Ethereum, the first blockchain platform to support smart contracts, has become a target for various cybercrimes, particularly financial frauds like Ponzi schemes. Ponzi schemes on Ethereum are known as Smart Ponzi Schemes (or Ponzi Contracts) and have caused huge financial losses. Current Ponzi contract detection models face three main challenges: simple opcode sequence processing does not effectively distinguish Ponzi from non-Ponzi contracts, single-feature-based models lack accuracy, and reliance on transaction records hinders early detection. To address these issues, this paper proposes a Multi-Feature Ponzi Scheme Detection Model (MFDPonzi). MFDPonzi tracks the changes in stack, memory, and storage parameters during the execution of smart contracts, reconstructing opcode sequences and extracting diverse features, including semantic and developer features. Finally, a multi-feature fusion algorithm is used to enhance model stability. Additionally, MFDPonzi can identify Ponzi contracts at the early stage of smart contract creation without relying on transaction data. Experimental results show that MFDPonzi achieves an 85.9% recall and an 88.7% F-score on Ethereum smart contracts, outperforming baselines in both performance and robustness.

Bayesian Personalized Ranking (BPR) is a widely used optimization function in GNN-based recommender systems, and negative samples are usually obtained through the Random Negative Sampling (RNS) method during BPR training. However, from the gradient perspective, RNS tends to select low-quality samples with minimal information, resulting in small gradients. These small gradients contribute little to BPR optimization, limiting the model’s ability to effectively distinguish between positive and negative samples. To alleviate this issue, we propose a general negative sample information enhancement method: Enhancing Random Negative Sampling (E-RNS), which constructs hard negative samples by enhancing the information in randomly selected negative samples. Specifically, in the Noise Injection step, it generates initial noise and injects a certain amount of noise in the same direction into the vector dimensions of positive samples to create enriched information. Then, in the Information Fusion step, this enriched information is mixed with the negative samples to synthesize new hard negative samples. Extensive experiments demonstrate that applying E-RNS to GNN-based recommender models significantly improves performance.

Deep neural networks have recently been applied to long-term time series forecasting (LTSF) tasks, aiming to capture dynamic temporal features from historical data for more accurate predictions. However, time series data inevitably contain noise, influencing prediction accuracy. Recent denoising studies suggest that high-frequency components in the frequency domain are the source of noise, and these studies use filters to remove them. While these filters effectively suppress noise, they also unintentionally discard valuable high-frequency information. To tackle this problem, we propose a Wavelet-Enhanced Long Short-Term Memory (WE-LSTM). Specifically, we introduce an Enhanced Spectral Denoising module that combines Fourier Transform and Discrete Wavelet Transform (DWT) for joint denoising. This module achieves effective denoising through adaptive wavelet decomposition with multiple wavelet bases. Additionally, this work presents an Enhanced LSTM module, which integrates exponential gating units into LSTM to capture complex patterns and enhance LTSF capabilities. Extensive experiments on seven real-world datasets with varying noise levels demonstrate that the WE-LSTM architecture achieves performance comparable to current state-of-the-art models in LTSF tasks, validating the superiority of our model.

In sequencial recommendation, the expansion of Transformer layers often results in over-smoothing phenomena, where the hidden representations of users become similar. The problem occurs because self-attention algorithms essentially act as a low-pass filter, insufficient for capturing high-frequency information. As a result, self-attention algorithms struggle to detect the rapidly changing interests of users in the short term. Residual connection promotes information flow by passing the rich low-frequency information within the model, compelling the model to focus on high-frequency details. Therefore, we propose a sequencial recommendation model named Hybrid Residual Connect attention network for sequential Recommendation(HRCRec). Hybrid Residual Connection utilizes a frequency separation module to separate low-frequency and high-frequency components, then employs feature rescaling to adaptively amplify the high-frequency components. Finially, adaptive aggregation is applied to balance the weights of high and low frequency information during training. Our experimental evaluation on four benchmark datasets indicates that our model surpasses other baseline methods in recommendation accuracy.

Social recommendation typically enhances user preference representation by integrating social connections among users. However, users’ intricate social behaviors may introduce noisy social connections for user preference representation modeling. Due to the absence of ground truth labels, existing social denoising models typically rely simply on user similarity as the denoising criterion. These models fail to accurately identify noisy social connections with low preference similarity and neglect the crucial role of opinion leaders in the propagation of social influence. In this paper, we propose a novel Social Influence and Preference-guided denoising enhancement framework(SIP) for social recommendation. The framework includes the Preference-guided Social Relationship Augmentation (PSRA) module and the Social Influence Augmentation (SIA) module. In the PSRA module, we reduce the impact of social connections with low preference similarity by strengthening the associations between users and their close friends. At the same time, considering the importance of opinion leaders in user preference modeling, we use the SIA module to narrow the gap in preference representation between users and opinion leaders. Extensive experiments on three real-world datasets demonstrate that our proposed framework effectively removes noise from social data and significantly enhances the performance of existing state-of-the-art social recommendation models.

Transformers have emerged as a crucial technology in long-term time series forecasting (LTSF). Recent studies have demonstrated that Transformers segment time series into patches, effectively capturing temporal patterns. However, single-scale patches still fail to adequately capture the complex short-term fluctuations and long-term trends in time series. To tackle this issue, we propose a novel framework, the Multi-scale Patch Parallel Prediction Network (MP3Net). Firstly, we propose a multi-scale patch module to extract local features and long-term correlations from the time series to effectively capture and identify data characteristics across different time scales. Secondly, we propose an adaptive fusion module to dynamically balance the weight of local detail features and long- term correlations from each branch. Thirdly, we introduce a hybrid loss function for model training aimed at enhancing the sensitivity and adaptability of the network towards different types of errors, consequently improving overall performance. Experiments conducted on five benchmark datasets demonstrate that MP3Net surpasses existing methods in LTSF and MP3Net presents an effective approach for addressing the challenge of modeling long-term dependencies in time series analysis.

