基于时空Transformer-encoder的跨社交网络用户匹配方法  

User matching method for cross social networks based on spatial-temporal Transformer-encoder

作　　者：张洋 马强[1] Zhang Yang;Ma Qiang(School of Information Engineering,Southwest University of Science&Technology,Mianyang Sichuan 621010,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621010

出　　处：《计算机应用研究》2024年第12期3742-3748,共7页Application Research of Computers

基　　金：国家自然科学基金面上项目(62071170)。

摘　　要：针对目前基于签到时空数据的跨社交网络用户匹配方法未充分利用时空信息之间的耦合关系,导致时空数据特征提取困难,匹配准确率下降的问题,提出了一种基于时空Transformer-encoder的跨社交网络用户匹配方法。该方法通过网格映射将签到时空信息转换为序列数据,生成签到序列;利用序列嵌入层将离散的签到序列映射到连续高维空间;然后借助多头注意力机制和卷积神经网络提取高维签到特征,并利用卷积神经网络实现优化多头注意力模块权重变换和特征融合;最后利用前馈神经网络实现分类,输出用户匹配得分。在两组真实社交网络用户数据集上进行大量用户匹配实验,与现有方法相比,准确率提升了0.40~10.53百分点,F_(1)值提升了0.43~9.5百分点。这验证了所提方法能够有效提取用户签到耦合特征,并提高用户匹配的性能。In response to the shortcomings of current cross social network user matching methods based on check-in spatial-temporal data that do not fully utilize the coupling relationship between spatial and temporal information,resulting in difficulty in feature extraction from spatial-temporal data and a decrease in matching accuracy.This paper proposed a cross social network user matching model based on spatial-temporal Transformer-encoder.This method converted check-in spatial-temporal information into sequential data through grid mapping,generated check-in sequences.It used sequence embedding layers to map discrete check-in sequences to a continuous high-dimensional space.Then,it used multi-head attention mechanism and convolutional neural network to extract high-dimensional check-in features,and used convolutional neural network to optimize multi-head attention module weight transformation and feature fusion.Finally,it used feedforward neural networks to implement classi-fication and outputting user matching scores.Extensive user matching experiments on two real social network user datasets show improvements in accuracy by 0.40 to 10.53 percentage point,and F_(1)value by 0.43 to 9.5 percentage point,compared to existing methods.The experiment validates that the proposed method can effectively extract user check-in coupling features and improve user matching performance.

关 键 词：跨社交网络 用户匹配 Transformer-encoder 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]