運用搜尋引擎查詢紀錄之景氣監測系統

Abstract

景氣指數及指標常被用來監控景氣週期轉換的情形。普遍來說，這些指數及指標是由許多經濟變數所組成，而這些不同的變數都是由不同政府部門來彙整。為了整合這些變數，必須經由大量複雜程序處理，造成景氣周期監控的延遲發佈。在這篇研究中，我們提出一個新的景氣周期監控系統來對景氣周期做預測，主要利用搜尋引擎上的關鍵字查詢記錄來建模型。為了要找出和景氣周期有高度相關的關鍵字，我們提出一個有效的特徵選取及過濾的方法。被選取的關鍵字及其查詢次數先做整合，接著為景氣周期的狀態做分類。為了要降低查詢次數造成的稀疏問題，我們導入離散化方法改進。 實驗主要是根據行政院經濟建設委員會發佈的五年資料集做測試，結果也顯示我們提出的系統有將景氣周期分類正確，而且所選取的關鍵字也反映出部分人類行為。和過往利用經濟變數的方法相比，因為關鍵字查詢記錄可以即時的從網路上取得，我們的系統能提供更及時的景氣周期資訊。

Business indices and indicators are used to monitor the regime shifts of business cycles. Generally, the indices and indicators are comprised of various economic variables that are compiled by different government departments. The compilation of the variables involves a great deal of data processing operation, which delays the monitoring of business cycles. In this paper, we propose a novel business cycle surveillance system that utilizes the query logs of search engines for business cycle modeling. The system employs an effective feature selection and pruning technique to identify query terms that are representative of business cycles. The selected terms and the frequency count of queries associated with the terms are then integrated to classify the status of business cycles. We use data discretization techniques to reduce the sparseness of query frequencies. Experimental results based on a five-year dataset show that the proposed system can classify the status of business cycles accurately, and the selected query terms reveal interesting human behavior patterns in different business cycles. Unlike economic variables, query logs are readily available through online Web services, so our system can provide business cycle information in a timely manner.