Xu Li

Xu Li

Ph.d candidate

Publications

WTAN: Wavelet Transform-based Attention Network for Tail Strike Prediction and Interpretable Analysis

Published in Submitted, 2025

Tail strike incidents are common flight safety events that typically occur during the landing phase. With the widespread deployment of Quick Access Recorders (QAR), large volumes of high-dimensional and heterogeneous flight data have become available. However, existing studies on flight safety often struggle to effectively capture multi-scale frequency temporal patterns and complex interrelationships among various flight parameters, resulting in suboptimal performance in tail strike prediction. To address the above issues, we propose a Wavelet Transform-based Attention Network, termed WTAN, to predict and interpret tail strike incidents, thereby enhancing flight safety. Specifically, we utilize Temporal Convolutional Networks (TCN) and Gated Recurrent Units (GRU) to standardize flight parameters with varying sampling frequencies to a unified sample length. To further enrich temporal frequency feature representation, we integrate wavelet transform with a multi-head attention mechanism, allowing the model to jointly exploit both temporal and frequency-domain information for more comprehensive pattern recognition. Additionally, we employ graph attention network to model the complex interrelationships between different flight parameters. We validate the effectiveness of WTAN on a real-world QAR dataset consisting of 44,808 Airbus A321 flight records. Experimental results demonstrate that WTAN significantly outperforms the baseline models in predictive accuracy and offers interpretable insights into the causes of flights with high tail strike risk by visualizing the contributions of different flight parameters.

Recommended citation: Xiong, D., Li, X., Song, E., et al. (2025). WTAN: Wavelet Transform-based Attention Network for Tail Strike Prediction and Interpretable Analysis.(Submitted)

Query-Guided Temporal Confidence Network for Noisy Temporal Knowledge Graph Reasoning

Published in Submitted, 2025

The presence of noise in real-world Temporal Knowledge Graphs (TKGs) severely hampers the performance of reasoning models. While prior research primarily addresses input noise—erroneous or irrelevant historical facts—by refining temporal aggregation, it largely overlooks the detrimental impact of label noise, where inaccurate future facts are used as training supervision. Furthermore, most existing methods treat all historical information equally, ignoring its varying relevance to the specific query, which weakens their ability to effectively filter noise. To address these issues, we propose QGTC-Net, a Query-Guided Temporal Confidence Network that robustly performs TKG reasoning under both input and label noise. QGTC-Net introduces two novel components: (1) a Query-Guided Learning Mechanism (QGLM) that dynamically extracts query-relevant information from historical snapshots via a dual-path architecture, enhancing targeted and query-sensitive history modeling; and (2) a Confidence-Aware Optimization Mechanism(CAOM) that evaluates the reliability of training labels by jointly assessing their temporal alignment and local structural plausibility, allowing the model to selectively optimize on high-confidence facts. Extensive experiments on four widely-used TKG benchmarks and their noisy counterparts demonstrate that QGTC-Net consistently outperforms state-of-the-art baselines across both clean and noisy settings, underscoring its effectiveness and robustness in real-world scenarios.

Recommended citation: Liao, L., Zheng, L., Shang, J., Li, X., Zhong, J., & Wei, K.. (2025). Query-Guided Temporal Confidence Network for Noisy Temporal Knowledge Graph Reasoning.(Submitted)

CCAN: Channel-wise Clustering and Attention Networks for Interpretable Hard Landing Prediction

Published in Submitted, 2025

Hard landing incidents are common flight safety events during the landing phase and are of significant concern in the aviation industry. Recent hard landing prediction methods tend to overemphasize temporal features while overlooking the landing process along the altitude dimension. In addition, they often fail to capture the interdependencies between contributing factors, limiting their ability to provide interpretable predictions. To address the above issues, we propose a novel Channel Clustering based Attention Network, termed CCAN, to predict hard landing incidents and identify their potential causes. Specifically, we resample and interpolate different flight parameters along the altitude dimension to align the landing process across different flights into a common reference frame. Subsequently, we design a channel clustering module that groups flight parameters into distinct clusters based on a predefined assignment threshold. Then, we employ graph attention network to capture the dependencies between different flight parameters within and across clusters. To further reveal the interactions between flight parameters throughout the landing process, we incorporate attention mechanism into Gated Recurrent Units (GRUs) to extract informative temporal features. We conducted experiments on a real-world QAR dataset with 44,729 Airbus A321 flights. Experimental results demonstrate that CCAN outperforms the baseline models in hard landing predictions and offers interpretability for hard landings by visualizing the dependencies between flight parameters and their interactions during the landing process.

