基于深度学习的科技信息文献推荐模型研究  

Research on Science and Technology Information Papers Recommendation Model Based on Deep Learning

作　　者：屈冰洋 王亚民[1] Qu Bingyang

机构地区：[1]西安电子科技大学经济与管理学院,陕西西安710171

出　　处：《情报理论与实践》2021年第11期160-165,共6页Information Studies:Theory & Application

摘　　要：[目的/意义]传统的科技信息文献推荐模型没有充分挖掘科技信息文献以及科研人员的本身特征,因此,文章结合深度学习技术获得健壮的用户及文献特征,提高推荐精度。[方法/过程]在传统概率矩阵分解模型中加入科研人员合作作者以及文献文本信息来构建推荐模型。首先,堆叠去噪自编码器(SDA)利用科研人员合作作者信息提取用户特征;其次,融合注意力机制的双向长短时记忆网络(BiLSTM-Attention)利用文献文本信息提取文献特征;最后,在概率矩阵分解方法中,加入用户特征矩阵和文献特征矩阵,从而预测科研人员偏好,实现个性化科技信息文献推荐。[结果/结论]通过实验证明,新提出的方法在模型上具有更好的拟合准确度,提高了推荐效果。[Purpose/significance] The traditional science and technology information papers recommendation models do not fully explore the characteristics of papers and researchers themselves.Therefore, this paper combines the deep learning technology to obtain robust user and paper characteristics, so as to improve the recommendation accuracy.[Method/process] The traditional probabilistic matrix factorization model is constructed by adding the co-authors of researchers and the papers text information.First, stacked denoising autoencoders(SDA) uses researchers co-author information to extract user feature.Secondly, bidirectional long short-term memory networks-attention(BiLSTM-Attention) utilizes basic information of papers to extract paper feature.Finally, in the probability matrix factorization method, the feature matrix of researchers and papers are added, so as to predict the preference of researchers and realize the recommendation of personalized scientific and technological information papers.[Result/conclusion] Experiments show that the new model has better fitting accuracy in the model and improves the recommendation effect.

关 键 词：深度学习 文献推荐 科技信息 概率矩阵分解 双向长短时记忆网络 堆叠去噪自编码器 推荐模型 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]