ACNN-FM: A novel recommender with attention-based convolutional neural network and factorization machines
ACNN-FM: A novel recommender with attention-based convolutional neural network and factorization machines
复制标题
ACNN-FM：一种具有基于注意力的卷积神经网络和分解机的新型推荐器
DOI：
10.1016/j.knosys.2019.05.029
复制
发表时间：
2019-10-01
影响因子：
8.8
通讯作者：
Qin, Xueyang
中科院分区：
文献类型：
--
作者：
Pang, Guangyao;Wang, Xiaoming;Qin, Xueyang
随着互联网的快速发展，网络购物、网络教育、数字娱乐等应用平台产生的数据量急剧增长，给互联网用户带来了严重的信息过载问题。传统的推荐方法是互联网用户从各种信息中提取有价值信息的关键。针对现有推荐系统存在的数据稀疏、冷启动、过度依赖人工特征提取等问题，提出了一种基于注意力的卷积神经网络和因子分解机的推荐系统（ACNN-FM），实现了带注释的推荐。首先，从局部到整体的角度，本文提出了词级注意机制和短语级注意机制，以提高卷积神经网络文本处理过程中记忆历史词汇（短语）重要性和顺序的能力。其次，它构建了一个模型，自动提取隐藏的用户和项目的自然语言形式的评论的特征。最后，利用因子分解机分析用户隐藏特征与项目之间的关联，实现基于关联的推荐。实验结果表明，ACNN-FM方法优于现有的NARR方法，在NARR、DeepCoNN、BCF和NMF方法中，ACNN-FM具有最高的数据利用率，从而在大规模数据环境下显著提高了推荐性能。(C)2019 Elsevier B. V.版权所有。
With the rapid development of the Internet, the data generated from application platforms such as online shopping, e-education, and digital entertainment has exhibited dramatical growth, which has caused serious information overload to Internet users. The traditional recommendation approaches are crucial for Internet users to extract valuable information from various information. However, there exist some problems such as sparse data, cold start, and over-reliance on manual extracted feature and so on. To address the above problems, this paper proposes a novel recommender with Attentionbased Convolutional Neural Network and Factorization Machines (ACNN-FM), which achieves the recommendation with comments. Firstly, from the perspective of local to overall, this paper proposes a word-level attention mechanism and a phrase-level attention mechanism to increase the ability to remember the importance and the order of historical vocabulary (phrase) in the process of text processing of convolutional neural networks. Secondly, it constructs a model to automatically extract hidden features of users and items from comments in the form of natural language. Finally, we utilize factorization machines to analyze the association between the hidden features of users and items, and implement the recommendation based on the association. Extensive experiments are conducted for demonstrating that ACNN-FM method outperforms state-of-the-art NARR method, and ACNN-FM has the highest data utilization among NARR, DeepCoNN, BCF and NMF methods, thus the recommendation performance is significantly improved in large-scale data environment. (C) 2019 Elsevier B.V. All rights reserved.