Recommended citation: Zhang, H., Sun, H., Liu, Y., Zhao, X., Li, X., Shang, J., Zheng, L.. (2025). CCAN: Channel-wise Clustering and Attention Networks for Interpretable Hard Landing Prediction.(Submitted)

Fine-Grained Time and Hidden Feature Learning for Interpretable Hard Landing Prediction Based on QAR Data

Published in IEEE Transactions on Intelligent Transportation Systems, 2025

Hard landings, as a common type of aviation incident, have consistently attracted the attention of airlines and aviation authorities. In recent years, the widespread adoption of Quick Access Recorder (QAR) systems has led numerous researchers to focus on predicting hard landing events through the analysis of QAR data. However, most studies treat QAR data as standard time series without fully accounting for its unique characteristics. Unlike typical time series, QAR data exhibits limited periodicity and trends, making it challenging for traditional modeling approaches to capture its complex patterns. Furthermore, model interpretability, as an essential aspect for practical deployment and decision-making, remains insufficiently explored. To address these issues, we propose a Fine-Grained Time and Hidden Feature Learning model for Interpretable Hard Landing Prediction based on QAR Data (TF-QAR). Specifically, we introduce a novel fine-grained temporal aggregation module, which dynamically extracts the importance of each time step through learnable parameters, to efficiently model the temporal dependencies in QAR data. Additionally, we develop a feature aggregation module that introduces a learnable adjacency matrix to model the significance of flight features and their interrelationships, revealing not only key parameters that directly influence hard landings, but also hidden parameters that are indirectly related. We conducted extensive experiments using a dataset of 37,929 real A320 flight segments in China. The results demonstrate that our model outperforms existing state-of-the-art baselines. Moreover, by visualizing the learnable parameters, TF-QAR provides interpretable insights valuable for pilot decision-making, offering practical support for the prevention and management of hard landing events.

Recommended citation: Cai, J., Shang, J., Li, X., Li, C., & Zheng, L. (2025). Fine-Grained Time and Hidden Feature Learning for Interpretable Hard Landing Prediction Based on QAR Data. IEEE Transactions on Intelligent Transportation Systems.
Download Paper

Crossperformer: Cross-period Transformer for Time Series Forecasting

Published in Submitted, 2025

Time series data often exhibit multiple periodic characteristics that are critical for accurate forecasting and understanding system dynamics. Recently, the effectiveness of existing Transformer-based time series forecasting methods has been challenged, these methods typically embed multiple variables at the same timestamp using Transformers, relying on self-attention mechanisms to predict future horizons based solely on correlations between embedded tokens. However, this approach results in attention maps that fail to capture the periodic nature of time series data effectively. To address this limitation, we propose cross-period Transformer (Crossperformer), an enhanced model designed specifically for multi-period time series data. We introduce a periodic segment embedding method and develop a cross-period attention mechanism, enabling Crossperformer to explicitly capture periodic patterns and effectively model cross-period dependencies among variables. To further advance temporal representation learning, we employ a feed-forward network with periodic activation functions to learn generalized periodic patterns, thereby enhancing both prediction accuracy and generalization capabilities of the framework. Experimental results on multiple real-world benchmark datasets demonstrate that Crossperformer achieves state-of-the-art performance in long-term time series forecasting tasks. Our findings indicate that Transformers are inherently unsuitable for time series forecasting, since they cannot handle the periodicity in time series data. This research opens new avenues for future exploration in this domain.

Recommended citation: * (2025). Crossperformer: Cross-period Transformer for Time Series Forecasting. (Submitted)

Causal Mamba for Multivariate Time Series Forecasting

Published in Submitted, 2025

Multivariate time-series forecasting (MTSF) is essential for decision-making in various domains. As modern systems grow in complexity, effectively modeling inter-variable dependency has become increasingly important. Mamba, a recent state space model known for its ability to capture long-range dependencies, has also been adapted to model correlations among variables. However, vanilla-Mamba ignores causal direction/sequential order between variables, which limits their effectiveness in discovering complex dependency patterns. To address this, we propose Causal Mamba, a novel Mamba model for MTSF tasks. Specifically, we incorporate a Granger causality and topological sorting method to extract the causal order of variables. Subsequently, we employ the causal order information to modify vanilla-Mamba into Dynamic Mamba, which captures directional inter-variable dependency. Then we use TCN and FFN to model temporal dependency. Comprehensive experiments conducted on six real-world datasets demonstrate the efficacy of our model in improving forecasting performance, and validate the effectiveness of causal order of variables compared to vanilla-Mamba.

