融合时间和知识信息的生成对抗网络序列推荐算法

Sequence Recommendation Algorithm Based on Generative Adversarial Network Integrating Time and Knowledge Information

序列推荐作为一种常用的推荐系统技术,通过对用户的历史交互序列进行建模来预测下一个可能交互的项目。现有的序列推荐方法主要利用用户交互序列和上下文信息进行推荐,忽略了序列中交互项目之间的时间间隔信息,交互项目之间的组合依赖以及上下文信息中存在噪声的问题,导致推荐结果受限。针对以上问题,提出一种基于生成对抗网络的序列推荐模型TKWGAN,该模型包含一个生成器和一个判别器。生成器结合了用户历史交互序列和各项目之间的时间间隔信息对用户偏好进行建模并生成预测,判别器则引入了知识图谱信息对项目进行语义扩充,从而能更准确地对生成器的预测进行合理性判断。针对用户交互序列和知识图谱信息中可能存在噪声的问题,提出一种基于小波变换的多核卷积神经网络来构造判别器,以更全面、准确地捕获用户的潜在兴趣,提高推荐的准确性。在MovieLens-1M、Amazon Books和Yelp2018这3个公开数据集上的实验结果表明,与8个序列化推荐算法相比,提出的TKWGAN模型在命中率(HR@N)和归一化折损累计增益(NDCG@N)指标上均有显著提升。

Sequential recommendation is a pivotal technique within the realm of recommendation systems and is employed to predict subsequent interaction items by modeling the users′historical sequences of engagements.Nonetheless,contemporary sequential recommendation methods predominantly emphasize user interaction sequences and contextual data in their recommendations.This approach inadvertently disregards the salient temporal interval information between interaction items,the intricate interdependencies among these items,and potential perturbations in contextual information.Consequently,the efficacy of the resulting recommendations is constrained.To address these inherent limitations,this study proposes an innovative sequential recommendation model called TKWGAN,which is founded on a Generative Adversarial Network(GAN).This model consists of a generator and a discriminator.The generator seamlessly integrates the users′historical interaction sequences with temporal interval data pertaining to items within the sequence.This synthesis of information enables the accurate modeling of user preferences and the subsequent generation of predictions.The discriminator introduces salient insights offered by the knowledge graph to enrich the semantic context of the items.This augmentation of the contextual information facilitates a judicious assessment of the predictions formulated by the generator.To address the potential noise inherent in both user interaction sequences and knowledge graph data,a multikernel Convolutional Neural Network(CNN)augmented with wavelet transformation is introduced for the discriminator.This discerning approach enables a comprehensive exploration of users′latent interests,consequently elevating the precision of the recommendations.An empirical evaluation conducted on widely recognized datasets,specifically MovieLens-1M,Amazon Books,and Yelp2018,underscores the superiority of the TKWGAN model.In contrast to eight state-of-the-art sequential recommendation algorithms,the TKWGAN model demonstrates a marked enhancement in the Hit Rate(HR@N)and Normalized Discounted Cumulative Gain(NDCG@N)indicators.