| Abstract: | 網路提供給消費者大量資訊. 網路使用者在網路上對於各種產品,如電影,音樂,餐廳, 商 品等給予的評分,常常構成相當大的網路資料. 關於這類評分一分到五分或一分到十分 的資料,目前網路上常見的呈現方式是以各產品所獲得的平均評分以圖表表示(例如以 五個星號代表五分). 然而,因為沒有考慮到評分者與評分者之間的差異,這樣簡單的平 均分數可能不客觀公平. Ho and Quinn (2008) 提出一個貝氏模型並以MCMC方法估計其中參數,該模型有 納入評分者與評分者之間的差異. Ho and Quinn (2008)並且以網路資料舉例說明 他們方法比平均分數能夠更合理的解釋評分資料. 可是, 該二位作者於文章結尾也指出, 當資料量很大,甚至於當新資料進來而需要重新估計模型參數, 以MCMC方法來計算 於實際應用上是不可行的. 本研究計畫的目的就是提出一個有效可行的方法來解決這 個問題.
The internet has offered consumers with a vast amount of information. One growing area of such information is ratings by internet users on various kinds of products such as movies, music, restaurants, commodities, etc. Consider rating data in which each product was rated on a scale of 1 to 5 by internet users. The current displays of each product`s preference are typically based on “average rating,” but it is well known that the average rating method ignores systematic differences across raters. Ho and Quinn (2008) proposed a Bayesian model and Markov chain Monte Carlo (MCMC) methods to take into account systematic differences across raters, and at the same time incorporate statistical uncertainty in the ratings. However, to work efficiently on an industrial scale and to adjust the parameters in real-time as new rating arrive, the MCMC methods may not be computationally feasible. The current project aims to provide a solution to this problem by using efficient approximation algorithm. |
