Differential Privacy implemented on the Recommender's System
INTRODUCTION
A recommender system attempts to predict a user’s potential likes and interests by analyzing the user’s historical transaction data. Currently recommender systems are highly successful on e-commerce web sites capable of recommending products users will probably like. Collaborative Filtering (CF) is one of the most popular recommendation techniques as it is insensitive to product details. This is achieved by analyzing the user’s historical transaction data with various data mining or machine learning techniques, e.g. k nearest neighbor rule, the probability theory and matrix factorization.
OBJECTIVE
The aim of the project was to solve the problem of inferring an individual’s rating, especially for the neighborhood-based methods using some background information of the individual. For example, an adversary can infer the rating history of an active user by creating fake neighbors based on background information.
EARLIER PRIVACY APPROACHES AND THEIR DRAWBACKS
A collaborative filtering method employs certain traditional privacy preserving approaches, such as cryptographic, obfuscation and perturbation.Among them, Cryptographic is suitable for multiple parties but induces extra computational cost. Obfuscation is easy to understand and implement, however the utility will decrease significantly. Perturbation preserves high privacy levels by adding noise to the original dataset, but the magnitude of noise is subjective and hard to control. Moreover, these traditional approaches suffer from a common weakness: the privacy notion is weak and hard to prove theoretically, thus impairing the credibility of the final result. In order to address these problems, differential privacy has been proposed.
DIFFERENTIAL PRIVACY
Differential privacy was introduced into CF by McSherry et al., who pioneered a study that constructed the private covariance matrix to randomize each user’s rating before submitting to the system.Machanavajjhala et al presented a graph link-based recommendation algorithm and formalized the trade-off between accuracy and privacy. We will applying this approach in our recoomender system.
RESOLUTION
The process of differential privacy solves this problem by using the fact that user’s rating should be inferred from the entire database of users by weighing each user with their “similarity score” with the reference user. Thus user’s rating comes from the entire population rather than some predefined set of users. This has two advantages – first we can never identify the ratings of that individual user, second, it is computationally very effective.
OVERVIEW
Application : Recommender systems
Input data : User-item rating matrix
Output data : Prediction
Challenges : High sensitivity
Solutions : Adjust sensitivity measurement
Selected mechanism : Group large candidate set
QUICK LINKS