Research On Differentially Private Algorithms For Product Recommendation

With the growth of data volume and the development of Internet commodity economy,the importance of recommendation algorithms is self-evident.Recommendation algorithms aim to provide users with personalized recommendations using many techniques like data mining.However,the product recommendation often needs to use users'feedback on products,which makes recommendation algorithms under the risk of leaking users'privacy information.In the field of privacy preserving,differential privacy has become a universal standard because of its high utility.In terms of the following two specific problems,this paper studies differentially private algorithms in product recommendation.First,privacy preserving in frequent itemset mining.In the product recommendation,when consumption data is abundant,it is often necessary to mine frequent itemsets to discover potential correlations between products,where the historical consumption and evaluations for products by users are private information.We study the problem of mining frequent itemsets in high-dimensional datasets under the requirement of protecting users' differential privacy,and propose the PrivBUD-Wise algorithm.Unlike traditional algorithms,PrivBUD-Wise does not reduce the dimensions of datasets by truncating or splitting,because that will cause additional information loss.PrivBUD-Wise does not make any changes to original datasets,and consumes privacy budgets in a better way to improve the utility.We propose a new differential privacy mechanism:SRNM,and provide it with a strict mathematical proof.In addition,PrivBUD-Wise firstly proposes a biased privacy budget allocation strategy,which enables the algorithm to take full advantage of the characteristics in frequent itemset mining problems.The comparative experiments on three real-world datasets verify the good performance of PrivBUD-Wise.Second,privacy preserving in multi-armed bandits.In the product recommendation,when the consumption data is not abundant,the recommendation for new users or new items needs to use reinforcement learning algorithms based on multi-armed bandit algorithms,where users' feedback on products is privacy information.Based on real-world application scenarios and stochastic bandit problems with side observations,we consider the situation of having side rewards in stochastic multi-armed bandit problems,and propose the UCB-Side algorithm with guaranteed upper bound for the regret.When introducing differential privacy to the UCB-Side algorithm,we propose the DP-UCB-Side algorithm based on traditional privacy protection technology,and then propose its improvement scheme:the DP-UCB-INT-Side algorithm.Through a large number of comparative experiments,we verify the effectiveness of the UCB-Side algorithm and the DP-UCB-INT-Side algorithm.