Improving the Accuracy for Privacy-Preserving Point-of-Interest Recommender Systems
Location
SU 214
Department
Computer Science
Abstract
In the digital era, an essential ingredient of numerous online vendors and various types of websites is a recommender system. This technology immensely reduces users’ search time when looking for the contents of their interest. Recommender systems make suggestions based on individual preferences learned from their users. Generally speaking, users have to unconditionally share personal search and/or purchase history with service providers, who also have full access to their private preferences. In this research, we have implemented a privacy-preserving point-of-interest recommender system based on a framework with three major components: a mobile app IncogniToGo, a vendor-hosted aggregate server, and a remote central server. Furthermore, we improved the system’s prediction accuracy by estimating each anonymous user’s GPS location and incorporating this information into the recommendation process. Our current model integrates the Google Cloud Platform (Maps and Firebase), a wireless communication standard called Wi-Fi Direct, and machine learning algorithms. IncogniToGo allows a user to rate a place in Google Maps. Internally, it computes the user’s current location based on their rated places and communicates this data using a random user ID via Wi-Fi Direct to the aggregator server. User groups are created on the aggregator server, and the corresponding group preferences are then sent to the Firebase (central server). Machine learning algorithms are performed on the server to extract latent features of the shared data. Finally, IncogniToGo pulls such features from Firebase and generates personalized recommendations locally on the user’s device, which prevents the server from learning users’ individual preferences.
Faculty Sponsor
Xiwei Wang, Northeastern Illinois University
Improving the Accuracy for Privacy-Preserving Point-of-Interest Recommender Systems
SU 214
In the digital era, an essential ingredient of numerous online vendors and various types of websites is a recommender system. This technology immensely reduces users’ search time when looking for the contents of their interest. Recommender systems make suggestions based on individual preferences learned from their users. Generally speaking, users have to unconditionally share personal search and/or purchase history with service providers, who also have full access to their private preferences. In this research, we have implemented a privacy-preserving point-of-interest recommender system based on a framework with three major components: a mobile app IncogniToGo, a vendor-hosted aggregate server, and a remote central server. Furthermore, we improved the system’s prediction accuracy by estimating each anonymous user’s GPS location and incorporating this information into the recommendation process. Our current model integrates the Google Cloud Platform (Maps and Firebase), a wireless communication standard called Wi-Fi Direct, and machine learning algorithms. IncogniToGo allows a user to rate a place in Google Maps. Internally, it computes the user’s current location based on their rated places and communicates this data using a random user ID via Wi-Fi Direct to the aggregator server. User groups are created on the aggregator server, and the corresponding group preferences are then sent to the Firebase (central server). Machine learning algorithms are performed on the server to extract latent features of the shared data. Finally, IncogniToGo pulls such features from Firebase and generates personalized recommendations locally on the user’s device, which prevents the server from learning users’ individual preferences.