Recommended citation: * (2025). Causal Mamba for Multivariate Time Series Forecasting. (Submitted)

MVAN: Towards Interpretable Joint Prediction of Multiple Flight Safety Events with Multi-View Attention Network

Published in Submitted, 2025

Recently, the use of multivariate flight data for safety event analysis has attracted significant attention from both academia and industry. However, most existing studies focus on individual safety events, overlooking the joint prediction and interpretation of multiple interrelated events, which hinders effective risk management given the inherent coupling among such events. Moreover, the complex nature of flight data, characterized by multi-scale temporal dynamics and intricate inter-parameter correlations, poses substantial analytical challenges. To address these issues, we propose MVAN, a novel Multi-View Attention Network that jointly predicts and interprets multiple flight safety events by modeling the flight data from temporal, parametric, and event-specific perspectives. MVAN introduces two specialized attention mechanisms: Time Trend Attention, which captures temporal patterns and abrupt changes through various convolutional operations, and Gaussian-Enhanced Attention, which models both global and local inter-parameter dependencies via a hybrid approach. Furthermore, we design a Multi-Task Variable Selection Module to dynamically identify shared and task-specific features, thereby enhancing interpretability. Extensive experiments on a real-world flight dataset demonstrate that MVAN not only outperforms baseline models in predicting hard landing and tail strike events, but also effectively uncovers the coupling relationships between these events. In-depth case studies further validate the interpretability of the model by revealing both shared and event-specific parameter dependencies, offering actionable insights for flight operations and risk mitigating strategies.

Recommended citation: Li, C., Li, X., Zheng, L., Shang, J., et al. (2025). MVAN: Towards Interpretable Joint Prediction of Multiple Flight Safety Events with Multi-View Attention Network.(Submitted)

A Dual Two-Stage Attention-based Model for interpretable hard landing prediction from flight data

Published in Engineering Applications of Artificial Intelligence, 2025

Hard landings are a significant safety concern in aviation, with potential consequences ranging from poor passenger experiences to serious injuries or fatalities. Predicting and explaining hard landing events are equally important for enhancing flight safety, the former makes it possible to give proactive warnings, while the latter helps pilots identify the reasons and refine their skills. However, existing studies generally lack a comprehensive consideration for the fine-grained characteristics of flight data containing both inter-temporal and inter-parametric relationships, resulting in suboptimal prediction performance. In addition, most of existing approaches aim at improving the prediction performance but fail to provide interpretability for the causes of hard landing. To address the above problems, we propose DUTSAM, a DUal Two-Stage Attention-based interpretable Model for hard landing prediction from quick access recorder (QAR) data. The model consists of dual parallel modules, each of which combines a convolutional feature encoder and a two-stage attention mechanism. The two encoders capture fine-grained characteristics by encoding multivariate data from temporal domain and parametric domain respectively. After that, the dual two-stage attention mechanism captures the inter-temporal and inter-parametric correlations in reverse order to predict hard landing and provide interpretation from both temporal and parametric perspectives. Experimental results on a real QAR dataset with 37,920 flights show that DUTSAM achieves better prediction performance compared with other state-of-the-art baselines in terms of Precision, Recall, and F1-score. Additionally, case study demonstrates that DUTSAM can uncover key flight parameters and moments strongly correlated to the hard landing events.

