Search

Embedding knowledge graphs (KGs) into continuous vector space is an essential problem in knowledge extraction. Current models continue to improve embedding by focusing on discriminating relation-specific information from entities with increasingly complex feature engineering. We noted that they ignored the inherent relevance between relations and tried to learn unique discriminate parameter set for each relation. Thus, these models potentially suffer from high time complexity and large parameters, preventing them from efficiently applying on real-world KGs. In this paper, we follow the thought of parameter sharing to simultaneously learn more expressive features, reduce parameters and avoid complex feature engineering. Based on gate structure from LSTM, we propose a novel model TransGate and develop shared discriminate mechanism, resulting in almost same space complexity as indiscriminate models. Furthermore, to develop a more effective and scalable model, we reconstruct the gate with weight vectors making our method has comparative time complexity against indiscriminate model. We conduct extensive experiments on link prediction and triplets classification. Experiments show that TransGate not only outperforms state-of-art baselines, but also reduces parameters greatly. For example, TransGate outperforms ConvE and RGCN with 6x and 17x fewer parameters, respectively. These results indicate that parameter sharing is a superior way to further optimize embedding and TransGate finds a better trade-off between complexity and expressivity.

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.

More and more work has focused on incorporating different kinds of literals into Knowledge Graph to promote the performance of knowledge embedding. These literals contain numeric literals, text literals, image literals and so on. These additional descriptions are connected to the entities through certain attributes. To incorporate numeric literals, some methods combine the embeddings of literals part with the traditional part - embeddings of entities. However, in the construction of literals embeddings, these existing methods consider the differences of these attributes: one dimension represents one attribute. But they ignore semantic meanings of attributes themselves. In this paper, we propose two methods to incorporate attributes semantics into knowledge graph embeddings from two perspectives: LiteralEAN and literalE-AT. They concatenate with the embeddings of numeric literals by different ways. Furthermore, their extension model LiteralE-C is also proposed as having a more comprehensive representation of attributes semantics. In an empirical study over two standard datasets FB15k and FB15k-237, we evaluate our models for link prediction. We demonstrate that they show an effective way to improve LiteralE and achieve state-of-the-art results. In ablation experiments, we find combined models do better than their singular counterparts in most cases.

In recent years, audiences can find their interested TV play or movie videos by labels easily. However, for finding shots with certain semantic content in these videos, it is still a problem to annotate video shots by labels. Some existing approaches train models with annotated shots which cost a lot in labeling manually. Some other methods in solving this kind of task assume that the content of a video is only limited in the labels of the video. They ignore that the labels of a video are too coarse-grained to cover all content of the video. In this paper, we propose a multi-label, multi-instance topic model to annotate video shots by video labels. In a multi-label, multi-instance framework, video shots can be regarded as instances and shot labels are learned from labels in video level which makes the cost of labeling cheaper. On the other hand, our model learns label semantics by controlling the relationship between video labels and shots to solve coarse-grained problem. Furthermore, we also learn keywords for every video. The experiments on a large-scale real-world dataset show that our model outperforms other baseline models substantially.

Live video comment generating task aims to automatically generate real-time viewer comments on videos like real viewers do. Like providing search suggestions by search engines, this task can help viewers find comments they want to post by providing generated comments. Previous works ignore the interactivity and diversity of comments and can only generate general and popular comments. In this paper, we incorporate post time of the comments to deal with the real-time related comment interactions. We also take the video type labels into consideration to handle the diversity of comments and generate more related and informative comments. To this end, we propose a pre-training based encoder-decoder joint model called PLVCG model. This model is composed of a bidirectional encoder to encode context comments and visual frames jointly as well as an auto-regressive decoder to generate real-time comments and classify the type of the video. We evaluate our model in a large-scale real-world live comment dataset. The experiment results present that our model outperforms the state-of-the-art on live video comment ranking and generating task significantly.

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 multi-modal 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 multi-modal 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.

