利用多重模型進行社群網路之地點分類

Abstract

有鑒於社群媒體的盛行,人們現在非常熱衷於在各個社群媒體上分 享自己的日常生活,而這些使用者的使用特性又與使用的地點有著強 烈的關聯性,因此,如何去了解分析地點的特性便成了一個社群媒體 上很重要的問題,並且在許多的應用上都扮演著重要的角色。但是, 一個具有挑戰性的困難點在於社群網路上單獨一篇的文章發佈往往不 足以蘊含足夠的資訊來分析地點的性質,為了對抗這樣的瓶頸,我們 同時地使用了一個地點的複數則文章以及其對應的相關資訊來進行 地點類別的分類。我們在 Instagram 上蒐集了百萬則量級的文章發表, 同時也蒐集了這些文章的相對應關聯資訊,包含但不限於發表時間、 GPS、以及使用者的資料,並且使用了 Foursquare 的預先定義地點類 別來做為我們的分類類別(包含了食物、夜生活、辦公、學校等......), 我們也提出了一個突破社群媒體資料收集限制的標籤數據蒐集方法。 在這裏,我們運用了來自複數則發表文章的豐富資訊來訓練我們的多 重模型,並且在超過十則文章的地點上,即使在充滿雜訊的條件下, 也達到了 73.18% 的分類準確度,而我們不只提出一個相當穩固的分類 模型,同時,我們也運用了社群媒體上的關聯性,來提升非熱門地點 的分類準確度,並且得到了超過 15% 的相對準確度提升。

Due to the popularity of social media, people are willing to share their daily life in social media (e.g., Instagram, Facebook). Since the strong connection between location and users' behavior, understanding the properties of locations has been widely required in several applications in social media. However, it is a challenging problem to estimate location properties from single post. To tackle this problem, we first collect million-scale posts (photos) from Instagram including all the contextual information (e.g., time, GPS, user profile). We observe that we can utilize additional and pre-defined categories (location types) from Foursquare as ground truth, and leverage social features---multi-modality features from social media---for measuring the remaining location types. Based on the statistics from our collected Instagram and Foursquare locations, only 16\% locations in Instagram can map to the locations in Foursquare (i.e., known location types such as food, nightlife spot, residence). Therefore, in this work, we focus on location type classification by leveraging rich and informative posts belong to the location. We not only propose a robust model by engaging multi-modality features, but also overcome the quandary of posts scarcity for unpopular (new discovered) locations by applying social features. We conduct experiments on our collected million-scale dataset. Even by using noisy and diverse data from Instagram, we still achieve an outstanding 73.18% accuracy for locations with more than 10 posts. Meanwhile, for unpopular locations, we can further achieve 15% relative improvement.