Recommended citation: Shang, J., Li, X., Zhang, R., Zheng, L., Li, X., Zhang, R., ... & Sun, H. (2025). A Dual Two-Stage Attention-based Model for interpretable hard landing prediction from flight data. Engineering Applications of Artificial Intelligence, 154, 110911.
Download Paper

MGAN: Multi-scale Graph Attention Network for Flight Safety Incidents Detection

Published in Submitted, 2025

The approach and landing phases are the stages with the highest frequency of safety incidents during the flight of a civil aviation aircraft. The factors influencing these incidents are often reflected in flight data, making the analysis of such data crucial for detecting incidents. The Quick Access Recorder (QAR), an essential onboard device in commercial aircraft, collects vast amounts of flight data generated by various Internet of Things (IoT) devices. The integration of QAR data with artificial intelligence technologies has made significant progress in the field of aviation safety analysis. However, existing detection approaches frequently struggle to accurately model the dependencies and interactions among flight variables, which exhibit high dimensionality and varying temporal frequencies, thereby undermining detection accuracy. Moreover, numerous approaches depend on opaque black-box models, which offer limited interpretability. To address these challenges, we introduce the Multi-scale Graph Attention Network (MGAN), a novel anomaly detection framework that integrates multivariate flight data sampled at varying frequencies to identify events and provide interpretable insights into their causes. MGAN proposes a Hybrid Temporal Network (HTN), which combines Gated Recurrent Units (GRU) with Dilated Causal Convolutions to capture long-range dependencies and complex interactions among flight variables. Additionally, the Cross-frequency Graph Attention Network (GAT) models dependencies across different sampling frequencies, enabling the extraction of multi-level feature representations. Extensive experiments on the QAR dataset of 44,808 Airbus A321 flights operated by Chinese airlines demonstrate that MGAN significantly outperforms state-of-the-art models in anomaly detection tasks, while also providing interpretative analyses of the underlying causes of safety incidents.

Recommended citation: Yan, W., Li, X., Zhao, X., et al. (2025). MGAN: Multi-scale Graph Attention Network for Flight Safety Incidents Detection.(Submitted)

IATCN: Interpretable Attention-Enhanced Temporal Convolutional Network for Multiple Flight Safety Incidents Prediction

Published in Submitted, 2025

Flight safety incidents, such as hard landings and tail strike risks, represent critical concerns during the landing phase. Although Quick Access Recorder (QAR) systems collect extensive multivariate flight data, previous studies have faced challenges in effectively modeling temporal dependencies and providing model interpretability, limiting their ability to predict multiple safety incidents simultaneously. To address this issue, we propose a novel model, named IATCN (Interpretable Attention-Enhanced Temporal Convolutional Network), for predicting multiple flight safety incidents, including both hard landing and tail strike risk incidents. Specifically, IATCN combines temporal convolutional networks (TCNs) with attention mechanisms, enabling the model to effectively capture temporal dependencies in flight data while enhancing the focus on key features. Additionally, we are the first to integrate attention mechanisms with Grad-CAM to improve the interpretability and visualization of flight safety incidents. This combination offers valuable insights into the model’s decision-making process by highlighting the importance of specific flight parameters. We conduct experiments on real-world QAR datasets of 37,904 Airbus A320 flight samples. Experimental results demonstrate that the IATCN model outperforms state-of-the-art baselines and provides effective interpretability by visualizing the importance of parameters, offering valuable insights for flight safety analysis.

Recommended citation: Song, L., Sun, H., Li, X., et al. (2025). IATCN: Interpretable Attention-Enhanced Temporal Convolutional Network for Multiple Flight Safety Incidents Prediction.(Submitted)

DGAN: Flight sensor data anomaly detection based on dual-view graph attention network

Published in IEEE Sensors Journal, 2025

A large number of flight data is collected by various sensors during flight, including altitude, speed, pitch, etc. Flight unsafe events, such as hard landings, can be reflected in flight anomaly data. Thus, flight data anomaly detection is crucial for aviation safety, enabling the identification of deviations from normal operations. However, due to the complexity and high dimensionality of the flight data, existing methods often fail to adequately extract temporal and variable information. As a result, these methods exhibit poor detection performance, especially when the anomaly data is scarce. To address these issues, we propose a novel Dual-view Graph Attention Network (DGAN) model for flight data anomaly detection. Specifically, two parallel Graph Attention (GAT) layers are employed to create embeddings that are rich in temporal correlations and variable dependencies. And TCN-based autoencoder module is proposed for data augmentation on these embeddings to extract two views consistent with the underlying data patterns. Furthermore, DGAN constructs pairs of positive and negative samples both within a single view and between different views. A dual-view contrastive training strategy is proposed to construct contrastive loss, which can learn a good representation of flight sequence and ensure effective anomaly detection even in cases of scarce anomaly data. We conduct experiments on a dataset comprising 44,808 real-world flight samples from the Airbus A321. The results demonstrate that the performance of DGAN exceeds state-of-the-art (SOTA) models. Moreover, we examine the effectiveness of DGAN through further visualizations and case studies.

