Research On Differentially Private Point-of-interest Recommendation

Recommendation System can filter out a mass of useless information for users and provide user the potential interest of content. The recommendation algorithm in point-of-interest(POI) recommendation system considered, three factors including user preferences to POI, the recommended from user’s friends and the influence of geographic location: At first, user’s check-in history can be used to describe the user preference for POI, and find the most matched user who have similar preference in POI by item-based collaborative filtering, and then provide user POI within high similarity user’s history; secondly, in order to provide user the recommendations from their friend, we defined the weight of friendship between two user as combining of the proportion of common check-in and the proportion of common checked POI; at last, the influence of geographic location are as important as other two factors due to people always tend to check in their nearby POI although the distant POI are even more attractive to them.Recommendation System may have user’s sensitive information leaked out while collecting their information. Up to now, there exist two types of risk of privacy exposure. First, users may immediately expose their accurate location right after they shared their check-in activity to their friends due to the accurate location of POI’s; secondly, the potential eavesdropper who hide in user’s all friends can deduce user’s check-in activity on the basis of recommended items even if user chose to check in POI privately and did not share this activity. To solve those two kinds of attack model, we proposed two privacy preserving algorithms.1) The accurate location for each POI are obfuscated to a virtual circle, the real POI may located in anywhere within the virtual circle. Compare to the exist algorithm <k,s>-Privacy’s limitation, we proposed <r,h>-Privacy algorithm, and proved the effectiveness theoretically of this algorithm.2) In order to reduce the correct probability of the attacker’s inference, we have to make those parameters more average, namely the lower variance. Accordingly, the differential privacy approach with Laplace Mechanism is proposed to bring appropriate noise into the weight of friendship. At last, we proved that this approach literally protected user’s privacy.In the end of this work, we applied these two algorithm to recommendation system, and regarded it as a whole to carry out theoretical analysis, then the formula about the degree of privacy preserving was presented. Next, we carried out substantial comparative experiments, the results said that our privacy algorithm indeed largely protect user’s geographic information and check-in information, in the mean time, the effect on recommendation algorithm is nearly cost-free and negligible.