Traditional attributed network embedding methods are designed to map structural and attribute information of networks jointly into a continuous Euclidean space, while recently a novel branch of them named binarized attributed network embedding has emerged to learn binary codes in Hamming space, aiming to save time and memory costs and to naturally fit node retrieval task. However, current binarized attributed network embedding methods are scarce and mostly ignore the local attribute similarity between each pair of nodes. Besides, none of them attempt to control the independency of each dimension(bit) of the learned binary representation vectors. As existing methods still need improving, we propose an unsupervised Neural-based Binarized Attributed Network Embedding (NBANE) approach. Firstly, we inherit the Weisfeiler-Lehman proximity matrix from predecessors to aggregate high-order features for each node. Secondly, we feed the aggregated features into an autoencoder with the attribute similarity penalizing term and the orthogonality term to make further dimension reduction. To solve the problem of integer optimization we adopt the relaxation-quantization method during the process of training neural networks. Empirically, we evaluate the performance of NBANE through node classification and clustering tasks on three real-world datasets and study a case on fast retrieval in academic networks. Our method achieves better performance over state- of-the-art baselines methods of various types.

User Alignment is to find users belonging to a same real person on different social networks and has become a fundamental task for many sequent applications such as cross-network recommendation systems. When matching users in multiple social networks, existing approaches always know some correctly matched users, which can be called seeds. Then, existing methods strongly depend on the neighboring users of each user to propagate alignment information from seeds and align probable matching users implicitly. However, the completeness and validity of original alignment information among seeds cannot be fully preserved when learning and aligning multiple user spaces. In this paper, we propose a unified framework named Jumping Seed Alignment Information Propagation (JSAIP) to flexibly leverage, for each user, complete and correct alignment information from seeds. Specifically, JSAIP learns a reasonable user space for each social network by preserving enough original network and label information. Then, JSAIP ensures the correct alignment among seeds and shared labels to reduce the diversity between different user spaces. Finally, JSAIP constructs jumping links from seeds to each user in each social network and ultilizes original seed alignment information to enhance or rectify the alignment information propagated from neighbors. Experiments on real world datasets demonstrate the effectiveness of our proposed JSAIP method compared to several state-of-the-art methods.

Point-of-interest (POI) recommendation has become an important service to help users discover attractive locations. A variety of available check-in data make it possible to build a personalized POI recommender system, but the extreme sparsity of check-in data poses a severe challenge for POI recommendation. Recent studies mainly utilize social information, categorical information and/or geographical information to supplement the highly sparse check-in data. However, these studies often apply shallow methods for the extra information and provide considerably limited improvements on POI recommendation. In this paper, we propose a fine-grained POI recommendation framework, called FGRec to capture the intrinsic influences of social, categorical and geographical information on the check-in behaviors of users. First, we study the social influence in depth by exploiting the multi-hop social friends and top-n nearest neighbor friends, not only the direct friends (i.e., 1-hop friends). Second, we investigate the categorical influence by factorizing both user-POI and user-category matrices simultaneously over the same user embedding space, rather than simply using the popularity of POI categories. Third, we explore the geographical influence by integrating two types of distance (i.e., the distance between user homes and POIs and the distance among POIs) into a unified probability distribution over check-in POIs, instead of modeling them separately. Finally, experimental results on two large-scale real-world datasets demonstrate the effectiveness and superiority of the proposed method.

Knowing exact demographic attributes of users is crucial for human-computer interaction, intelligent marketing and automatic advertising. Ubiquitous sensor devices yield massive volumes of temporal data which hide a lot of valuable demographic information. In this paper, we bridge the gap between sensor data and demographic prediction to obtain real attributes of users from popular sensor devices: pedometer, which is widely used in mobile devices. We propose a novel model named Separated Embedding and Correlation Learning (SECL) for demographic prediction. Specifically, SECL first process the input data with a separated embedding layer to disentangle task-specific features for interference eliminating, and then capture the hidden correlations between different tasks via a correlation learning layer, finally the refined task-specific features are fed into a multi-task prediction layer to predict demographic attributes. Experimental results show impressive performance of our model on a real-world pedometer dataset, which is made publicly available on https://github.com/deepdeed/SECL.

Users’ real demographic attributes is impressively useful for intelligent marketing, automatic advertising and human-computer interaction. Traditional method on attribute prediction make great effort on the study of social network data, but ignore massive volumes of disparate, dynamic, and temporal data derived from ubiquitous mobile and sensor devices. For example, daily walking step counts produced by pedometer. Multiple demographic prediction on temporal data have two problems. First one is that differential effectiveness of different time period data for prediction is unclear. And another one is how to effectively learn the complementary correlations between different attributes. To address the above problem, we propose a novel model named Correlation-Aware Neural Embedding with Attention (CANEA), which first directly separates different attribute oriented feature using separated embedding layer, and use attention mechanism to assign a higher weight to dominant time point. Then it captures informative correlations using correlation learning layer. Finally we obtain the refined task-specific representations with optimal correlation information for predicting certain attributes. Experimental results show the effectiveness of our method.

