Autoencoder-based privacy-preserving collaborative filtering

Privacy-preserving collaborative filtering systems are effective approaches that provide personalized recommendations to users without violating their privacy. However, data disguising approaches based on randomized perturbation techniques cause corruption in genuine user data and negatively affect the system's success in generating recommendations. Especially in data disguising processes to achieve high privacy levels, the losses in recommendation accuracy are quite high. In this study, an autoencoder-based recommendation generation approach was used to alleviate the recommendation accuracy losses of genuine user ratings disguised by randomized perturbation techniques. The recommendation accuracies of disguised data sets produced at different privacy levels using randomized perturbation techniques on the data set produced with Book-Crossing as a reference were analyzed with traditional memory-based neighborhood algorithms and autoencoder-based prediction generation approaches. The autoencoder-based prediction generation system was tested with various hidden layer numbers (2, 3, and 4) and activation functions (tanh, elu, selu, and linear). As a result of experimental studies, it was shown that the autoencoder-based prediction generation approach on the disguised data set using randomized perturbation technique significantly increases the prediction accuracy compared to traditional neighborhood-based approaches. When the mean absolute error values were examined via varying privacy levels, the error values at the lowest and highest privacy levels were 1.395 and 2.249, respectively, in the traditional neighborhood-based collaborative filtering approach, while these values were 1.208 and 1.313 in the autoencoder-based collaborative filtering approach. In conclusion, the autoencoder-based collaborative filtering system can better tolerate the distortions that occur in the data set due to increasing privacy levels and can produce high-accuracy recommendations when high privacy levels are provided to the user.