應用特徵擷取法找尋資訊市場中具影響性的事件

Influential Event Analysis of Information Markets Using Statistical Feature Selections

陳政彥(Cheng-Yen Chen)

指導教授 ： 陳建錦

國立臺灣大學/管理學院/資訊管理學研究所/碩士(2009年)

https://doi.org/10.6342/NTU.2009.01775

摘要

自從1990年代開始，資訊市場 (Information Markets) 已經被證實是一套有效的公共事務集體智慧整合工具。透過市場運作與交易機制，資訊市場可以蒐集與整合群眾的集體智慧，並將集體智慧轉換為市場價格，該價格反映了群眾對一公共議題的看好度與可行性，進而輔助議題決策者制定合理的公眾政策。對於每一個預測市場議題，我們可依其價格與時間的變化關係畫出一價格走勢圖，在本論文中，我們將提出一套自動化的方法來分析資訊市場價格走勢圖，藉由分析資訊市場價格波動與相關的新聞文件，我們可以找出影響議題看好度的重要新聞事件，這些重要的新聞事件將可以協助群眾及社會學者了解影響公共議題的重要因子且分析社會現象。
在我們的方法裡，我們會先找出一資訊市場價格走勢圖內顯著且重要的價格波動期間，接著我們分析該期間內與議題相關的新聞文件，我們會使用事件偵測與追蹤(topic detection and tracking)的相關技術來將新聞文件集結成許多內容相似的文件群，而每一文件群代表一新聞事件。最後，我們會使用一些統計上常用的特徵篩選法(feature selection)來挑選該期間最獨特的新聞事件來代表價格波動的影響因子。實驗結果證明，我們所提出的方法能有效的找出影響資訊市場價格波動的重要事件。此外，我們也實作出一套web-based的資訊市場分析系統，透過視覺化的操作介面，使用者可快速的瞭解一公共議題曾出現了哪些具影響性的新聞事件。

關鍵字

資訊市場 ； 重要事件偵測 ； 價格波動 ； 漸進式分群 ； 統計式特徵擷取法

並列摘要

Since the 1990s, information markets have proved effective in predicting the outcome of public issues. A public issue is represented by an information market and all its possible outcomes form the contracts of the market. Through the mechanisms of market trading and rewarding, information markets are able to collect public’s opinions and convert the opinions into prices. The price of a contract indicates the feasibility of an outcome from public perspectives. High prices indicate that the public recognizes the outcome. By contrast, low prices mean that the outcome may not be feasible from the public’s viewpoint. The prices thus can help governments and policy makers establish reasonable policies to public issues. For each contract, we collect its price in a daily basis and propose a method to analyze the price sequence. We associate fluctuations in the price sequence with news documents and identify news events that cause the fluctuations. The identified events can help the public or social science scholars comprehend influential factors of public issues and social phenomena.
In the proposed approach, we first identify time periods accompanying significant price fluctuations. Next, we apply techniques of topic detection and tracking to cluster news documents in the periods. A cluster groups content-similar news documents and represents a news event. Finally, statistical feature selection methods are employed to identify events highly associated with the periods. The events then represent the cause of the price fluctuations and are influential to the corresponding public issue. Experiments based on a real world dataset demonstrate that the proposed method can identify influential events of information markets effectively. We also develop a prototype system based on the proposed method. The system graphically labels influential events of information markets that help users comprehend the development of public issues easily.

並列關鍵字

information markets ； event detection ； price fluctuation ； incremental clustering ； statistical feature selection

參考文獻

[3] J. Berg, R. Forsythe, F. Nelson, and T. Rietz, “Results from a dozen years of election futures markets research,” Technical Report, University of Iowa, 2003.

[4] J. Berg, R. Forsythe, T. Rietz, “What makes markets predict well? Evidence from the Iowa Electronic Markets,” Understanding Strategic Interaction: Essays in Honor of Reinhard Selten, pages 444-463, 1996.

[6] J. Berg, F. Nelson and T. Rietz, “Prediction Market Accuracy in the Long Run,” International Journal of Forecasting, Volume 24, Issue 2, pages 285-300, 2008.

[7] G. Boyle, "A Primer on Information Markets," ISCR, Victoria University of Wellington, 2005.

[8] G. Caldeira, “Expert Judgment versus Statistical Models, Explanation versus Prediction,” Perspectives on Politics, Volume 2, Issue 04, pages 777-780, 2004.