Transformer has achieved excellent results in the field of time series forecasting. However, some recent studies have pointed out that the existing Transformer models are ineffective in preserving the temporal order information. To solve this problem, we propose a Linear Weighted Attention mechanism (LinWA), a new, simple, and effective attention weights calculation method. Linwa calculates the linear weights and the attention weights, respectively, and combines them to reduce the proportion of the original attention weights to use the order information effectively and improve the forecasting performance. Additionally, we introduce an adaptive parameter adjustment mechanism to adjust the proportion between linear weights and attention weights dynamically. Experimental results on six publicly available datasets show that LinWA preserves temporal order well and improves performance on many Transformer models.

In order to cope with the increasingly severe global energy conservation and emission reduction problems, research on urban carbon emission prediction is of great significance. The existing methods mainly use time series analysis to predict urban carbon emission, but there is a strong spatial correlation between the carbon emission data of several cities. Therefore, this paper designs a multi-scale spatial-temporal feature screening attention network to predict target cities' future carbon emission data. Firstly, this paper combines the daily carbon emission data of the near-neighbouring cities and the daily homologous emission data of the target city to analyze the urban carbon emission data from a spatio-temporal perspective. Then, this paper designs a multi-scale spatial interactive convolution module and a multi-scale temporal convolution module to extract multi-scale spatio-temporal features effectively. In addition, the feature screening module is designed to reduce the adverse effects of redundant features. Finally, the multi-scale features are used to predict the future carbon emissions of the target city through a predictor. The experimental results show that our prediction model is superior to the existing methods in six datasets.

Abstract As domain knowledge evolves, new concepts (entities) continuously emerge, leading to a decrease in the coverage of existing taxonomies with hierarchical structures, thus necessitating the continual expansion of these taxonomies to include new concepts. Due to the relationships (“contain”, “disjoint”, and “intersect”) between the boxes, which can effectively represent asymmetric hierarchies, box embeddings have been successfully applied in taxonomy expansion. However, existing models that use box embeddings for taxonomy expansion have the following shortcomings: (1) the size of the boxes is not restricted, and the model produces meaningless boxes; (2) the model does not fully utilize the geometric information of the boxes. To address the above shortcomings, this paper proposes a taxonomy expansion model based on projecting entities as boxes: PEB-TAXO. Firstly, PEB-TAXO employs modified L1 regularization to constrain the box sizes in all dimensions, pushing the box sizes towards the preset minimum, thereby avoiding the generation of meaningless boxes by the model. Secondly, the model utilizes a box inclusion inference method: it infers the relationship between two entities through the relationship between two boxes in geometric space, thus fully exploiting the geometric information of the boxes for more accurate inferences. Finally, we conducted extensive experiments on two public datasets and verified that PEB-TAXO greatly improves performance over mainstream taxonomy expansion methods.

The core problems of Web Service Composition (WSC) are to satisfy user preferences while at the same time facilitate reasonable construction for WSC. This poses two problems. First, the user real preferences has not been fully expressed when qualitative preference is used to measure composite servi ce. Second, when considering whether the WSC is reliable, some studies use the trust value as the reference attribute of the composite service. However, this is not sufficient as an evaluation index. To solve these problems, we first use the neural network to adjust initial weights in qualitative preference, making qualitative preference accurately measure user preferences when it changes. Second, we redefine the service invocation structures based on the travel plan. Third, we propose a new indicator: availability. Based on the service invocation structures, global availability of WSC is obtained from availability of a single service. Finally, WSC in three aspects: qualitative, quantitative and availability are used to get the final optimal composite service by multi-objective optimization algorithm. Results show that our method is reasonable and efficient compared with other counterparts.

Aiming at compensating the lack of affect interaction between teachers and students in e-Learning systems/environments, this paper presents an architecture of interactive text-oriented affect compensation Mechanism in e-Learning. Based on which, a hierarchical emotion regulation agent (ERA) is proposed to perceive the emotion state of an individual or group from sentences, make decisions on how to adjust their emotions in a way of on-line discussion group, and control the compensation behaviors. In which, inspired by the strategies that have been widely used in the fields of psychology and sociology, the emotion regulation polices are designed for textual interaction applications. Moreover, two curves for emotion intensity change over time are introduced into predict the emotion intensity for individual and when a group effort takes, respectively. Finally, an emotion regulation chatroom is designed and implemented to validate and verify the proposed model and mechanism. Different from the video- and audio-data based emotion classification, our method is suitable for interactive texts that widely exist in e- Learning. The emotion regulation strategies can especially benefit the interacting sides in the text-based communication.

E-learning provides an unprecedented flexibility and convenience for learners via breaking the limitation of space-time. Most researchers are only concerned about the learner's cognitive and construct a great amount of substantive digital learning resources, however they neglect of the learners' affect in current e-learning systems. In this paper, we focus primarily on the negative affect of learners, and propose an emotion compensation mechanism associated with trust and personality traits in traditional recommender technology. First, we analyze the difference between emotion compensation and traditional recommender. Next, the score of trust is calculated with historical behavior, otherwise depend on similarity of personality traits without historical experience. We use trustworthiness to replace similarity as prediction weight in trust filtering process. At last we do experiments with data collected in previous system named emotion-chatting. Compared with results of experiments between traditional recommender and trust-personality recommender, the average of accuracy is improved 4 points in percentage.