Split learning (SL) aims to protect user data privacy by distributing deep models between the client-server and keeping private data locally. In SL training with multiple clients, the local model weights are shared among the clients for the local model updates. This paper first reveals data privacy leakage exacerbated by local weight sharing among the clients in SL through model inversion attacks. Then, to reduce the data privacy leakage issue, we propose and analyze privacy-enhanced SL (P-SL) (or SL without local weight sharing). We further propose a parallelized P-SL to expedite the training process by duplicating multiple server-side model instances without compromising accuracy. Finally, we explore P-SL with late participating clients and devise a server-side cache-based training method to address the forgetting phenomenon in SL when late clients join. The experiment results demonstrate that P-SL helps reduce up to$50\%$of client-side data leakage, which essentially achieves a better privacy-accuracy trade-off than the current trend by using differential privacy mechanisms. Moreover, P-SL and its cache-based version achieve comparable accuracy to baseline SL under various data distributions, while incurring lower costs for computation and communication. Additionally, caching-based training in P-SL mitigates the negative effect of forgetting, stabilizes learning, and enables practical and low-complexity training in a dynamic environment with late-arriving clients.

In this paper, we aim at developing a new collaborative filtering recommender system using soft ratings, which is capable of dealing with both imperfect information about user preferences and the sparsity problem. On the one hand, Dempster-Shafer theory is employed for handling the imperfect information due to its advantage in providing not only a flexible framework for modeling uncertain, imprecise, and incomplete information, but also powerful operations for fusion of information from multiple sources. On the other hand, in dealing with the sparsity problem, community context information that is extracted from the social network containing all users is used for predicting unprovided ratings. As predicted ratings are not a hundred percent accurate, while the provided ratings are actually evaluated by users, we also develop a new method for calculating user-user similarities, in which provided ratings are considered to be more significant than predicted ones. In the experiments, the developed recommender system is tested on two different data sets; and the experiment results indicate that this system is more effective than CoFiDS, a typical recommender system offering soft ratings.

In this paper, we address the problem of combining information in recommender systems (RSs) based on Dempster-Shafer theory (DST). We first discuss the characteristics of this problem, and then analyze six popular combination methods in the context of RSs. Based on the analysis, we propose two new mixed combination methods which can be considered as useful tools for fusing information in the systems. To evaluate the proposed methods, we integrate them into a typical RS based on DST, and then measure recommendation performances on MovieLens data set. The experimental results show that, comparing to the baselines, the new methods outperform with regards to DS-MAE and DS-Recall, and can be comparable in terms of DS-Precision and DS-F1.

This paper proposes a new method for combining information in recommender systems based on Dempster-Shafer theory. Within this method, focal elements whose probabilities are less than or equal to an infinitesimal threshold are considered as noise that may be caused by the process of fusing information, and then eliminated. Comparing with two baselines, known as 2-points focused and 2-probabilities focused combination methods, the new method is more effective because it does not eliminate focal elements whose probabilities are not very small. Moreover, the new method is tested in a wide range of experiments on MovieLens data set, and the experimental results indicate that the proposed method is much better than the baselines in accuracy of recommendations.