Search results

Money laundering is the process of concealing unlawfully obtained funds by presenting them as coming from a legitimate source. Criminals use crypto money laundering to hide the illicit origin of funds using a variety of methods. The most simplified form of bitcoin money laundering leans hard on the fact that transactions made in cryptocurrencies are pseudonymous, but open data gives more power to investigators and enables the crowdsourcing of forensic analysis. With the motive to curb these illegal activities, there exist various rules, policies and technologies collectively known as anti-money laundering (AML) tools. When properly implemented, AML restrictions reduce the negative effects of illegal economic activity while also promoting financial market integrity and stability, but these bear high costs for institutions. The purpose of this work is to motivate the opportunity to reconcile the cause of safety with that of financial inclusion, bearing in mind the limitations of the available data. The authors use the Elliptic dataset; to the best of the authors' knowledge, this is the largest labelled transaction dataset publicly available in any cryptocurrency.

AML in bitcoin can be modelled as a node classification task in dynamic networks. In this work, graph convolutional decision forest will be introduced, which combines the potentialities of evolving graph convolutional network and deep neural decision forest (DNDF). This model will be used to classify the unknown transactions in the Elliptic dataset. Additionally, the application of knowledge distillation (KD) over the proposed approach gives finest results compared to all the other experimented techniques.

The importance of utilising a concatenation between dynamic graph learning and ensemble feature learning is demonstrated in this work. The results show the superiority of the proposed model to classify the illicit transactions in the Elliptic dataset. Experiments also show that the results can be further improved when the system is fine-tuned using a KD framework.

Existing works used either ensemble learning or dynamic graph learning to tackle the problem of AML in bitcoin. The proposed model provides a novel view to combine the power of random forest with dynamic graph learning methods. Furthermore, the work also demonstrates the advantage of KD in improving the performance of the whole system.

A smart city is a potential solution to the problems caused by the unprecedented speed of urbanization. However, the increasing availability of big data is a challenge for transforming a city into a smart one. Conventional statistics and econometric methods may not work well with big data. One promising direction is to leverage advanced machine learning tools in analyzing big data about cities. In this paper, the authors propose a model to learn region embedding. The learned embedding can be used for more accurate prediction by representing discrete variables as continuous vectors that encode the meaning of a region.

The authors use the random walk and skip-gram methods to learn embedding and update the preliminary embedding generated by graph convolutional network (GCN). The authors apply this model to a real-world dataset from Manhattan, New York, and use the learned embedding for crime event prediction.

This study’s results show that the proposed model can learn multi-dimensional city data more accurately. Thus, it facilitates cities to transform themselves into smarter ones that are more sustainable and efficient.

The authors propose an embedding model that can learn multi-dimensional city data for improving predictive analytics and urban operations. This model can learn more dimensions of city data, reduce the amount of computation and leverage distributed computing for smart city development and transformation.

Remaining useful life (RUL) prediction for power transformer maintenance is a challenging task on heterogeneous data. Monitoring data of power transformers are not always compatible or in an identical format; therefore, RUL predictions traditionally work separately on different data. Moreover, chemical molecules used in RUL prediction can be transformed into each other under different conditions, thus forming a complete graph with uncertain adjacency matrix (UAM). This study aims to find and evaluate a new model to achieve better results of RUL prediction than the other baselines.

In this work, the authors propose a spatiotemporal complete graph convolutional network (STCGCN) for RUL prediction in two branches, in which daily and hourly features are extracted from correlated heterogeneous data separately. This study provides a thorough evaluation of the proposed model on real-world data and compare the proposed model with state-of-the-art RUL prediction models.

By using the multibranch structure and EucCos similarity aggregation, STCGCN was able to capture the dynamic spatiotemporal patterns on a variety of heterogeneous data and obtain more accurate prediction results, compared to other time series prediction methods.

In this work, the authors propose a novel multibranch structure to compute feature maps from two heterogeneous data sources efficiently and a novel similarity aggregation method to compute the spatial UAM within the complete graph. Compared with traditional time series prediction models, the model pays attention to the spatial relationships in time series data.

Given the importance of spatial effects in improving the accuracy of hotel demand forecasting, this study aims to introduce price and online rating, two critical factors influencing hotel demand, as external variables into the model, and capture the spatial and temporal correlation of hotel demand within the region.

