以帶標題資訊層疊推論影響力及偏好之方法

Abstract

隨著使用者愈來愈樂意在社群網路中分享資料，商人也持續地從可獲取的鉅量資料中來對潛在消費者試著做分析。其中一個方式是經由觀察使用者頻繁分享的內容來推論他們的主題偏好，故個人化的廣告得以實施；另一種想法是建模探討使用者之間的影響力關係，以達到最大程度的病毒式行銷。本論文所提出在資訊層疊上的研究可以同時適用於前述兩項應用。其關鍵的概念是，一個資訊是否會被該資訊的接收者所分享，端看於該資訊傳送者與接收者之間的影響力關係，以及接收者對於該資訊的主題偏好。在這篇論文中，我們提出一個聯合優化問題以帶標題的資訊層疊推論影響力及使用者偏好，並展現其相對於相關文獻上的優勢。在合成資料及真實資料上的實驗評估中，皆顯示了本模型對於未看過的資訊層疊資料有更好的預測能力。

As data shared between users keep exploding in online social network, marketers keep trying to get more information about their potential customers over the available big data. One way is to infer the preference of users from their sharing/visiting content so that proper advertisement could be given, and another idea is to model the influence between users to maximize the viral marketing power. We argue that our approach to the study of information cascades could satisfy both applications at the same time. The main point here is whether a piece of information will be shared by a receiver is jointly determined by (a) the influence between the sender and the receiver, and (b) the receiver's preference to the particular topic of the received information. In this thesis, we propose a joint optimization problem on inferring influence and users' preference through topic-sensitive information cascades, and demonstrate several advantages over the related literature. Experiments with synthetic data and real data show that out model has better predicting power on unseen information cascades.