Recommended citation: Yan, W., Li, X., Zheng, L., Shang, J., Wu, L., Liu, H., ... & Qian, Y. (2025). DGAN: Flight data anomaly detection based on dual-view graph attention network. IEEE Sensors Journal.
Download Paper

MultiSafe: Multiple Flight Safety Events Prediction Based on Interpretable Deep Multi-Task Learning

Published in Submitted, 2025

Flight safety remains a central concern in the civil aviation industry. Recently, increasing attention has been given to leveraging high-dimensional temporal flight data, typically collected by Quick Access Recorders (QAR), to predict safety events. However, most existing studies focus on individual safety events, overlooking the latent and often complex correlations among multiple events. For instance, during landing, a maneuver intended to reduce the risk of one event may inadvertently raise the risk of another. Predicting multiple events introduces three main challenges: (1) handling flight parameters recorded at inconsistent sampling rates, (2) learning task-specific parameter importance to enhance model interpretability, and (3) modeling complex temporal dependencies to improve prediction accuracy. To address these issues, we propose MultiSafe, a deep multi-task learning model for predicting and interpreting multiple flight safety events. First, to process parameters with heterogeneous frequencies, we introduce a Multi-Scale Shared Encoder using convolutional layers with adaptive kernel sizes to generate unified representations across tasks. Next, a Parameter Selector based on gating networks learns task-specific parameter importance, enabling interpretable predictions by identifying key operational features. Finally, a Temporal Decoder with a fine-grained attention mechanism captures intricate temporal dependencies among parameters. Experiments on a dataset of 37,518 A320 flight records show that MultiSafe outperforms state-of-the-art baselines in prediction accuracy. Moreover, its interpretability offers actionable insights to assist pilots in enhancing flight safety.

Recommended citation: Huang, Y., Shang, J., Li, X., et al. (2025). MultiSafe: Multiple Flight Safety Events Prediction Based on Interpretable Deep Multi-Task Learning.(Submitted)

ERD-Net: Modeling entity and relation dynamics for Temporal Knowledge Graph reasoning

Published in Knowledge-Based Systems, 2025

Temporal Knowledge Graph (TKG) has garnered significant attention and applications due to its immense potential and impact in event prediction. Existing approaches primarily focus on learning low-dimensional embeddings of entities and relations to predict valid triples. Despite notable advancements, these methods face challenges such as inadequate modeling of relation dynamics and limited exploration of non-repeated patterns. To address these issues, we propose the Entity and Relation Dynamic representation learning Network (ERD-Net), a novel framework that combines dynamic representation learning with a copy-generation mechanism to enhance TKG reasoning. ERD-Net begins by learning intrinsic embeddings for entities and relations, capturing their time-invariant properties. These embeddings are subsequently frozen during the dynamic learning phase to ensure stability. In the dynamic representation learning stage, we extend beyond entity dynamics to explore relation dynamics, distinguishing between global and local categories to evaluate their distinct contributions. Furthermore, a copy-generation mechanism is introduced in the decoder, enabling simultaneous modeling of both historical repeated patterns and novel non-repeated patterns. Comprehensive experiments on public benchmarks validate the efficacy of our approach. The results demonstrate that ERD-Net achieves state-of-the-art performance in TKG reasoning by effectively integrating dynamic representation learning with copy-generation mechanisms. Notably, our findings reveal that improving the prediction of non-repeated patterns can significantly enhance performance, highlighting a promising direction for future research in TKG reasoning.

Recommended citation: Liao, L., Zheng, L., Chen, F., Shang, J., Li, X., & Li, W. (2025). ERD-Net: Modeling entity and relation dynamics for Temporal Knowledge Graph reasoning. Knowledge-Based Systems, 317, 113404.
Download Paper

MDGNN: Multiple Flight Safety Incidents Prediction Model Based on Dynamic Graph Neural Networks