For high practical implications, the authors conduct the case study in Xiamen, China, where the hotel industry is prosperous. Based on the daily demand data of 118 hotels before and during the COVID-19 period (from January to June 2019 and from January to June 2021), the authors evaluate the prediction performance of the proposed innovative model, that is, a deep learning-based model, incorporating graph convolutional networks (GCN) and gated recurrent units.

The proposed model simultaneously predicts the daily demand of multiple hotels. It effectively captures the spatial-temporal characteristics of hotel demand. In addition, the features, price and online rating of competing hotels can further improve predictive performance. Meanwhile, the robustness of the model is verified by comparing the forecasting results for different periods (during and before the COVID-19 period).

From a long-term management perspective, long-term observation of market competitors’ rankings and price changes can facilitate timely adjustment of corresponding management measures, especially attention to extremely critical factors affecting forecast demand, such as price. While from a short-term operational perspective, short-term demand forecasting can greatly improve hotel operational efficiency, such as optimizing resource allocation and dynamically adjusting prices. The proposed model not only achieves short-term demand forecasting, but also greatly improves the forecasting accuracy by considering factors related to competitors in the same region.

The originalities of the study are as follows. First, this study represents a pioneering attempt to incorporate demand, price and online rating of other hotels into the forecasting model. Second, integrated deep learning models based on GCN and gated recurrent unit complement existing predictive models using historical data in a methodological sense.

Traffic flow prediction has always been a top priority of intelligent transportation systems. There are many mature methods for short-term traffic flow prediction. However, the existing methods are often insufficient in capturing long-term spatial-temporal dependencies. To predict long-term dependencies more accurately, in this paper, a new and more effective traffic flow prediction model is proposed.

This paper proposes a new and more effective traffic flow prediction model, named channel attention-based spatial-temporal graph neural networks. A graph convolutional network is used to extract local spatial-temporal correlations, a channel attention mechanism is used to enhance the influence of nearby spatial-temporal dependencies on decision-making and a transformer mechanism is used to capture long-term dependencies.

The proposed model is applied to two common highway datasets: METR-LA collected in Los Angeles and PEMS-BAY collected in the California Bay Area. This model outperforms the other five in terms of performance on three performance metrics a popular model.

(1) Based on the spatial-temporal synchronization graph convolution module, a spatial-temporal channel attention module is designed to increase the influence of proximity dependence on decision-making by enhancing or suppressing different channels. (2) To better capture long-term dependencies, the transformer module is introduced.

The purpose of this paper aims to model interaction relationship of traffic agents for motion prediction, which is critical for autonomous driving. It is obvious that traffic agents’ trajectories are influenced by physical lane rules and agents’ social interactions.

In this paper, the authors propose the social relation and physical lane aggregator for multimodal motion prediction, where the social relations of agents are mainly captured with graph convolutional networks and self-attention mechanism and then fused with the physical lane via the self-attention mechanism.

The proposed methods are evaluated on the Waymo Open Motion Dataset, and the results show the effectiveness of the proposed two feature aggregation modules for trajectory prediction.

This paper proposes a new design method to extract traffic interactions, and the attention mechanism is used in each part of the model to extract and fuse different relational features, which is different from other methods and improves the accuracy of the LSTM-based trajectory prediction method.

With the continuous growth of online education, the quality issue of online educational videos has become increasingly prominent, causing students in online learning to face the dilemma of knowledge confusion. The existing mechanisms for controlling the quality of online educational videos suffer from subjectivity and low timeliness. Monitoring the quality of online educational videos involves analyzing metadata features and log data, which is an important aspect. With the development of artificial intelligence technology, deep learning techniques with strong predictive capabilities can provide new methods for predicting the quality of online educational videos, effectively overcoming the shortcomings of existing methods. The purpose of this study is to find a deep neural network that can model the dynamic and static features of the video itself, as well as the relationships between videos, to achieve dynamic monitoring of the quality of online educational videos.

The quality of a video cannot be directly measured. According to previous research, the authors use engagement to represent the level of video quality. Engagement is the normalized participation time, which represents the degree to which learners tend to participate in the video. Based on existing public data sets, this study designs an online educational video engagement prediction model based on dynamic graph neural networks (DGNNs). The model is trained based on the video’s static features and dynamic features generated after its release by constructing dynamic graph data. The model includes a spatiotemporal feature extraction layer composed of DGNNs, which can effectively extract the time and space features contained in the video's dynamic graph data. The trained model is used to predict the engagement level of learners with the video on day T after its release, thereby achieving dynamic monitoring of video quality.

