Abstract

With the rapid growth of public cloud offerings, how to design effective prediction models that provide appropriate recommendations for potential users has become more and more important. In dynamic cloud environment, both of user behaviors and service performance are sensitive to contextual information, such as geographic location information. In addition, the increasing number of attacks and security threats also brought the problem that how to protect critical information assets such as sensitive data, cloud resources and communication in a more effective and secure manner. In view of these challenges, we propose a privacy-preserving and sparsity-aware location-based prediction method for collaborative recommender systems. Specifically, our method is designed as a three-phase process: Firstly, two privacy-preserving mechanisms, i.e., a randomized data obfuscation technique and a region aggregation strategy are presented to protect the private information of users and deal with the data sparsity problem. Then a location-aware latent factor model based on tensor factorization is applied to explore the spatial similarity relationships between services. Finally, predictions are made based on both global and spatial nearest neighbors. Experiments are designed and conducted to validate the effectiveness of our proposal. The experimental results show that our method achieves decent prediction accuracy on the premise of privacy preservation.

Introduction

Recommendation has been a hot research topic with the rapid growth of cloud services [1-2]. Great efforts have been done both in industry and academia to develop effective prediction models for recommender systems, which mainly aim at exploiting available information to provide users with satisfying recommendations [3-4]. With the popularity of mobile applications and devices, most cloud services could be invoked everywhere [5]. Because of the dynamics of cloud environment, most cloud services become region-sensitive. Actually, user preferences, quality of service (QoS) and the popularity of services are all varying with the change of user’s geographic location. Location information plays an increasingly important role in both users’ behaviors and service performance, especially in dynamic cloud environment and real-world applications. Although there have been some researches focusing on studying location influence to recommendation models [6-8]. Most of them merely focused on the location influence on user preferences. Few work paid attention to the location influence on QoS performance of services. Compared with traditional internet services, QoS of cloud services is more sensitive to location due to the dynamics of their environment. Both of QoS of cloud services and user behaviors are usually changing over geographic location. Thus it is still a fundamental task for recommender systems to provide the most beneficial suggestions to potential users with the consideration of location information. Moreover, data sparsity is always a serious threat that deteriorates the performance of recommendation methods [9-10], where users may only use a small number of services and provide limited QoS records. Under a data-sparsity scenario, existing collaborative recommendation models fail to capture the similarity relationships between users or services effectively. Factorization technique has been a successful prediction model used in recommender systems and proved to be an effective way to address the data sparsity problem [11-12].