基于用户相关性的动态网络媒体数据无监督特征选择算法

Unsupervised Feature Selection Algorithm for Dynamic Network Media Data Based on User Correlation

移动互联网、社交媒体的快速发展,极大推动了各个领域对文本、图像、视频等网络媒体数据处理的需求.该类数据具有高维度、动态更新、内容复杂的特性,增加了特征计算以及分类难度.同时,当前网络媒体数据的特征选择方法主要针对静态数据,并且对数据格式规范性要求较高.针对上述问题,为保证对动态网络媒体数据的实时特征提取,该文提出了一种基于用户相关性的动态网络媒体数据无监督特征选择算法(Unsupervised Feature Selection Algorithm for Dynamic Network Media Based on User Correlation,UFSDUC).首先,对社交网络中的交互用户进行关系分析,作为无监督特征选择的约束条件.然后,利用拉普拉斯算子构建用户相关性的特征选择模型,量化相关用户之间的关系强弱,通过拉格朗日乘子法给出特征模型中最优用户关系的数学方法.最后,基于梯度下降法设定动态网络媒体数据的阈值,用以计算非零特征权值来更新最优特征子集,达到对网络媒体数据进行有效分类的目的.该算法可在保证用户在相关性完整的基础上对动态网络媒体数据进行准确、实时的特征选择.该文采用3个标准网络媒体数据集,同时与5种目前较为流行的同类型算法进行对比以验证算法的有效性.

With the rapid development of the mobile network and social media,more and more Internet multi-media data including texture,image,video and others produce continuously at all times,meanwhile,requirements that learn and apply such data have growth.However,feature calculation and classification efficiency are severely limited,because of the high-dimensional,the complex content and dynamic updating characteristics of Internet multi-media data.Moreover,traditional algorithms mainly solve the feature extraction and classification problem for static multi-media data,and these algorithms require that data format need to conform the specific standard.Aiming to above problems,we proposed an efficient unsupervised feature selection algorithm based on user correlation that is called by UFSDUC(Unsupervised Feature Selection Algorithm for Dynamic Network Media Based on User Correlation)to ensure the feature extraction in real time for the dynamic multi-media data.Firstly,we analyzed user relationships in social networks,and combine the potential social factor to abstract three kinds of relational models including MFS(Multi-user Follow Same user),SFM(Same user Follow Multi-user),FEO(Follow Each Other).Take such models as the constraint condition for the unsupervised feature selection processing.Secondly,we use Laplace operator with the strength of relationship between users to building the relationship model,and then the lagrangian multiplier method is utilized to obtain the mathematical expression of the optimal relationship in the feature model.Moreover,in the proposed algorithm quantifies the strength of between users,which the more strength of the correlation may be gets the more similar information of the feature of between users.Therefore,our algorithm achieved the optimum solution for the multi-media data of the social network.Finally,we set the threshold of the multi-media data of the social network by utilizing the gradient descent method.This threshold is used to obtain the nonzero feature value,and then update the best subset of features to achieve the efficient performance to classify the multi-media data of the social network.In this paper,contributions of the proposed algorithm can be summarized as follows:(1)different traditional feature select algorithms that each sample need get the classification label,the proposed unsupervised feature selection algorithm can define the feature relationship according to different standards without labeling samples,for instance,the similarity of between samples and the distribution of the local information;(2)the correlative information of users is more stable than the self-users of information,such as the circle of friends once established will stably live in Internet always.Therefore,the proposed method can provides the important constraint condition for the feature extraction of the multi-media data by utilizing the user relevance;(3)the proposed algorithm realizes the feature selection efficiently at real time when the complete user relevance as a precondition.In this paper,we utilize three stander multi-media datasets to verify the proposed algorithm including Sina Weibo dataset,Flicker dataset,Blog Catalog dataset from‘Datatang’.These datasets have many characteristic enhancing the difficult of the feature extraction,such as amount of users,the complex relationship of between users,various categories of users.Moreover,we compare with five popular algorithms to evaluate the performance.