Published in IEEE Transactions on Intelligent Transportation Systems, 2025

Flight safety incidents, such as hard landings and tail strike risks, represent critical concerns during the landing phase. Although Quick Access Recorder (QAR) systems collect extensive multivariate flight data, previous studies have faced challenges in effectively modeling the complex interdependencies between flight parameters, which has limited their ability to predict multiple safety incidents simultaneously. To address this issue, we propose a novel model, named MDGNN, to capture hidden spatio-temporal dependencies and predict both hard landing and tail strike risk incidents. Specifically, we employ temporal convolutional networks (TCNs) to extract both localized representations and long-term temporal trends from multivariate flight data, ensuring the standardization of flight parameters across varying frequencies. Additionally, we are the first to construct a dynamic graph to model temporal relationships, applying a dynamic graph neural network and a temporal convolution module to accurately capture intricate spatial and temporal dependencies. Extensive experiments conducted on 37,904 Airbus A320 flight samples demonstrate that the MDGNN model surpasses state-of-the-art baselines with high prediction accuracy. Furthermore, a case study visualizing key flight parameters highlights the model’s ability to reveal the root causes of safety exceedances, offering valuable insights for flight safety analysis.

Recommended citation: Song, L., Li, X., Liu, H., Wu, L., Sun, H., Zheng, L., & Shang, J. (2025). MDGNN: Multiple Flight Safety Incidents Prediction Model Based on Dynamic Graph Neural Networks. IEEE Transactions on Intelligent Transportation Systems.
Download Paper

MRI: Memory Retention Mechanism for Interpretable Time Series Forecasting

Published in Submitted, 2025

Interpretability in time series forecasting is crucial for building reliable models in mission-critical applications. While most studies prioritize improving prediction accuracy through complex model designs, few explore how models derive predictions from data patterns, limiting the understanding of model behavior across diverse datasets. Addressing this gap requires solving two key challenges: enhancing interpretability by retaining critical information and maintaining non-linearity for high forecasting performance. To this end, we propose MRI, a Memory Retention mechanism for Interpretable time series forecasting. Inspired by human brain’s forgetting mechanism, we design a novel memory retention module to discard irrelevant information and a memory recall module to retain sparse but important data, which largely improve model interpretability. To preserve non-linearity, we introduce learnable nonlinear functions with theoretical convergence guarantees, eliminating the need for manual selection of activation function. The combination of linear weights and learnable nonlinear functions, with well-defined relevance importance, enables both model-level and instance-level interpretations. Extensive experiments on nine real-world datasets demonstrate that MRI achieves state-of-the-art performance with competitive computational efficiency. Additionally, MRI demonstrates strong robustness against Gaussian noise and effectively captures periodic and trend patterns, providing meaningful model-level and instance-level explanations. The nature-inspired approach offers a new paradigm for building interpretable and robust deep learning models.

Recommended citation: Li, X., Zheng, L., Shang, J., Liao, L., Zhang, J., & Zhong, J. (2025). MRI: Memory Retention Mechanism for Interpretable Time Series Forecasting. IEEE Transactions on Pattern Analysis and Machine Intelligence.(Submitted)

Boosting Low-budget Active Learning with Label Calibration and Unsupervised Representations

Published in Submitted, 2025

In active learning research, low-budget active learning poses a significant challenge, primarily due to the scarcity of labeled examples available for training the target model. This limitation often leads to the model’s tendency to make approximately stochastic predictions for examples with confusing category features, a phenomenon known as catastrophic forgetting. To address this challenge, this paper proposes a novel and unified framework that leverages label calibration and unsupervised representations, aiming to boost the performance of active learning algorithms in low-budget scenarios. Specifically, to cope with the scarcity of labeled data, we perform pre-training on the entire unlabeled dataset to generate unsupervised representations that are rich in geometric information. These representations are then used by diversity-based active learning algorithms to query examples for labeling. To address model instability, we creatively calibrate pseudo-labels of unlabeled examples using statistical information derived from model outputs during training. These calibrated pseudo-labels, combined with the unsupervised representations, are then fed into a KNN classification model to generate predictions for test examples. Theoretical analysis reveals that our approach exhibits a tighter bound on generalization error, and empirical results demonstrate its superior effectiveness in boosting the performance of active learning algorithms under low-budget constraints.

