非穩定環境下之旁道信息推論

Abstract

在本篇論文中，我們研究下列的問題：在非穩定環境下，如何利用其他觀測到的異質變數預測目標時間序列。在資訊爆炸的時代，資料產生機制處於非穩定的狀態在許多情況下皆可遇見，尤其是受人類行為影響的情況下。因此，不斷更新並從資料中學習到最新的觀念(concept) 是不可避免的。基於通訊系統中的信道 (channel)，我們視不同時間序列的資料為不同旁道 (side channel) 產生的輸出，並利用消息理論來萃取其中包含我們所關心的目標的資訊。除了萃取資訊的機制，我們更進一步對其篩選，並將選出確實包含訊息的資訊結合以達到預測的目的。

In this thesis, we research into the problem: how to predict a time series variable with heterogeneous sources under time varying environment. In this information explosive era, the non-stationary issue of the data generating mechanism is ubiquitous, especially data with human activities involved, and making learned concept catch up the change is thus inevitable. Basing on the concept of communication channel, we view the different data sources as output from different side channels, and applying information theoretic methods to extract information about the target we are interested in from these channels. With extracted information, a wrapper type selecting mechanism for sifting out non-informative ones, followed by an information combining procedure for fusing the information.