Knowledge graphs (KGs) have proven to be effective to improve the performance of recommendation. However, with the tremendous increase of users and items, existing methods still face several challenging problems: (1) path-based methods rely heavily on manually designed meta-path; (2) embedding-based methods lack sufficient considerations of user personality. To overcome the shortcomings of previous works, we propose a novel model, named KCER, short for leveraging Knowledge Context to Enhance Recommendation. Firstly, KCER generates the representation of knowledge context associating with specific user-item pairs. Then to obtain enriched user representations, we leverage a gated attention network to extracted meaningful information from the associated knowledge context and user dedicated ID embedding. We conduct extensive experiments on three real-world datasets to evaluate the model. The experimental results show the superiority of KCER compared with other state-of-the-art methods.

There is a significant style overfitting problem in neural-based character recognition: insufficient generalization ability to recognize characters with unseen styles. To address this problem, we propose a novel framework named Style-Melt Nets (SMN), which disentangles the style and content factors to extract pure content feature. In this framework, a pair of parallel style net and content net is designed to respectively infer the style labels and content labels of input character images, and the style feature produced by the style net is fed to the content net for eliminating the style influence on content feature. In addition, the marginal distribution of character pixels is considered as an important structure indicator for enhancing the content representations. Furthermore, to increase the style diversity of training data, an efficient data augmentation approach for changing the thickness of the strokes and generating outline characters is presented. Extensive experimental results demonstrate the benefit of our methods, and the proposed SMN is able to achieve the state-ofthe-art performance on multiple real world character sets.

With the popularity of smart devices such as mobile phones and tablets, the security problem of the widely used ARMv8-A processor has received more and more attention. Flush+Reload and Flush+Flush cache attacks have become two of the most important security threats due to their low noise and high resolution. In order to resist Flush+Reload and Flush+Flush attacks, researchers proposed many defense methods. However, these existing methods have various shortcomings. The runtime defense methods using hardware performance counters cannot detect attacks fast enough, effectively detect Flush+Flush or avoid a high false positive rate. Static code analysis schemes are powerless for obfuscation techniques. The approaches of permanently reducing the resolution can only be utilized on browser products and cannot be applied in the system. In this paper, we design two more secure collaborative APIs—flush operation API and high resolution time API—which can resist Flush+Reload and Flush+Flush attacks. When the flush operation API is called, the high resolution time API temporarily reduces its resolution and automatically restores. Moreover, the flush operation API also has the ability to detect and handle suspected Flush+Reload and Flush+Flush attacks. The attack and performance comparison experiments prove that the two APIs we designed are safer and the performance losses are acceptable.

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.

ARM-FPGA embedded SoCs have been widely used in the fields of drones, embedded and IoT devices due to its high performance and hardware design flexibility. However, ARM-FPGA embedded SoC suffers various types of security threats, one of which is flush-based cache attack. The proposed defense schemes either lead to a high false positive rate or a large performance loss. Due to the importance of high resolution time APIs in the system, schemes that permanently reduce the resolution of time APIs can only be implemented in specific applications such as browsers. Moreover, the method of protecting high resolution timers in software cannot defend against an attacker with root privileges. In this paper, we propose a more secure timer which is a hardware/software co-design on ARM-FPGA embedded SoC. When a software process calls the flush operation, the timer adaptively reduces its resolution and recover after a short period of time. In the case that the flush operation is not called, the impact of the timer on system performance is almost negligible. This hardware/software co-design guarantees the availability of a high resolution time API while defend against attackers with root privileges. The results of the attack experiments show that the success rates of Flush+Reload and flush-based Spectre attacks can be reduced to less than 1% when using the timer. Performance test results show that the timer access latency is 9.5% slower than the fastest PMCCNTR but 5% faster than the global timer of Cortex-A9 MPCore. The modified flush operation API for the design only increases the time consumption by about 12%.

Big data brings much convenience to our daily life, while it also raises concerns to personal information leakage. It is important for data owner to protect sensitive data from being attacked. In this paper, we elaborate on the possible attacks that may be suffered at various stages of personal data life cycle and the corresponding privacy protection methods. Based on these methods, a differential attribute desensitization system (DADS) for personal information protection has been proposed. Data owner can define sensitive attribute level in the DADS. And then DADS can automatically identify sensitive data, and take differentiated data desensitization measures for structured and unstructured multi-attribute information. Experimental results show that DADS can automatically adapt to different types of sensitive information and protect personal sensitive information effectively.

