引用本文
  • 杨振宇,刘国敬,王钰.基于细粒度特征交互的推荐模型[J].信息安全学报,2021,6(5):144-155    [点击复制]
  • YANG Zhenyu,LIU Guojing,WANG Yu.Recommendation model based on fine-grained feature interaction[J].Journal of Cyber Security,2021,6(5):144-155   [点击复制]
【打印本页】 【在线阅读全文】【下载PDF全文】 查看/发表评论下载PDF阅读器关闭

←前一篇|后一篇→

过刊浏览    高级检索

本文已被:浏览 1148次   下载 719 本文二维码信息
码上扫一扫!
基于细粒度特征交互的推荐模型
杨振宇1,2, 刘国敬2, 王钰2
0
(1.中国矿业大学信息与控制工程学院 徐州 中国 221116;2.齐鲁工业大学(山东省科学院)计算机科学与技术学院 济南 中国 250353)
摘要:
随着电子商务的兴起,用户在网购的同时留下了大量的评论。用户评论通常包含丰富的用户兴趣和项目属性等语义信息,反应了用户对项目特征的偏好。近年来,许多基于深度学习的方法通过利用评论进行推荐,并取得了巨大成功。这些工作主要是采用注意机制来识别对评分预测很重要的词或方面。它们单一的从评论中提取特征信息,并通过用户和物品的特征交互得到预测分数。然而,过度的聚合可能会导致评论中细粒度信息的丢失。此外,现有的模型要么忽略了用户和项目评论的相关性,要么只在单个粒度上构建评论特性交互,这导致用户和项目的特征信息不能被有效而全面地捕获。针对上述问题,在本文我们考虑通过从评论的多个粒度捕获特征信息,然后为用户和物品进行多粒度下的特征交互,可以实现更好的评分预测和解释性。
为此,我们提出了一种新的用于评分预测的细粒度特征交互网络(FFIN)。首先,模型并没有将用户的所有评论聚合成一个统一的向量,而是将用户和物品的每条评论单独建模,通过堆叠的扩展卷积分层地为每个评论文本构建多层次表示,充分地捕获了评论的多粒度语义信息;其次,模型在每个语义层次上构建用户和物品评论的细粒度特征交互,这有效避免了单粒度交互导致的次级重要信息被忽略的问题;最后,由于用户的评论行为通常是主观且个性化的,我们没有使用注意力机制来识别重要信息,而是通过类似于图像识别的层次结构来识别高阶显著信号,并将其用于最终的评分预测。我们在6个来自Amazon和Yelp的具有不同特征的真实数据集上进行了广泛的实验。我们的结果表明,与最近提出的最先进的模型相比,所提出的FFIN在预测精度方面获得了显著的性能提升。进一步的实验分析表明,多粒度特征的交互,不仅突出了评论中的相关信息,还大大提高了评分预测的可解释性。
关键词:  特征交互  评论推荐  多粒度  卷积神经网络
DOI:10.19363/J.cnki.cn10-1380/tn.2021.09.11
投稿时间:2021-04-20修订日期:2021-07-31
基金项目:本课题得到山东省重大科技创新工程(No.2020CXGCO10102)资助。
Recommendation model based on fine-grained feature interaction
YANG Zhenyu1,2, LIU Guojing2, WANG Yu2
(1.School of lnformation and Control Engineering, China University of Mining and Technology, Xuzhou 221116, China;2.School of Computer Science and Technology, Qilu University of Technology (Shandong Academy of Sciences), Jinan 250353, China)
Abstract:
With the rise of e-commerce platforms, users left a large number of reviews while shopping online. These reviews usually contain rich semantic information about user interests and item attributes, reflecting user preferences for item features. In recent years, many deep learning-based solutions have been proposed by exploiting reviews for recommendation, and have achieved great success. The attention mechanism is mainly adopted in these works to identify words or aspects that are important for rating prediction. They singularly extract feature information from reviews, and obtain prediction ratings through the feature interaction between users and items. However, excessive aggregation may result in the loss of fine-grained information in reviews. In addition, the existing models either ignore the relevance of user and item reviews or construct review feature interactions only at a single granularity, which results in the features information of users and items not being captured efficiently and comprehensively. In order to address the above issue, in this paper we consider that by capturing feature information from multiple granularities of each review, and then performing feature interactions at multiple granularities for users and items could enable a better rating prediction with explainability.
To this end, we propose a novel fine-grained feature interaction network (FFIN) for rating prediction in this paper. In the first place, instead of aggregating all the user's reviews into a uniform vector, the model deals with each review for the user and the item separately, and construct a multi-level representation for each review text hierarchically by stacked extended convolution, which adequately captures the multi-granularity semantic information of the review. In the second place, the model constructs the fine-grained feature interaction of user and item reviews at each semantic level, which effectively avoids the problem that secondary important information is overlooked due to single-grained interaction. In the end, the user's review is usually subjective and personalized, instead of using the attention mechanism to identify important informations, the high-order salient signals are identified by a hierarchical structure similar to image recognition and used in the final rating prediction. Extensive experiments are conducted over six real-world datasets with diverse characteristics. Our results demonstrate that the proposed FINN obtains substantial performance gain over recently proposed state-of-the-art models in terms of prediction accuracy. At the same time, further experimental analysis shows that the interaction of multi-granularity features not only highlights the relevant information in the reviews, but also greatly improves the interpretability of the rating prediction.
Key words:  feature interaction  review recommendation  multi-granularity  convolutional neural network