Travel user interest discovery from visual shared data in social networks

• Published in: 2017 Sudan Conference on Computer Science and Information Technology (SCCSIT) pp. 1 - 7
• Main Authors: Yassin, Fatima Mohamed, Lazzez, Onsa, Ouarda, Wael, Alimi, Adel M.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-11-2017
• Subjects: CNN; Computer architecture; Facebook; Object recognition; Ontologies; Ontology; Social networks; social visual information; travel user interest; Visualization
• DOI: 10.1109/SCCSIT.2017.8293057

Social networks are virtual environments where users express their opinions, preferences and interests. All shared information in Social Networks like Facebook, Twitter, LinkedIn can be very useful to understand citizen's interest. The visual shared data in social networks can express many latent knowledge about user' interests in such topics especially the travel topic. This study aims to discover the travel user interest through the huge number of social images shared daily. In this paper, we propose a comparison between approaches based on Feedforward learning of Convolutional Neural Network (CNN) architectures GoogleNet and VGG'19 trained on Places365 Dataset for visual object Recognition. Once objects are recognized in images, we propose an Ontology based decision system for travel user interest prediction. We have constructed a new database of shared images in Sudanese and Tunisian Facebook accounts. We achieve classification rates of 93% and 87% using respectively GoogleNet and VGG'19 CNN architectures on Sudanese Facebook users compared to the opinion of the expert 78%, and classification rates of 80% and 77% using respectively GoogleNet and VGG'19 CNN architectures on Tunisian Facebook users compared to the opinion of the expert 82%. The travel interest is hard task done by expert then our approaches based on CNN, GoogleNet and VGG'19 architectures can facilitate the interest of travel topic make it easy to discover is the user interest in travel or not, other thing the GoogleNet architecture leading to a better performance and improving the classification accuracy than VGG'19.