Binoclt: A New Binomial Classification Scheme for Long-Text Mining in Online Social Network

With the rapid development of information technology, articles with long-text content are very common among social networks. These articles generally focus on specific areas, whose contents can be further categorized into two kinds - articles spreading instant news or personal opinions, defined as `event', and articles sharing interesting anecdotes, defined as `emotion'. Better classifying these large-scaled and fast-updated online articles brings great benefits to many relevant studies, such as topic detection and user behavior analysis. However, existing methods, such as deep neural network, require massive computing resources and long training time to classify articles. To overcome these shortcomings, we present an innovative online binomial classification scheme for long-text mining in social networks - Binoclt. First, we propose an online topic clustering algorithm based on keyword frequency to combine the extended TextRank with weighted titles and the improved GloVe methods. Then we define two kinds of features, namely Maximum Keyword Score (MKS) and Cluster Item Number (CIN), and use them to train our classifier to distinguish between articles as `events' or `emotion'. To verify the efficiency our classifier, we conduct extensive and comprehensive experiments using the real dataset from WeChat, which is the largest social networking platform in China. Then we compare Binoclt with other advanced neural network methods. The results show that our method contributes to a greater classification accuracy while it also reduces the complexity of work and shortens the training time.