Recently, network embedding has attracted a surge of attention due to its ability to automatically extract features from graph-structured data. Though network embedding method has been intensively studied, most of the existing approaches pay attention to either structures or attributes. In this paper, we propose a novel attributed network embedding method based on a Siamese neural network, named SANE, to capture both the network structure and node attribute information in a principled way. Specifically, to preserve local semantic proximity, we adopt a Siamese neural network, which can directly learn the similarity of paired nodes with their attributes as input. Then, a skip-gram module is connected with the final shared hidden layer to capture high-order proximity based on the latent representation of node attributes. Thus, we can learn the complex interrelations between nodes. Empirically, we evaluate our model on several real-world datasets and the experimental results have verified the effectiveness of our proposed approach.

This paper addresses attributed network embedding which maps the structural information and multi-modal attribute data into a latent space. Most existing network embedding algorithms concentrate on either node-oriented modeling or edge-oriented modeling, resulting in unilaterally capturing information from nodes or edges. However, there is no effective method to bilaterally extract node attributes cooperated with edge attributes, which delineates the outline and detail of social network. To this end, we propose a novel Node-Edge Bilateral Attributed Network Embedding method named NEBANE. Regarding each edge as a specific node, we construct a pioneering node-edge-node triangular structure for bilateral information modeling on both nodes and edges. Furthermore, we envisage a pairwise loss which maximizes the likelihood of connected node pairs and of connected node-edge pairs to measure the node-node and node-edge similarity. Empirically, experiments on two real-world datasets, including link prediction and node classification, are conducted in this paper. Our method achieves substantial performance gains compared with state-of-the-art baselines (e.g., 4.21%–13.65% lift by AUC scores for link prediction).

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.

Demographic attributes are crucial for characterizing different types of users in developing market strategy. However, in retail scenario, individual demographic information is not often available due to the difficult manual collection process. Several studies focus on inferring users’ demographic attribute based on their transaction histories, but there is a common problem. Hardly work has introduced knowledge for purchase data embedding. Specifically, purchase data is informative, full of related knowledge entities and common sense. However, existing methods are unaware of such external knowledge and latent knowledge-level connections among items. To address the above problem, we propose a Knowledge-Aware Embedding (KAE) method that incorporates knowledge graph representation into demographic prediction. The KAE is a multi-channel and item-entity-aligned knowledge-aware convolutional neural network that fuses frequency-level and knowledge-level representations of purchase data. Through extensive experiments on a real world dataset, we demonstrate that KAE achieves substantial gains on state-of-the-art demographic prediction models.

User Identity Linkage is to find users belonging to the same real person in different social networks. Besides, anchor users refer to matching users known in advance. However, how to match users only based on network information is still very difficult and existing embedding methods suffer from the challenge of error propagation. Error propagation means the error occurring in learning some users’ embeddings may be propagated and amplified to other users along with edges in the network. In this paper, we propose the Anchor UseR ORiented Accordant Embedding (AURORAE) method to learn the vector representation for each user in each social network by capturing useful network information and avoiding error propagation. Specifically, AURORAE learns the potential relations between anchor users and all users, which means each user is directly connected to all anchor users and the error cannot be propagated without paths. Then, AURORAE captures the useful local structure information into final embeddings under the constraint of accordant vector representations between anchor users. Experimental results on real-world datasets demonstrate that our method significantly outperforms other state-of-the-art methods.

There is a significant style overfitting problem in traditional content supervision models of character recognition: insufficient generalization ability to recognize the characters with unseen font styles. To overcome this problem, in this paper we propose a novel framework named Style and Content Supervision (SCS) network, which integrates style and content supervision to resist style overfitting. Different from traditional models only supervised by content labels, SCS simultaneously leverages the style and content supervision to separate the task-specific features of style and content, and then mixes the style-specific and content-specific features using bilinear model to capture the hidden correlation between them. Experimental results prove that the proposed model is able to achieve the state-of-the-art performance on several widely used real world character sets, and it obtains relatively strong robustness when the size of training set is shrinking.

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.

