Utility-Based Differentially Private Recommendation System.

Area of Science:

• Computer Science
• Data Science
• Artificial Intelligence

Background:

• Recommendation systems depend on user data, raising privacy concerns that deter users.
• Existing privacy-preserving methods often sacrifice recommendation accuracy (utility).
• A balance between high privacy and high utility is crucial for recommender system adoption.

Purpose of the Study:

• To develop a robust recommendation system that offers both high accuracy and strong user privacy.
• To address the limitations of existing private model-based collaborative filtering algorithms.
• To improve the utility of privacy-preserving recommendation techniques.

Main Methods:

• Proposed an enhanced private model-based collaborative filtering algorithm.
• Combined imputation of missing ratings using 'l'-injection with differential privacy.
• Introduced a random differential privacy approach to Alternating Least Squares (ALS) for enhanced utility.

Main Results:

• Identified data sparsity as a key vulnerability in recommender frameworks.
• The proposed private noisy Random ALS algorithm demonstrated superior performance over non-noisy ALS.
• Experimental results on benchmark datasets confirmed improved privacy and utility.

Conclusions:

• The novel approach effectively balances privacy and utility in recommendation systems.
• The private noisy Random ALS algorithm offers a promising solution for secure and accurate recommendations.
• Addressing data sparsity is vital for enhancing privacy in recommender frameworks.