RGFN: Recurrent Graph Feature Network for ClickBait Detection

With the booming development of social media, clickbaits are widespread in the online world. Online users can get easily infected by these deceptive titles, which has brought many network security problems. Clickbait detection is a sub-problem in practical text classification tasks with strongly unbalanced training set. Previous researches tend to extract implicit text features automatically through an universal deep learning model, which limits the model to capture the text features for lack of positive samples. In addition, recent studies also applied graph-based models to capture the global relationships among documents via the word edges for text classification. These works use graph convolution layer to aggregate the connected words disorderly, and ignore the word position information when doing classification. In this paper, we propose Recurrent Graph Feature Network (RGFN), a graph-based model combines text feature and word position for clickbait detection. RGFN introduces the traditional text features such as characterizing phrase, named entity as feature nodes into a title-word heterogeneous graph structure. RGFN also use GRU aggregation function and fixed-length sampling method to capture the word position information. Experiments show that RGFN achieves higher performance than existing models on the clickbait detection task.

I. Introduction

In the past few decades, with the booming development of new media, the traditional media has been gradually replaced by the network texts and the Clickbait have been appearing widespread in the online world. Clickbaits, in social media, are misleading headlines and links aim to mislead the reader to ‘click’ them. The content behind the link can be a bundled software or even a virus aims to attack users’ computer which has affected the current network security. Detecting and filtering Clickbaits has become one of the most important tasks for Internet enterprises. This paper aims to design the algorithm for early detecting clickbait titles from the headlines. Clickbait detection task is a sub-problem of text classification, and deep learning models are widely used in these tasks, such as textCNN [1] and textRNN [2]. Recently, the text classification model based on the graph neural network (GNN) has also attracted extensive attention such as TextGCN [3], TensorGCN [4] and TextING [5]. Graph-based model can capture the global word co-occurrence of the corpus by aggregating the information of documents with similar words, which achieved state-of-the-art performance on several data sets.