Obtaining demographic attributes of online users is of great significance for retail marketing, targeted advertisement and many other scenarios. Users' wanderings on various websites and applications contains user preference on different items, and can be leveraged to infer one's private attributes. Existing studies usually focus on manually defined features, relationships in online social networks, or modeling global user preferences. However, attribute inference from the most common behavioral data (e.g., browsing history, shopping cart) is recently overlooked, and still requires further research. In this work, we propose a Topic-Enhanced Attentive Recurrent Neural Network (TEA-RNN) model to capture both local neighborhood-based features (with attentive RNN) and global patterns (with topic model) within user behaviors, and apply multi-task learning mechanism with weighted losses to further leverage the latent relationships within demographics. Experimental results on real-world datasets demonstrates the effectiveness of TEA-RNN by comparing with several commonly used baselines.

Detecting local topics from social media is an important task for many applications, ranging from event tracking to emergency warning. Recent years have witnessed growing interest in leveraging multimodal social media information for local topic detection. However, existing methods suffer great limitation in capturing comprehensive semantics from social media and fall short in bridging semantic gaps among multimodal contents, i.e., some of them overlook visual information which contains rich semantics, others neglect indirect semantic correlation among multi-modal information. To deal with above problems, we propose an effective local topic detection method with two major modules, called IEMM-LTD. The first module is an image-enhanced multi-modal embedding learner to generate embeddings for words and images, which can capture comprehensive semantics and preserve both direct and indirect semantic correlations. The second module is an embedding based topic model to detect local topics represented by both words and images, which adopts different prior distributions to model multi-modal information separately and can find the number of topics automatically. We evaluate the effectiveness of IEMM-LTD on two real-world tweet datasets, the experimental results show that IEMM-LTD has achieved the best performance compared to the existing state-of-the-art methods.

When users in various web and mobile applications enjoy the convenience of recommendation systems, they are vulnerable to attribute inference attacks. The accumulating online behaviors of users (e.g., clicks, searches, ratings) naturally brings out user preferences, and poses an inevitable threat of privacy that adversaries can infer one's private profiles (e.g., gender, sexual orientation, political view) with AI-based algorithms. Existing defense methods assume the existence of a trusted third party, rely on computationally intractable algorithms, or have impact on recommendation utility. These imperfections make them impractical for privacy preservation in real-life scenarios. In this work, we introduce BiasBooster, a practical proactive defense method based on behavior segmentation, to protect user privacy against attribute inference attacks from user behaviors, while retaining recommendation utility with a heuristic recommendation aggregation module. BiasBooster is a user-centric approach from client side, which proactively divides a user's behaviors into weakly related segments and perform them with several dummy identities, then aggregates real-time recommendations for user from different dummy identities. We estimate its effectiveness of preservation on both privacy and recommendation utility through extensive evaluations on two real-world datasets. A Chrome extension is conducted to demonstrate the feasibility of applying BiasBooster in real world. Experimental results show that compared to existing defenses, BiasBooster substantially reduces the averaged accuracy of attribute inference attacks, with minor utility loss of recommendations.

The collaborative filtering (CF)-based recommender systems provide recommendations by collecting users’ historical ratings and predicting their preferences on new items. However, this inevitably brings privacy concerns since the collected data might reveal sensitive information of users, when training a recommendation model and applying the trained model (i.e., testing the model). Existing differential privacy (DP)-based approaches generally have non-negligible trade-offs in recommendation utility, and often serve as centralized server-side approaches that overlook the privacy during testing when applying the trained models in practice. In this paper, we propose PrivRec, a user-centric differential private collaborative filtering approach, that provides privacy guarantees both intuitively and theoretically while preserving recommendation utility. PrivRec is based on the locality sensitive hashing (LSH) and the teacher-student knowledge distillation (KD) techniques. A teacher model is trained on the original user data without privacy constraints, and a student model learns from the hidden layers of the teacher model. The published student model is trained without access to the original user data and takes the locally processed data as input for privacy. The experimental results on real-world datasets show that our approach provides promising utility with privacy guarantees compared to the commonly used approaches.

Local topic detection is an important task for many applications such as local event discovery, activity recommendation and emergency warning. Recent years have witnessed growing interest in leveraging spatio-temporal social media (eg. Twitter) for local topic detection. However, existing methods overlook the continuity of time and location, which is quite important and useful for local topic detection. For example, tweets posted at adjacent time and location should be considered correlated instead of isolated. To address this challenge, we propose a multi-layer heterogeneous network based embedding learner to preserve vicinity correlation as well as co-occurrence correlation, and map all the location, time, and keywords into a same latent space. Based on the heterogeneous network embedding, we develop a Bayesian mixture model to find local topics without specifying the number of topics in advance. Moreover, tweets are frequently updated, thus, we adopt an incremental update strategy to process continuous tweet stream in real time. The extensive experiments on real-world data sets demonstrate that our method outperforms the state-of-the-art existing methods.

Local topic detection from spatio-temporal social media (e.g. Twitter) plays an important role in many applications, therefore, it has attracted a surge of research attention in recent years. However, most existing studies consider time and location separately, they often assume tweets are correlated as long as their time is adjacent or their location is neighboring. But in reality, only tweets posted at both adjacent time and neighboring location tend to talk about the same local topic. To address this issue, we propose a network based embedding model to capture the correlation between time and location, and jointly model spatio-temporal information and semantic information together. This embedding model can ensure that the generated keyword embeddings are semantically coherent and spatio-temporally close. Based on the keyword embeddings, we present a novel topic model to obtain high-quality local topics. This topic model presumes that each tweet is represented by only one topic and takes the background mode of words into consideration to address the concise and noisy problem of Twitter. The experiments demonstrate that the effectiveness and efficiency of our method have been improved significantly compared to the state-of-the-art existing methods.

Detecting local topic from social media is an important task for many applications, such as local event discovery and activity recommendation. Recent years have witnessed growing interest in utilizing spatio-temporal social media for local topic detection. However, conventional topic models consider keywords as independent items, which suffer great limitations in modeling short texts from social media. Therefore, some studies introduce embedding into topic models to preserve the semantic correlation among keywords of short texts. Nevertheless, due to the lack of rich contexts in social media, the performance of these embedding based topic models still remain unsatisfactory. In order to enrich the contexts of keywords, we propose two network based embedding methods, both of which can generate rich contexts for keywords by random walks and produce coherent keyword embeddings for topic modeling. Besides, processing continuous spatio-temporal information in social media is also very challenging. Most of the existing methods simply split time and location into equal-size units, which fall short in capturing the continuity of spatio-temporal information. To address this issue, we present a hotspot detection algorithm to identify spatial and temporal hotspots, which can address spatio-temporal continuity and alleviate data sparsity. Finally, the experiments show that the performance of our methods has been improved significantly compared to the state-of-the-art methods.