探勘 Instagram 社群使用者移動模式

Abstract

近年來，照片分享社群如Instagram、Flckr、Pinterest等等蓬勃發展，因此，我們提出一套方法，利用模糊C均值分群演算法(FCM)，探勘Instagram上的使用者移動模式。我們提出的方法共包含4個步驟，首先，我們在FCM中加上地點加權的概念，將鄰近的地點叢集在一起成為一個群集，我們稱每個群集為一個地標，並將每個地點轉換成一個空間的特徵向量。接著，我們使用隱含狄利克雷分佈模型，將使用者貼文內的標籤分成多個主題，並將每個地點轉換成語意的特徵向量。然後，我們使用動態規劃方法，計算使用者間移動路徑間的距離，並修改FCM將使用者移動路徑分成不同群集。最後，我們利用第一、二步的結果，將使用者移動路徑轉換成地標和主題的序列，並修改FCM將轉換後的序列分成不同群集。也就是說，我們將使用者移動路徑分成兩個層級的群集，在第三步，我們利用空間和語意的特徵向量，將使用者移動路徑分群，我們稱之為地點層級；而在第四步，我們則利用轉換後的序列來分群，我們稱之為地標層級。地點層級的分群結果，可以幫助我們分析使用者細部的移動方式，而地標層級的分群結果，則可以讓我們了解使用者概觀的移動方式。實驗結果顯示，我們所提出的方法可以有效地探勘出使用者在Instagram上的移動模式。

Photo sharing social networks have drawn considerable attention recently, such as Instagram, Flickr, VSCO, and Pinterest. Therefore, in this study, we propose a framework to mine users’ movements on Instagram based on the fuzzy c-means method (FCM). The proposed method contains four phases. First, we incorporate the concept of weighted locations into the FCM method to group neighboring locations into a cluster (or landmark), and then convert each location into a spatial feature vector. Second, we apply the Latent Dirichlet Allocation (LDA) model to cluster the hashtags into location topics, and then convert each location into a semantic feature vector. Then, we transform each trajectory into two sequences of spatial and semantic feature vectors. Third, we apply a dynamic programming approach to compute the distance between converted trajectories, and modify the FCM method to cluster them together. Finally, we transform each trajectory into a sequence of the landmark and topic identifiers obtained in the first and second phases and then further modify the FCM method to cluster transformed sequences together. That is, we cluster trajectories together by spatial and semantic feature vectors in the third phase (location layer), and then cluster them together by transformed sequences in the last phase (landmark layer). The clusters formed in the location layer can provide us to analyze users’ movements in a fine-grained view while those in the landmark layer can complement movement analyses in a coarse-grained view. The experiment results show that the proposed framework can effectively find users’ movements on Instagram.