一個應用矩陣分解的即時QoS預測法

Abstract

一個現有的應用常會藉由已存在的服務去組成。而一般來說，通常會採用QoS的預測值來進行挑選合適的服務，現今已有許多不同的預測QoS方法被提出來。然而當新產生的QoS數據被平台蒐集時，這些預測方法所推論的QoS預測值大多都無法更快速的產生或有效率的更新。本文中，我們提出一個基於時間的QoS預測方法，應用矩陣分解方法的定期背景更新，同時利用並結合線上平台的優勢，在時間的進行下，當更多新的資料持續的被蒐集時，系統變會即時去做部分優化。最後我們會參考時間變化下數據變化的趨勢，來決定並選擇一個較佳的QoS預測值，來提升整體預測的準確度。而實驗數據顯示出我們的方法在加入時間的考量因素後，在系統運行一段時間後，準確度會慢慢提高並逐漸優於其他現有的QoS預測方法。

An application (service) can be composited with many existing services. Generally, an appropriate service might be selected according to the predicted QoS values, and there are several approaches of which each introduces a distinct prediction method. However, the QoS values derived by these approaches may not be generated quickly or update efficiently when new QoS records is collected. In this thesis, we proposed a time-aware QoS prediction method. The method applies matrix factorization approach to update the QoS matrix periodically. And another part takes the advantage of the online platform, by using the new collected QoS records to partially update some QoS values in the predicted matrix in last time slice. Then these two different QoS predicting methods may generate different results. By considering the trend of past QoS records, a better predicted QoS can be chosen and save to the database. The experiment results show that when data collected in different time slices is considered, the accuracy of QoS values derived by our method is gradually better than other approaches.