假新聞偵測：圖神經網路模型

Abstract

近年來社群媒體興起加速了假新聞的傳播，因此假新聞偵測逐漸被視為一個重要的任務。有別於過去的研究大多將機器學習模型應用在假新聞偵測並以文字、圖片作為訓練資料，我們認為社群參與資料(Social Interaction Data)也能提供重要資訊。本研究利用線上新聞分享的社群參與資料(Online News Sharing Data)建立一個異質圖(Heterogeneous Graph)，並以圖神經網路模型(Graph Neural Network Model)進行假新聞預測。我們使用了三個推特(Twitter)上的假新聞資料集進行實驗，其中的資料包含原推文、回應推文、推文關係的結構、推文作者與原推文正確性標記等資訊。不同於文獻中的方法，將推文的文字內容特徵與在異質圖上的網路特徵分開進行預測，我們的模型將兩資料同時作為圖神經網路模型的訓練資料。實驗結果顯示，在假新聞偵測這個任務上，比起分開利用各項特徵進行預測，運用圖神經網路與社群參與資料有更突出的表現。

The rise of social media accelerates fake news propagation and worsen the fake news problem. Even though there have been many machine learning approaches proposed on fake news detection, most of them focused on text and image data. To leverage online social interaction data for fake news detection, we proposed a graph neural network model using a heterogeneous graph that captures online news sharing activities. We evaluated our models with three fake news datasets constructed based on news-worthy events on twitter. The datasets contain variables such as source tweets, reply tweets, tweets relation structures, and manually labeled ground truth for fake news. Except from previous works predicting fake news with text content features or embedding features on heterogeneous graph separately, our models combine both features as the training data and predict with graph neural network model. Experimental results show that our models perform better comparing with other methods predicting with text features and graph embedding features separately. The results suggest that graph neural network model and online news sharing data are helpful for fake news detection tasks.