POI (point-of-interest) recommendation as an important location-based service has been widely utilized in helping people discover attractive locations. A variety of available check-in data provide a good opportunity for developing personalized POI recommender systems. However, the extreme sparsity of check-in data and inefficiency of exploiting unobserved feedback pose severe challenges for POI recommendation. To cope with these challenges, we develop a heterogeneous graph embedding-based personalized POI recommendation framework called HRec. It consists of two modules: the learning module and the ranking module. Specifically, we first propose the learning module to produce a series of intermediate feedback from unobserved feedback by learning the embeddings of users and POIs in the heterogeneous graph. Then we devise the ranking module to recommend each user the ultimate ranked list of relevant POIs by utilizing two pairwise feedback comparisons. Experimental results on two real-world datasets demonstrate the effectiveness and superiority of the proposed method.

Knowledge representation learning aims to transform entities and relationships in a knowledge base into computable forms, so that an efficient calculation can be realized. It is of great significance to the construction, reasoning and application of knowledge base. The traditional translation-based models mainly obtain the finite dimension vector representation of entities or relationships by projecting to finite dimensional Euclidean space. These simple and effective methods greatly improve the efficiency and accuracy of knowledge representation. However, they ignore a fact that the semantic space develops and grows forever with the passage of time. Finite dimensional Euclidean space is not enough in capacity for vectorizing infinitely growing semantic space. Time is moving forward forever, so knowledge base would expand infinitely with time. This determines that the vector representation of entities and relationships should support infinite capacity. We fill the gap by putting forward TransI (Translating Infinite Dimensional Embeddings) model, which extends knowledge representation learning from finite dimensions to infinite dimensions. It is trained by Trend Smooth Distance based on the idea of continuous infinite dimension vector representation. The Training Efficiency of TransI model is obviously better than TransE under the same setting, and its effect of Dimension Reduction Clustering is more obvious.

In recent years, deep neural networks have been widely applied to recommender systems. Although there are extensive explorations of deep neural networks on the collaborative filtering problem in item recommendation, most of the existing methods employ a similar loss function, i.e., the prediction loss of user-item interactions, and only change the form of the input, which may limit the model’s performance. To address this problem, we present a novel framework, named DCAR, short for Deep Collaborative Autoencoder for Recommendation. Specifically, with the implicit feedback matrix as the input, we employ the autoencoder module to obtain the latent representations of users and items respectively. Then, to predict the matching score of corresponding user-item pairs, an interaction prediction module is designed based on the neural network architecture. The two parts are coupled together and employ alternating training to learn. We conduct extensive experiments on several real-world datasets and the results empirically verify the superior performance of DCAR on item recommendation. The code related to this paper is available at: https://github.com/strange-jiong/DCAR.

Aligning users belonging to the same person in different social networks has attracted much attention. Recently, embedding methods have been proposed to represent users from different social networks into vector spaces with same dimension. To handle the challenge of vector space diversity, existing methods usually make vectors of known aligned users closer/consistent and overlap different vector spaces. However, compared to large amount of users in each social network, the consistence constraint on aligned users is not enough to reduce the diversity. Besides, missing edges/labels may provide incorrect information and affect the effect of the overlap between learned vector spaces. Therefore, we propose the OURLACER method, i.e, jOint UseR and LAbel ConsistencE Representation, to jointly represent each user and label under the consistence constraints of know aligned users and shared labels. Specifically, OURLACER utilizes collective matrix factorization to complete missing edges and labels for each user, which can provide sufficient information to distinguish each user. Moreover, OURLACER adds the consistence constraint on shared labels in different social networks. Because each user has own labels, label consistence can restrict each user and greatly reduce the diversity between learned vector spaces. Extensive experiments conducted on real-world datasets demonstrate that our method significantly outperforms other state-of-the-art methods.

Attributed network embedding, which aims to map the structural and attribute information into a latent vector space jointly, has attracted a surge of research attention in recent years. However, a vast majority of existing work explores the correlation between node structure and attribute values whereas the attribute type information which can be potentially complementary is ignored. How to effectively model the nodes, attribute types and attribute values as well as their relations in a unified framework is an open yet challenging problem. To this end, we propose a translation-based attributed network embedding method named TransANE. In our approach, the whole attributed network is considered as a coupled network which consists of two components, i.e., node relation network and attribute correlation network. We construct attribute correlation network by the co-occurrence of attribute values. Each node-attribute relation is regarded as an attributional triple, e.g., (Tom, Gender, Male). We introduce knowledge representation method to model the mapping between nodes, attribute types and attribute values. Empirically, experiments on two real-world datasets including node multi-class classification and network visualization are conducted to evaluate the effectiveness of our method TransANE in this paper. Our method achieves significant performance compared with state-of-the-art baselines.

