運⽤社群媒體並具效率性之顯著趨勢預測:以股市分析為例

在各個領域中，預測商品未來的趨勢都具有⾼度的應⽤價值，不論是在股市中預測股價會突然上漲的股票，或者是預測在電商中哪些商品的銷售量會有⼤幅成⻑，都能為我們帶來可觀的效益。⽽⽬前的研究中，已經證實了社群媒體可以反映出現實世界的事件影響以及⼤眾的情緒意⾒。並且這些特徵與其他商品的變化有⾼度相關性，包含股價起伏、⾳樂與書籍銷售量以及電影票房，因此在預測商品趨勢的研究中，也有部分研究已經結合社群媒體來提升效能，然⽽，在⽬前的商品趨勢研究中，⼤部分研究都僅預測趨勢⽅向。僅知道趨勢⽅向⽽沒有變化強度，在應⽤上的價值就會受限，因此在本論⽂中，我們設計了⼀套具⾼效率的顯著預測趨勢架構，可以有效的結合社群媒體資料以及商品的歷史資料，同時我們藉由分析這些社群媒體的特徵，可以快速地找出潛在的顯著趨勢之商品，就能提前對於預測商品進⾏處理，有效的過濾那些不太可能有顯著趨勢之商品。經由⼀系列真實資料的實驗評估，整體來說本研究所提出的⽅法在預測準確度及效率上都可顯著優於其他⽅法，驗證了我們的模型在商品顯著趨勢之預測上可以同時達到準確性及效率性之要求。

關鍵字

趨勢預測；社群媒體；分類問題；意⾒探勘

並列摘要

In many fields, predicting future trends of items is highly valuable. Predicting which stock prices will suddenly rise on stock market, or which sales of products in commerce will grow, brings us high benefit. A number of researches have already shown that the characteristics of social media are highly relevant to the trends of items in the various domains, like stock prices, book sales, and movie box office revenue. They predict trends of items with social media data to improve accuracy. However, their goals are to predict the direction of the movement during the given period. In actual practice, only knowing the direction of the trend without the variation will limit the value. Therefore, in this thesis, we design a novel framework that can efficiently predict significant trends and effectively combine historical data of items with social media data. We can extract features from social media and then quickly find out the potential items which have significant trends. It reduces the unnecessary computation and achieves efficiency. Through empirical evaluation on real data, our proposed method outperforms than the other methods on efficiency and effectiveness. The result indicates that in the prediction of significant trends, our method can achieve both of efficiency and effectiveness at the same time.