以排序演算法整合多異質社群網路使用者

Abstract

隨著科技和技術的發展，社群網路越趨成熟及普及，並已和人們的生活密不可分，無論是基本資料、活躍的時段、出現的地點或是與朋友間的互動等資訊，都可以透過社群網路獲得，進而了解一個人。然而，當人們使用社群網路的習慣從單一變成多元，要從單一社群網路全面了解一個人變得更加困難。因此近期，有些研究開始探討，如何將多社群網路上的帳號進行連結，預測這些帳號是否來自真實世界同一個人。透過帳號連結，服務提供者可以完整地了解使用者，提供更準確的服務以及推薦。本研究針對任兩個帳號進行三個面向的特徵選取，包括基本資訊比對、社群關係以及行為一致性，並且透過排序學習法搭配一對一配對的限制，來進行帳號配對的預測。最後以現實世界中的資料，兩個著名的社群網路 (Google+和Twitter) 進行實驗，並且以查準率、查全率、準確率以及F值進行評估，可以發現本方法超越許多現行的方法，最後我們也將分析特徵的影響力，以及效能提升的關鍵。

With the prevalence of mobile communication techniques, people now spend a lot of time diving in online social networks for various purposes. Social networks thus contain abundant information of users, such as active periods, emerging locations, and interactions with friends. However, as the services social networks provided are orthogonal, no single network comprehensively depicts a user. Recently, a number of researches start to discover the alignments between entities from different social networks. The discovered alignments are valuable as they reveal intentions of users from different perspectives and are helpful to service providers to offer customized services. In this paper, we investigate the alignment problem of users between different social networks. Three aspects of features including profile matching, social relationship and behavior consistency, and techniques of learning to rank with mapping constraints are applied. We resolve the class skewness problem which generally exists in social networks due to the lack of sufficient negative links by learning to rank. Extensive experiments based on two popular heterogeneous social networks (Google+ and Twitter) with evaluation metrics, precision, recall, accuracy and f-measure, are applied to illustrate performance of our method. Also analysis on the features and constraint gives practical implications for future research.