Exploring heterogeneous social information networks for recommendation
- Publication Type:
- Issue Date:
A basic premise behind our study of heterogeneous social information networks for recommendation is that a complex network structure leads to a large volume of implicit but valuable information which can significantly enhance recommendation performance. In our work, we combine the global popularity and personalized features of travel destinations and also integrate temporal sensitive patterns to form spatial-temporal wise trajectory recommendation. We then develop a model to identify representative areas of interest (AOIs) for travellers based on a large scale dataset consisting of geo-tagged images and check-ins. In addition, we introduce active time frame analysis to determine the most suitable time to visit an AOI during the day. The outcome of this work can suggest relevant personalized travel recommendations to assist people who are arriving in new cities. Another important part of our research is to study how “local” and “global” social influences exert their impact on user preferences or purchasing decisions. We first simulate the social influence diffusion in the network to find the global and local influence nodes. We then embed these two different kinds of influence data, as regularization terms, into a traditional recommendation model to improve its accuracy. We find that “Community Stars” and “Web Celebrities”, represent “local” and “global” influence nodes respectively, a phenomenon which does exist and can help us to generate significantly better recommendation results. A central topic of our thesis is also to utilize a large heterogeneous social information network to identify the collective market hyping behaviours. Combating malicious user attacks is also a key task in the recommendation research field. In our study, we investigate the evolving spam strategies which can escape from most of the traditional detection methods. Based on the investigation of the advanced spam technique, we define three kinds of heterogeneous information networks to model the patterns in such spam activities and we then propose an unsupervised learning model which combines the three networks in an attempt to discover collective hyping activities. Overall, we utilize the heterogeneous social information network to enhance recommendation quality, not only by improving the user experience and recommendation accuracy, but also by ensuring that quality and genuine information is not overwhelmed by advanced hyping activities.
Please use this identifier to cite or link to this item: