Online social networks are playing a great role in our daily life by providing a platform for users to present themselves, articulate their social circles, and interact with each other. Posting image is one of the most popular online activities, through which people could share experiences and express their emotions. Intuitively, there must exist a connection between images and their associated hashtags. In this paper, we focus on systematically describing this relationship and using it to improve downstream tasks. First, we use a two-sample Kolmogorov-Smirnov test on an Instagram dataset to show the existence of the relationship at a significance level of α = 0.001. Second, in order to comprehensively explore the relationship and quantitatively analyse it, we adopt a graph-based approach, utilising the semantic information of hashtags and graph structure among images, to mine meaningful features for both hashtags and images. At last, we apply the extracted features about the relationship to improve an image multi-label classification task. Compared to a state-of-the-art method, we achieve a 12.0% overall precision gain.
展开▼