Models with spatiotemporal feature extraction layers consisting of four types of DGNNs can accurately predict the engagement level of online educational videos. Of these, the model using the temporal graph convolutional neural network has the smallest prediction error. In dynamic graph construction, using cosine similarity and Euclidean distance functions with reasonable threshold settings can construct a structurally appropriate dynamic graph. In the training of this model, the amount of historical time series data used will affect the model’s predictive performance. The more historical time series data used, the smaller the prediction error of the trained model.

A limitation of this study is that not all video data in the data set was used to construct the dynamic graph due to memory constraints. In addition, the DGNNs used in the spatiotemporal feature extraction layer are relatively conventional.

In this study, the authors propose an online educational video engagement prediction model based on DGNNs, which can achieve the dynamic monitoring of video quality. The model can be applied as part of a video quality monitoring mechanism for various online educational resource platforms.

The purpose of this paper is to summarize the existing point cloud target detection algorithms based on deep learning, and provide reference for researchers in related fields. In recent years, with its outstanding performance in target detection of 2D images, deep learning technology has been applied in light detection and ranging (LiDAR) point cloud data to improve the automation and intelligence level of target detection. However, there are still some difficulties and room for improvement in target detection from the 3D point cloud. In this paper, the vehicle LiDAR target detection method is chosen as the research subject.

Firstly, the challenges of applying deep learning to point cloud target detection are described; secondly, solutions in relevant research are combed in response to the above challenges. The currently popular target detection methods are classified, among which some are compared with illustrate advantages and disadvantages. Moreover, approaches to improve the accuracy of network target detection are introduced.

Finally, this paper also summarizes the shortcomings of existing methods and signals the prospective development trend.

This paper introduces some existing point cloud target detection methods based on deep learning, which can be applied to a driverless, digital map, traffic monitoring and other fields, and provides a reference for researchers in related fields.

The paper aims to transfer the item image of a given clothing product to a corresponding area of the user image. Existing classical methods suffer from unconstrained deformation of clothing and occlusion caused by hair or poses, which leads to loss of details in the try-on results. In this paper, the authors present a details-oriented virtual try-on network (DO-VTON), which allows synthesizing high-fidelity try-on images with preserved characteristics of target clothing.

The proposed try-on network consists of three modules. The fashion parsing module (FPM) is designed to generate the parsing map of a reference person image. The geometric matching module (GMM) warps the input clothing and matches it with the torso area of the reference person guided by the parsing map. The try-on module (TOM) generates the final try-on image. In both FPM and TOM, attention mechanism is introduced to obtain sufficient features, which enhances the performance of characteristics preservation. In GMM, a two-stage coarse-to-fine training strategy with a grid regularization loss (GR loss) is employed to optimize the clothing warping.

In this paper, the authors propose a three-stage image-based virtual try-on network, DO-VTON, that aims to generate realistic try-on images with extensive characteristics preserved.

The authors’ proposed algorithm can provide a promising tool for image based virtual try-on.

The authors’ proposed method is a technology for consumers to purchase favored clothes online and to reduce the return rate in e-commerce.

Therefore, the authors’ proposed algorithm can provide a promising tool for image based virtual try-on.

Automatic anatomical therapeutic chemical (ATC) classification is progressing at a rapid pace because of its potential in drug development. Predicting an unknown compound's therapeutic and chemical characteristics in terms of how it affects multiple organs and physiological systems makes automatic ATC classification a vital yet challenging multilabel problem. The aim of this paper is to experimentally derive an ensemble of different feature descriptors and classifiers for ATC classification that outperforms the state-of-the-art.

The proposed method is an ensemble generated by the fusion of neural networks (i.e. a tabular model and long short-term memory networks (LSTM)) and multilabel classifiers based on multiple linear regression (hMuLab). All classifiers are trained on three sets of descriptors. Features extracted from the trained LSTMs are also fed into hMuLab. Evaluations of ensembles are compared on a benchmark data set of 3883 ATC-coded pharmaceuticals taken from KEGG, a publicly available drug databank.

Experiments demonstrate the power of the authors’ best ensemble, EnsATC, which is shown to outperform the best methods reported in the literature, including the state-of-the-art developed by the fast.ai research group. The MATLAB source code of the authors’ system is freely available to the public at https://github.com/LorisNanni/Neural-networks-for-anatomical-therapeutic-chemical-ATC-classification.

This study demonstrates the power of extracting LSTM features and combining them with ATC descriptors in ensembles for ATC classification.