Recommendation engines based on location help users find visually appealing Points-of-Interest (POI). Next POIs recommendation is very significant and advantageous. In this paper, we propose a network for next POIs prediction, called “Where to Go Next (WTG)”, employing a hybrid deep learning model based on residual neural network and vision trans-former. WTG method models multiple features, including spatio-temporal, sequential, social, and context features. Users' mobility behavior is first modeled as images and a new architecture is then used to capture the spatio-temporal patterns while taking into account the social relations connecting users. The sequential features that are necessary to characterize the user behaviors are then extracted using an the vision transformer. Moreover, context feature extraction is integrated into the model. We demonstrate our model's significant superiority over eight existing methods by evaluating it on two real location-based networks: Foursquare and Gowalla. The source code supporting the findings of this study can be accessed and reproduced using the materials provided at the following repository: https://anonymous.4open.science/r/WTG-Next-ASCC.

This study introduces DeepTran, a novel deep-learning framework designed to predict vehicle energy consumption across urban road networks. DeepTran's hybrid architecture, which integrates a Residual Neural Network (ResNet) into Vision Transformer, captures both spatial patterns and temporal dependencies while incorporating real-time traffic conditions and static vehicle characteristics. Extensive experiments conducted on Michigan's road network using a diverse fleet of vehicles demonstrate DeepTran's superior performance over existing algorithms. This research contributes to the understanding of vehicle energy consumption patterns and has potential applications in urban planning, traffic management, and environmental policy formulation. To ensure reproducibility, all code, datasets, and experimental details are publicly available on GitHub at https://github.com/VanishingLight/vehicle_energy_consumption.

Stronger defenses against adversarial attacks are subsequently broken by a more advanced defense aware attack. We propose a stronger defense to achieve state-of the-art detection performance on both standard and defense-aware attacks. To this end, we propose a new architecture, called the Residual Vision Transformer (RVT) for attack based-image classification. RVT introduces the Residual Networks to Vision Transformers (ViT) to improve ViT in performance and efficiency and yield the best of both designs. The classic architecture of ViT is mainly modified by: (i) a hierarchy of Transformers containing a new residual token embedding, and (ii) a residual Transformer block leveraging a residual projection. Moreover, the positional encoding, a crucial component in existing vision transformers, can be safely removed in the RVT model, simplifying the design for higher resolution vision tasks. In this paper, the RVT model is evaluated for its robustness against adversarial attacks and is found to perform great in detecting adversarial images. Extensive experiments conducted on CIFAR-10 and SVHN datasets, show that our proposed deep learning algorithm significantly outperforms the state-of-the-art performance over other Vision Transformers and ResNets. RVT is first pre-trained on larger datasets (e.g. ImageNet-22k) and then fine-tuned to downstream tasks achieving a top-1 accuracy of 89.3% proving that the RVT model is a promising approach for image classification offering improved performance, efficiency, and robustness against adversarial attacks. To make the results reproducible, the code, the used data and details of the experimental setup are made available online at ">https://github.com/Robotoks/RVT .

Transformers have recently dominated a wide range of tasks in natural language processing. To obtain competitive image classification performance, the Vision Transformer (ViT) is the first computer vision model to rely exclusively on the Transformer architecture. Despite the success of vision Transformers at large scale, the performance is still below similarly sized convolutional neural network (CNN) counterparts (e.g., ResNets). We present in this paper a new architecture, named Residual Vision Transformer (Res-Vit), that improves Vision Transformer (ViT) in performance and efficiency by introducing Residual networks into ViT to yield the best of both designs. The classic architecture of ViT is mainly modified by: (i) a hierarchy of Transformers containing a new residual token embedding, and (ii) a residual Transformer block leveraging a residual projection. Moreover, the positional encoding, a crucial component in existing vision transformers, can be safely removed in the ResVit model, simplifying the design for higher resolution vision tasks. We validate Res-Vit by conducting extensive experiments, showing that this approach achieves state-of-the-art performance over other Vision Transformers and ResNets on ImageNet-1k. In addition, performance gains are maintained when pretrained on larger datasets (e.g. ImageNet-22k) and fine-tuned to downstream tasks. Pretrained on ImageNet-22k, our Res-Vit obtains a top-1 accuracy on the ImageNet-1k proving that the Res-ViT model is a promising approach for image classification offering improved performance, efficiency, and robustness.

