Social media like Twitter has been widely adopted for information dissemination due to its convenience and efficiency. However, false information and rumors on social media are undermining its utility as a valuable real-time information source. Existing works for information credibility analysis are based on offline batch analysis, often incurring a long lag since the event first occurs. In this paper, we develop a generative probabilistic model for real-time event credibility prediction in Twitter. We propose an online prediction algorithm based on streaming tweets, without storing or reprocessing the past tweets. We evaluate both the offline batch prediction and online streaming prediction performance of the proposed model on the Twitter dataset. The empirical results show that its batch prediction performance outperforms other algorithms based on aggregation analysis, and the online prediction performance quickly approaches that of the batch prediction with only a few hundred tweets.

Online social networks are becoming important platforms where users make social connections and share information. However, they are vulnerable to malevolent activities by malicious users. Hence, it necessitates effective automatic methods to predict user trustworthiness. The existing works mostly predict the trustworthiness of individual users separately from other users, ignoring the fact that users are related to each other through online social relationships. In this paper, we propose a probabilistic model based on Pairwise Markov Random Field (PMRF) that takes into account both user features and social relationships. In addition, we apply the Belief Propagation (BP) algorithm to perform inference efficiently in PMRF. The complexity of the algorithm grows only linear in the number of users. The experiment results on the Twitter datasets show that the proposed PMRF model can effectively exploit the social relationships to significantly improve the prediction performance.

Social recommender systems exploit the historic user data as well as user relations in the social networks to make recommendations. However, users are increasingly concerned with their online privacy, and hence, they are not willing to reveal their personal data to the general public. In this paper, we propose a social recommendation algorithm for top-N recommendation using only implicit user preference data. In particular, we model users' consumption behavior in the social network with Bayesian networks, using which we can infer the probabilities for items to be selected by each user. We develop an Expectation Propagation (EP) message-passing algorithm to perform approximate inference efficiently in the constructed Bayesian network. The original proposed algorithm is a central scheme, in which the user data are collected and processed by a central authority. However, it can be easily adapted for a distributed implementation, where users only exchange messages with their directly connected friends in the social network. This helps further protect user privacy, as users do not release any data to the public. We evaluate the proposed algorithm on the Epinions dataset, and compare it with other existing social recommendation algorithms. The results show its superior top-N recommendation performance in terms of recall.

Recommender systems have been widely used in e-commerce websites to suggest items that meet users' preferences. Collaborative filtering, which is the most popular recommendation algorithm, is vulnerable to shilling attacks, where a group of spam users collaborate to manipulate the recommendations. Several attack detection algorithms have been developed to detect spam users and remove them from the system. However, the existing algorithms focus mostly on rating patterns of users. In this paper, we develop a probabilistic inference framework that further exploits the target items for attack detection. In addition, the user features can also be conveniently incorporated in this framework. We utilize the Belief Propagation (BP) algorithm to perform inference efficiently. Experimental results verify that the proposed algorithm significantly improves detection performance as the number of target items increases.

Recent studies have shown the existence of considerable amount of packet-level redundancy in the network flows. Since application-layer solutions cannot capture the packet-level redundancy, development of new content-aware approaches capable of redundancy elimination at the packet and sub-packet levels is necessary. These requirements motivate the redundancy elimination of packets from an information-theoretic point of view. For efficient compression of packets, a new framework called memory-assisted universal compression has been proposed. This framework is based on learning the statistics of the source generating the packets at some intermediate nodes and then leveraging these statistics to effectively compress a new packet. This paper investigates both theoretically and experimentally the memory-assisted compression of network packets. Clearly, a simple source cannot model the data traffic. Hence, we consider traffic from a complex source that is consisted of a mixture of simple information sources for our analytic study. We develop a practical code for memory-assisted compression and combine it with a proposed hierarchical clustering to better utilize the memory. Finally, we validate our results via simulation on real traffic traces. Memory-assisted compression combined with hierarchical clustering method results in compression of packets close to the fundamental limit. As a result, we report a factor of two improvement over traditional end-to-end compression.

In this paper, we develop a Belief Propagation (BP) algorithm for similarity computation to improve the recommendation accuracy of the neighborhood method, which is one of the most popular Collaborative Filtering (CF) recommendation algorithms. We formulate a probabilistic inference problem as to compute the marginal posterior distributions of similarity variables from their joint posterior distribution given the observed ratings. However, direct computation is prohibitive in large-scale recommender systems. Therefore, we introduce an appropriate chosen factor graph to express the factorization of the joint distribution function, and utilize the BP algorithm that operates in the factor graph to exploit the factorization for efficient inference. In addition, since the high degree at the factor node incurs an exponential increase in computational complexity, we also propose a complexity-reduction technique. The overall complexity of the proposed BP algorithm on a factor graph is linear in the number of variables, which ensures scalability. Finally, through experiments on the MovieLens dataset, we show the superior prediction accuracy of the proposed BP-based similarity computation algorithm for recommendation.