With the boom of online video uploading, video tagging becomes an important way for video indexing. However, text-based video tagging methods ignore either genre labels or temporal differences of videos, which makes results defective. Fortunately, a new type of videos called time-sync commented videos which contains large amounts of information commented by the users helps videos tagging. In this paper, we propose a supervised dynamic Latent Dirichlet Allocation model utilizing the variational topics of time-sync comments to extract both genre labels and keywords as tags. We also implement experiments on large scale real-world datasets and the effectiveness of our model are proved both in genre label classification and keyword extraction compared with baseline models.

Network representation learning (NRL), which has become an focus of current research, learns low-dimensional vertex representations to capture network information. However, conventional NRL models either largely neglect the rich semantic information on edges and fail to extract good features of relations, or employ complex models that have rather high space and time complexities. In this work, we present an efficient NRL model, MultNet, for Social Relation Extraction (SRE) task, which evaluates the ability of NRL models on modeling the relationships between vertices. We conduct extensive experiments on several public data sets and experiments on SRE indicate that MultNet outperforms other baseline models significantly.

Societal event forecasting plays a significant role in crisis warning and emergency management. Most traditional prediction methods focus on predicting whether specific events would happen or not. However, the results of these methods are not always informative for the policy makers due to excessive frequency, lack of details and supportive evidence about the predictive events. In this paper, we focus on the problem of integrated fine-grained event forecasting which is to predict the attributes of events and find out related precursors. Given a collection of news sequences, we transform the problem into a Multi-Instance Multi-Label learning (MIML) framework. Considering the sequential influence of events and hybridity of news, we implement the MIML framework based on the Long Short-Term Memory (LSTM) neural network, and propose the model called MIML-LSTM to extract three levels of deep features which represent news article, daily status and news sequence respectively. Based on this hierarchical representation, we design a compositional objective function for joint training of each part. Taking multiple types of protest event prediction as a demonstration, we evaluate the proposed model on news streams from three countries in Latin America, and the experimental results show the effectiveness of our model on integrated fine-grained event forecasting.

User identity linkage is to identify all the users belonging to the same individual in different networks and has been widely studied along with the increasing popularity of diverse social media sites. Generally, a pair of probable corresponding users on different networks may form a true “Anchor Link”. Most existing methods identify a user based on unique features (username, interests, friends, etc.) and neglect the importance of users local network structure. Therefore, one challenging problem is how to address the user identity linkage problem if only structural information is available. In this paper, we explore techniques for dealing with the fundamental and accumulated information from neighbouring anchor links. Furthermore, we design a Trustworthy Predicting Approach (TPA) for computing the authority of an anchor link, inferring the trustworthiness of a candidate anchor link being true and predicting whether an anchor link is able to be veritably formed. Experiments illustrate the effectiveness of our proposed algorithm.

Recently, the convolutional neural networks (CNNs) show the great power in dealing with various image classification tasks. However, in the task of Chinese character recognition, there is a significant problem in CNN-based classifiers: insufficient generalization ability to recognize the Chinese characters with unfamiliar font styles. We call this problem the Style Overfitting. In the process of a human recognizing Chinese characters with various font styles, the internal skeletons of these characters are important indicators. This paper proposes a novel tool named Skeleton Kernel to capture skeleton features of Chinese characters. And we use it to assist CNN-based classifiers to prevent the Style Overfitting problem. Experimental results prove that our method firmly enhances the generalization ability of CNN-based classifiers. And compared to previous works, our method requires a small training set to achieve relatively better performance.

Attributed network embedding, which aims to map structural and attribute information into a latent vector space jointly, has attracted a surge of research attention in recent years. However, existing methods mostly concentrate on either the local proximity (i.e., the pairwise similarity of connected nodes) or the global proximity (e.g., the similarity of nodes’ correlation in a global perspective). How to learn the global and local information in structure and attribute into a same latent space simultaneously is an open yet challenging problem. To this end, we propose a Neural-based Attributed Network Embedding (NANE) approach. Firstly, an affinity matrix and an adjacency matrix are introduced to encode the attribute and structural information in terms of the overall picture separately. Then, we impose a neural-based framework with a pairwise constraint to learn the vector representation for each node. Specifically, an explicit loss function is designed to preserve the local and global similarity jointly. Empirically, we evaluate the performance of NANE through node classification and clustering tasks on three real-world datasets. Our method achieves significant performance compared with state-of-the-art baselines.