Skeleton-based human action recognition has received increasing attention in recent years. It aims at extracting features on top of human skeletons and estimating human pose. However, existing methods capture only the action information while in a real world application such as cognitive assessment, we need to measure the executive functioning that helps psychiatrists to identify some mental disease such as Alzheimer, Schizophrenia and ADHD. In this paper, we propose a skeleton-based action recognition named Deep-Cogn for cognitive assessment. Deep-Cogn integrates a pose estimator to extract the human body joints and then automatically measures the executive functioning employing the distance and elbow angle calculation. Three score functions were designed to measure the executive functioning: the accuracy score, the rhythm score and the functioning score. We evaluate our model on two different datasets and show that our approach significantly outperforms the existing methods.

Location-based recommendation tools assist users in discovering attractive Points-of-Interest (POIs). Next POIs recommendation is of great importance and benefit. In this paper, we propose an attention-CNN based network named Deep-POIs for next POIs prediction by modeling several features such as spatio-temporal, sequential, social and context features. In Deep-POIs model, users’ mobility behavior is modeled as images and then a CNN architecture is employed to capture the spatio-temporal patterns considering the social relations connecting users. An attention mechanism is then employed to extract the sequential features which are essential to describe the user behaviors. The model also integrates context feature extraction. We evaluate our model on two real location-based networks, Foursquare and Gowalla, and show that it significantly outperforms eight existing methods.

The status of air pollution is serious all over the world. Analysing and predicting vehicle energy consumption becomes a major concern. Vehicle energy consumption depends not only on speed but also on a number of external factors such as road topology, traffic, driving style, etc. Obtaining the cost for each link (i.e., link energy consumption) in road networks plays a key role in energy-optimal route planning process. This paper presents a novel framework that identifies vehicle/driving environment-dependent factors to predict energy consumption over a road network based on historical consumption data for different vehicle types. We design a deep-learning-based structure, called DeepFEC, to forecast accurate energy consumption in each and every road in a city based on real traffic conditions. A residual neural network and recurrent neural network are employed to model the spatial and temporal closeness, respectively. Static vehicle data reflecting vehicle type, vehicle weight, engine configuration and displacement are also learned. The outputs of these neural networks are dynamically aggregated to improve the spatially correlated time series data forecasting. Extensive experiments conducted on a diverse fleet consisting of 264 gasoline vehicles, 92 Hybrid Electric Vehicles, and 27 Plug-in Hybrid Electric Vehicles/Electric Vehicles drove in Michigan road network, show that our proposed deep learning algorithm significantly outperforms the state-of-the-art prediction algorithms. To make the results reproductible, the code, the used data and details of the experimental setup are made available online at https://github.com/ElmiSay/DeepFEC.

Quality of service (QoS) has been considered as a significant criterion for selecting among functionally similar application software (AS). Choosing an AS hinges not only on price and functionality, but also on user preferences as well. The skyline queries have attracted tremendous amount of attention as they are a popular example of preference queries and they can retrieve the most interesting objects from a dataset. However, existent approaches are not sufficient where the delivered QoS attributes are inherently uncertain and incomplete. In this paper, we tackle the problem of the efficient skyline computing on uncertain and incomplete QoS. We represent each QoS attribute of an AS using an evidence distribution. We then develop appropriate algorithms to efficiently compute the skyline of an AS set. Finally, we present our experimental results that show the efficiency of the proposed algorithms.

Accurate taxi fare forecasting in complex and crowded scenarios is an important building block to enabling intelligent transportation systems in a smart city. Given the observation, increasing popularity of taxi services such as Uber and Didi Chuxing in China, unable to collect large-scale taxi fare data continuously. Traditional taxi fare prediction methods mostly rely on time series forecasting techniques, which fail to model the complex non-linear spatial and temporal relations. To address those issues, we propose a Deep Multi-View Network called Temporal ResNet (TRES-Net) framework. Specifically, our proposed model consists of three views: (i) temporal view: modeling correlations between future taxi fare values with near time points, (ii) spatial view: to model deep spatial correlations, we further introduce a spatial similarity matrix that can learn from spatially similar taxi trips and capture the multi-modality of the motion patterns, and (iii) semantic view: to extract more taxi fare patterns, we integrate more factors such as trip distance, travel time, passenger count, tolls amount, tip amount, etc.. Extensive experiments on more than 700 millions NYC trips over several fare prediction benchmarks demonstrate that our method is able to predict taxi fare in complex scenarios and achieves state-of-the-art performance. Our large scale evaluation demonstrates that our system is (a) accurate—with the mean fare error under 1 US dollar and (b) capable of real-time performance.

