候选新闻的点击概率分数的采样策略实现数据回放,同时借助知识蒸馏策略转移已学知识。实验结果表明,本文方法可有效提升新闻推荐模型在增量训练过程中的整体推荐性能,明显缓解灾难性遗忘问题。Personalized news recommendation technology is important to alleviate information overload and improve user experience.News recommendation models are usually iteratively trained based on fixed data sets.However,in real scenarios,news recommendation models need to constantly learn to adapt to new users and news.Therefore,incremental learning has been proposed to help models perform incremental updates.The main challenge of the incremental learning of news recommendation models is the catastrophic forgetting problem,where a trained model forgets the user preferences it has previously learned.In view of this,this paper proposes SSL-DR,an incremental learning method of news recommendation models based on self-supervised learning and data replay.SSL-DR firstly adds the self-supervised learning task to the news recommendation task to obtain the user's stable preference,which effectively reduces the problem of catastrophic forgetting.To consolidate the learned knowledge,SSLDR further implements a sampling strategy based on the user's click probability scores for candidate news to achieve data replay and transfer the learned knowledge through a knowledge distillation strategy.The experimental results show that,our method can effectively improve the overall recommendation performance of the news recommendation model in the process of incremental training,and significantly alleviate the problem of catastrophic forgetting.

关 键 词：新闻推荐 增量学习 灾难性遗忘 自监督学习 深度学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]