基于变分持续贝叶斯元学习的推荐算法  

Variational Continuous Bayesian Meta-learning Based Algorithm for Recommendation

作　　者：朱文韬 刘威[1,3] 梁上松 朱怀杰 印鉴 ZHU Wentao;LIU Wei;LIANG Shangsong;ZHU Huaijie;YIN Jian(School of Computer Science and Engineering,Sun Yat-sen University,Guangzhou 511436,China;School of Artificial Intelligence,Sun Yat-sen University,Zhuhai,Guangdong 519082,China;Guangdong Key Laboratory of Big Data Analysis and Processing,Guangzhou 510006,China)

机构地区：[1]中山大学计算机学院,广州511436 [2]中山大学人工智能学院,广东珠海519082 [3]广东省大数据分析与处理重点实验室,广州510006

出　　处：《计算机科学》2023年第7期66-71,共6页Computer Science

基　　金：国家自然科学基金(U1911203,61902439,61902438,62002396);广东省基础与应用基础研究基金(2021A1515011902,2020A1515011251,2019A1515011159,2019A1515011704)。

摘　　要：元学习方法近年被引入推荐系统以缓解冷启动问题。现有元学习算法只能提高算法处理一组静态分布的数据集(任务)的能力。当面对多个服从非平稳分布的数据集时,现有模型往往会出现负知识转移以及灾难性遗忘问题,导致算法推荐性能大幅下降。探索了基于变分持续贝叶斯元学习(Variational Continuous Bayesian Meta-Learning, VC-BML)的推荐算法。首先,算法假设元参数服从动态混合高斯模型,使其具有更大的参数空间,提高了模型适应不同任务的能力,缓解了负知识转移问题。然后,VC-BML的任务集群数量由中国餐馆过程(Chinese Restaurant Process, CRP)来灵活确定,使得模型在不同的混合分量中存储不同任务分布的知识,并在类似任务出现时调用这些知识,有助于缓解传统算法中的灾难性遗忘问题。为了估计模型参数的后验概率,算法采用了一种更稳健的结构化变分推理方法来近似后验值,以避免遗忘知识。最后,VC-BML在4个非平稳分布的数据集上的表现均优于基准算法。与基于点估计的基准算法相比,VC-BML提高了模型的稳健型,有助于缓解灾难性遗忘问题。Meta-learning methods have been introduced into recommendation algorithms in recent years to alleviate the problem of cold start.The existing meta-learning algorithms can only improve the ability of the algorithm to deal with a set of statically distributed data sets(tasks).When faced with multiple data sets subject to non-stationary distribution,the existing models often have negative knowledge transfer and catastrophic forgetting problems,resulting in a significant decline in algorithm recommendation performance.This paper explores a recommendation algorithm based on variational continuous Bayesian Meta-learning(VC-BML).Firstly,the algorithm assumes that the meta parameters follow the dynamic mixed Gaussian model,which makes it have a larger parameter space,improves the ability of the model to adapt to different tasks,and alleviates the problem of negative knowledge transfer.Then,the number of task clusters in VC-BML is flexibly determined by Chinese restaurant process(CRP),which enables the model to store knowledge of different task distributions in different mixed components and invoke this know-ledge when similar tasks occur,helping to alleviate the catastrophic forgetting problem in traditional algorithms.To estimate the posterior probabilities of model parameters,a more robust structured variational reasoning method is used to approximate the posterior values to avoid forgetting knowledge.Finally,VC-BML outperforms the baselines on all four datasets with non-stationary distributions.Compared with point-based estimation methods,VC-BML improves the robustness of the model and helps alle-viate the catastrophic forgetting problem.

关 键 词：推荐算法 冷启动 元学习 动态混合高斯模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]