應用向量自回歸模型於多媒體平台相互影響之研究

Abstract

隨著網路與資訊科技的盛行，大量的資料即時的在網路上流竄，使得使用者能以簡單快速之方法獲取最新消息。近年來因社群媒體之興起，有越來越多的使用者藉由社群平台即時分享或修改內容來抒發情緒，然而使用者於平台之選用上常會因不同類型事件而選擇不同的平台，使得各社群平台擁有其各自不同之類型。故本研究希望提出一個有效的方法來對平台彼此互動狀況進行探討，藉此找出平台間相互影響之特性。 本研究以向量自回歸模型 (Vector autoregression model) 為基礎，針對社群時間序列中含有每日季節循環樣式 (Pattern) 之問題以片段選取落後期數之方式，發展出 Vector autoregression with Segmented lags (VarSeg) 模型，期望藉由此模型能提供較有效之預測與分析能力於處理時間序列資料中含有季節因子之問題。 本研究針對三大主流網路平台 Twitter、Wikipedia 與 Google 之資料進行蒐集，並以每小時為單位來建置議題之時間序列，最後從三平台中選擇了 351 個趨勢議題來進行實驗。實驗結果證明 VarSeg 能夠對於資料當中含有當日循環樣式之議題提供良好的變數選用方法，並且在格蘭傑因果關係檢定上亦證明，一平台過去事件發生之情況，能夠用於預測其他平台事件未來之走勢。於衝擊反應函數之分析亦顯示各平台對於外部衝擊到完全反應的時間差大約 2 ~ 14 小時。另外，藉由與 VAR 模型之預測比較，發現 VarSeg 於議題之預測上，能降低預測的誤差值。故本實驗建構之 VarSeg 模型可作為往後研究多變量時間序列之基礎，並且可將其用於處理資料中含有季節性樣式之問題。

With the popularity of Internet and IT technology and a large amount of real-time information which diffuses in the Internet, users can obtain the latest news in a simple way. In recent years, many users use social media to share and edit the content that reflects their attention and thoughts. However, users often choose to interact with different platforms depending on specific triggering events. Therefore each platform has their own trending topic type. In our work, we develop an effective way to examine how social media platforms interact with each other and find the interplay between one platform and the other. Our research model is constructed based on the vector autoregression Model. We modified the model to handle the intraday cyclical patterns in social media time series and developed the vector autoregression model with segmented lags model (VarSeg). Our experimental results based on 351 trending topics from the three platforms suggest that VarSeg is able to capture the intraday dynamics of trending topics. Moreover, our Granger causality tests provide stronger empirical evidence that historical time series in one platform can help predict the future traffic on another platform. In addition, impulse response functions show that the time delay to fully response to external shock is around 2 to 14 hours. After applying VarSeg to predict future values, we compare with standard VAR, the results show that VarSeg provided reduced MSE for event predicting.