With the rapid development of IoT devices in the direction of multifunction and personalization, All Programmable SoC has been used more and more frequently because of its unrivaled levels of system performance, flexibility, and scalability. On the other hand, this type of SoC faces a growing range of security threats. Among these threats, cache timing attacks and power/elctromagnetic analysis attacks are two considerable ones which have been widely studied. Although many countermeasures have been proposed to resist these two types of attacks, most of them can only withstand a single type but are often incapable when facing multi-type attacks. In this paper, we utilize the special architecture of All Programmable SoC to implement a secure AES encryption scheme which can efficiently resist both cache timing and power/electromagnetic analysis attacks. The AES implementation has a beginning software stage, a middle hardware stage and a final software stage. Operations in software and start/end round of hardware are all randomized, which allow our implementation to withstand two types of attacks. To illustrate the security of the implementation, we conduct the three types of attacks on unprotected software/hardware AES, shuffled software AES and our scheme. Furthermore, we use Test Vector Leakage Assessment (TVLA) to test their security on encryption times and power/electromagnetic traces. The final result indicates that our encryption implementation achieves a high secure level with almost 0.86 times data throughput of the shuffled software AES implementation.

The key whitening is a technique intended to enhance the strength of a block cipher. Although some research work involves DPA attacks against the key whitening layer in the compact architecture, there are no literatures dedicated in the influence of the key whitening layers in the loop architecture from the standpoint of DPA. In this paper, we propose a practical chosen message power analysis approach against the loop architecture of ciphers with the key whitening layers, thus proving that the key whitening technique does not enhance the security of ciphers regard to DPA. Our approach follows a reduction strategy: we recover the whitening key in the general cipher with the key whitening layer and reduce other complicated key whitening layers to the general case. In order to further manifest the validity of the new approach, we carry extensive experiments on two ISO standardized ciphers CLEFIA and Camellia implemented in loop architecture on FPGA, and the keys are recovered as expected.

Improved Rotating S-box Masking (RSM2.0 for short) is a well-known countermeasure designed and implemented by DPA Contest V4.2 committee to provide security protection for AES-128. By combining both 1st-order masking and shuffling techniques, improved RSM claims to offer at least non-profiled resistance for its software implementation and up to now no systematic research has been published to challenge such security claim yet. To study the practical security of RSM2.0 against non-profiled attacks, we first propose an analytical methodology to guide the detection of the exploitable vulnerabilities in RSM2.0. On the basis of the methodology, several potential flaws hidden in both the algorithm design and detailed implementation of RSM2.0 are discovered and we make use of them to design six attacking schemes in total, all of which belong to non-profiled attacks. Four representative attacks are eventually implemented and submitted to DPA Contest V4.2 for official evaluation and the results show that all the submitted attacks are both practical and feasible. Among them, the best attack scheme requires only 257 power traces to crack the complete 128-bit master key with 80% success rate. To further improve the security level of RSM2.0, we also discuss some possible strategies to eliminate or mitigate the threats proposed by us.

Low-entropy masking schemes and shuffling technique are two common countermeasures against traditional side-channel analysis. Improved Rotating S-box Masking (RSM) is a combination of both countermeasures and is implemented by DPA contest committee to improve the software security level of AES-128. Compared with the original version, improved RSM mainly introduces both the offset and shuffle array as security foundations to counteract the existing attacks. In this paper, we first point out a general vulnerability referred to as "leakage fingerprints" and make use of it to successfully crack the offset array with 100% accuracy, which breaks down the masking countermeasure in the first step. Then, we show that cracking the shuffle array is still feasible but not necessary since several other vulnerabilities in the implementation level can be exploited to bypass the shuffle countermeasure directly. By selectively combining all these vulnerabilities, a dozen of attacks can be put forward, and we perform two of them as examples to verify their effectiveness. Official evaluation results show that, both attacks submitted by us are practical and feasible, and also operate with high efficiency. In terms of two major performance metrics, our best scheme requires 4 traces to reveal the AES master key with 80% Global Success Rate (GSR) and only 2 traces are enough to reduce the Maximum Partial Guessing Entropy (PGE) under 10.