Emotionally Representative Image Discovery for Social Events

With the emerging social networks, images have become a major medium for emotion delivery in social events due to their infectious and vivid characteristics. Discovering the emotionally representative images can help people intuitive-ly understand the emotional aspects of social events. Prior works focus on finding the most visually representative im-ages for the target queries or social events. However, the emotionally representative image should not only visually relevant with the social event, but also has a strong emotion-al appeal among people. In this paper, we propose an emo-tionally representative image discovery framework by jointly considering textual, visual and social factors. In particular, we build a hybrid link graph for images of each social event, where the weight of each link is measured by textual emotion information, visual similarity and social similarity. Then we propose the Visual-Social-Textual Rank (VSTRank) algo-rithm to calculate the importance score for each image, so that the emotionally representative images can be discovered under the constraint of textual, visual and social represen-tativeness. To evaluate the effectiveness of our approach, we conduct a series of experiments with 15 social events extract-ed from real social media dataset, and evaluate the proposed method with both quantitative criterions and user study