Forecasting accurate traffic speed is of great importance to advanced traffic management systems but challenging problem as it is affected by many complex factors, such as events, inter-road traffic, traffic lights, period and weather conditions. Traffic speed prediction based on deep learning techniques has received much attention in recent years. However, the power of deep learning methods has not yet fully been exploited in traffic prediction in terms of the depth of the model architecture. This paper designs a deep-learning-based structure, called BiRNet, to forecast accurate traffic speed in each and every region in a city. More specifically, we employ the residual neural network and the bi-directional recurrent neural network to model the spatial and temporal closeness, respectively. A look-up layer is introduced to model the spatial scale of the prediction area. BiRNet learns to dynamically aggregate the output of these neural networks which is further combined with external factor learning to improve the spatially correlated time series data forecasting. Our extensive experiments on real-world trip data-sets generated in Singapore and NYC’s road network, show that our proposed deep learning algorithm significantly outperforms the state-of-the-art learning algorithms.

Cities are the main poles of human and economic activity. Analyzing cities data is very important to improve the city economy as well as the life quality of the citizens. Since location based services and GPS devices can easily connect users located in different positions, it is worthwhile to optimize the efficiency of their shifting to a common location according to their preferences. For this reason, the support of advanced analysis queries such as the skyline operator has become important. This later finds the interesting objects according to a user preferences. However, data in such application can be uncertain, imprecise and incomplete. In this paper, we propose an imperfect spatial skyline query for users located in different positions. Detailed experimental analysis are reported. In addition, the theoretical properties developed in this paper help to devise efficient techniques to compute the spatial skyline over uncertain data fora set of users. Our extensive experiments show that the proposed algorithms provide quick initial response time.

In the last decade, skyline queries have attracted the interest of several researchers in the database field due to their ability to retrieve interesting objects among a large set of objects. Skyline analysis is a powerful tool in a wide spectrum of real applications including multi-criteria optimal decision making, preference answering and many applications where uncertain, imprecise and noisy data inherently exist. As large amounts of distributed data over Internet are communicated and shared, an important problem is to retrieve the global skyline from all the distributed local sites. In addition, though the skyline queries can control selection, there exist not much works that can handle skyline queries under database updates. In this paper, based on the marginal points notions, we introduce new methods to efficiently compute the global skyline from distributed local sites and over frequently updated databases. The efficiency and effectiveness of our proposal are verified by extensive experimental results.

Uncertain, imprecise and noisy data arise in a number of domains including sensor networks and data integration. Skyline analysis is a powerful tool in a wide spectrum of real applications involving multi-criteria optimal decision making. The skyline operator aims at returning the most interesting objects in a database. Previous researches showed that the skyline size over uncertain data is too large to be exploited. In this paper, we propose an advanced skyline analysis over uncertain databases where uncertainty is modelled by the evidence theory. We particularly tackle the following two important issues: (1) model the skyline query over an evidential database (2) rank the evidential skyline result and retrieve k skyline objects that are expected to have the highest score with considering the confidence level of the objects. We also study its impact on the top-k result. The efficiency and effectiveness of our proposal are verified by extensive experimental results.

Thanks to their ability to return interesting objects in a database, the skyline queries have received considerable attention from the database community over the last few years. Skyline analysis is a powerful tool in a wide spectrum of real applications including multi-criteria optimal decision making, preference answering and many applications where uncertain, imprecise and noisy data inherently exist. As large amounts of distributed data over an Internet are communicated and shared, an important problem is to retrieve the global skyline from all the distributed local sites. In this paper, based on the skyline query over centralised imperfect data where imperfection is modeled by the evidence theory, we propose to efficiently compute the global skyline from distributed local sites. The efficiency and effectiveness of our proposal are verified by extensive experimental results.