微網誌平台下推薦朋友予組織之研究

Abstract

自2009年起，台灣吹起一陣微網誌風潮，數以萬計的人加入噗浪，並在上面交友、分享資訊及閒聊八卦。噗浪儼然成為一個資訊流通的高速公路。一件大眾有興趣的事情，可以在一天內流傳到全台灣。例如戴爾公司於2009年6月發生網站標錯商品價格的事件，噗浪就扮演一個推波助瀾的角色，將資訊快速傳播到所有人手上。

對台灣組織來說，噗浪是一個達成組織目的之前哨站。防禦方面，可以快速偵測緊急突發狀況，以期有更多時間處理。主動方面，可以在噗浪認識更多的客戶或朋友，並且藉由互動獲得第一手即時資訊。

本論文首先對噗浪做特性分析，著重在情感、語言類別、和即時性主題上。接著建造一組個人化組織推薦系統，依照每一個組織的特性，抽取合適於組織的特徵，並且將這些特徵回饋給組織當作推薦的理由，來輔助組織做出交友決策。此外，也提出一個組織社群的概念，讓實作此系統時，不需擁有整個噗浪使用者的資料，只需要組織社群範圍的使用者資料就可以開始運作，而組織社群限制還可以提升推薦效能。 本論文使用特徵選擇機制，對每家組織選擇個人化組織特徵，並且使用支持向量機(Support Vector Machine)做訓練和預測，並由不同角度對推薦效能進行分析比較，例如特徵選擇機制是否有效能變化、考慮機器人服務對推薦的影響、以及組織社群效應等。最後提出一些議題供未來研究，包含推薦理由的設定和推薦系統即時化。

Microblog becomes popular in Taiwan from 2009. There are tens of thousands of users to join plurk. They make friends, share information and gossip with one another in plurk. Plurk which can be regarded as an information highway can rapidly spread news. For example, Dell marks the wrong price to products in official homepage in June 2009 and plurk plays an important role in information dissemination. For organizations in Taiwan, plurk is an outpost to achieve organizational purposes. In one way, plurk can detect and deal with emergency events quickly and unhurriedly. In another way, organizations can know more customers and friends, and obtain the first-hand real-time information by interaction with them. This thesis analyzes properties in plurk at first including emotional topic, linguistic topic and real-time property. Then it builds a personalized friend recommendation system which selects suitable features for each organization and those features will form recommendation reasons to support making friends. Besides, this thesis also presents a concept of organization community. That can make the implementation feasible for each organization. This thesis proposes a feature selection mechanism to select the customized features and uses Support Vector Machine (SVM) to train and predict data. Furthermore, this thesis makes experiments to discuss the influences of feature selection, robot services, and community in recommendation. Finally this thesis provides some issues for future works such as recommendation reasons and real-time recommendation.