Recommending items in pervasive scenarios: models and experimental analysis

In this paper, we propose and investigate the effectiveness of fully decentralized, collaborative filtering techniques. These are particularly interesting for use in pervasive systems of small devices with limited communication and computational capabilities. In particular, we assume that items are tagged with smart tags (such as passive RFIDs), storing aggregate information about the visiting patterns of users that interacted with them in the past. Users access and modify information stored in smart tags transparently, by smart reader devices that are already available on commercial mobile phones. Smart readers use private information about previous behavior of the user and aggregate information retrieved from smart tags to recommend new items that are more likely to meet user expectations. Note that we do not assume any transmission capabilities between smart tags: Information exchange among them is mediated by users' collective and unpredictable navigation patterns. Our algorithms do not require any explicit interaction among users and can be easily and efficiently implemented. We analyze their theoretical behavior and assess their performance in practice, by simulation on both synthetic and real, publicly available data sets. We also compare the performance of our fully decentralized solutions with that of state-of-the-art centralized strategies.