Recommended citation: Han, Y., Shang, J., Li, R., Li, X., Liao, L., Zheng, L., & Min, G.. (2025). Boosting Low-budget Active Learning with Label Calibration and Unsupervised Representations. IEEE Transactions on Pattern Analysis and Machine Intelligence. (Submitted)

Context-Aware Learning and Pattern Decomposition for Temporal Knowledge Graph Reasoning

Published in Submitted, 2025

Graph neural network-based approaches have achieved remarkable success in temporal knowledge graph (TKG) reasoning. Despite these advances, two critical challenges remain: (1) inadequate modeling of local contextual dynamics, which limits the adaptability of entity and relation representations to specific queries; and (2) inadequate mechanisms for handling emerging patterns, i.e., novel interactions absent from historical data, which reduces predictive performance in dynamic environments. To address these limitations, we propose TCDR-PD, a Temporal and Contextual Dynamic Representation Network with Pattern Decomposition. TCDR-PD introduces a Temporal and Contextual Dynamic Representation (TCDR)module to capture both global temporal trends and query-specific contextual dynamics, enabling more precise embeddings. Additionally, the Pattern Decomposition (PD) prediction module explicitly disentangles the prediction of recurring and emerging patterns, enabling tailored strategies to improve reasoning performance. Experiments on four benchmark datasets demonstrate that TCDR-PD outperforms state-of-the-art methods, effectively supporting stable reasoning over evolving TKGs.

Recommended citation: * (2025). Context-Aware Learning and Pattern Decomposition for Temporal Knowledge Graph Reasoning.(Submitted)

DUVET: Dual View Enhanced Transformer for Multivariate Flight Time series Anomaly Detection

Published in Submitted, 2024

In recent years, the prevalence of modern flight data recording systems such as Quick Access Recorder (QAR), has made it viable to improve flight safety by analyzing large volumes of flight parameters. However, existing studies mostly concentrate on specific safety events and heavily depend on expert-labeled data where the labels are highly scarce and imbalanced. To fully utilize valuable information from the tremendous unlabeled data, in this paper we focus on uncovering abnormal flight patterns and define it as an self-supervised time series anomaly detection problem, enabling us to discover potential and unexpected safety risks. To this end, we propose DUVET, a reconstructionbased Dual View Enhanced Transformer model for multivariate flight time series anomaly detection. Specifically, DUVET adopts Patching and Channel Independence techniques to reduce model parameters and improve Transformer’s computational efficiency of attention mechanism while preserving its semantic representation capacity. Based on the observation that flight safety usually relates to both long-term patterns and short-term deviations, we pioneer a new Bell Attention mechanism, and combine it with Transformer’s self-attention, to capture local and global dependencies respectively, which largely enhances the model reconstruction ability. We first evaluate DUVET on six public real-world datasets and one QAR dataset, demonstrating that DUVET achieves state-of-the-art performance in general time series anomaly detection task, and the ablation study clearly shows the vital role of Bell Attention. Further experiments on the QAR dataset show that DUVET meets key criteria of anomaly detection for flight safety in terms of stability, training efficiency and detection accuracy. Case study shows that DUVET can effectively uncover hidden flight safety risks.

Recommended citation: Li, C., Li, X., Chen, H., Zheng, L., Sun, H., Shang, J. (2024). DUVET: Dual View Enhanced Transformer for MultivariateFlight Time series Anomaly Detection.

ATPF: An Adaptive Temporal Perturbation Framework for Adversarial Attacks on Temporal Knowledge Graph

Published in IEEE Transactions on Knowledge and Data Engineering, 2024

Robustness is paramount for ensuring the reliability of knowledge graph models in safety-sensitive applications. While recent research has delved into adversarial attacks on static knowledge graph models, the exploration of more practical temporal knowledge graphs has been largely overlooked. To fill this gap, we present the Adaptive Temporal Perturbation Framework (ATPF), a novel adversarial attack framework aimed at probing the robustness of temporal knowledge graph (TKG) models. The general idea of ATPF is to inject perturbations into the victim model input to undermine the prediction. Firstly, we propose the Temporal Perturbation Prioritization (TPP) algorithm, which identifies the optimal time sequence for perturbation injection before initiating attacks. Subsequently, we design the Rank-Based Edge Manipulation (RBEM) algorithm, enabling the generation of both edge addition and removal perturbations under black-box setting. With ATPF, we present two adversarial attack methods: the stringent ATPF-hard and the more lenient ATPF-soft, each imposing different perturbation constraints. Our experimental evaluations on the link prediction task for TKGs demonstrate the superior attack performance of our methods compared to baseline methods. Furthermore, we find that strategically placing a single perturbation often suffices to successfully compromise a target link.

Recommended citation: Liao, L., Zheng, L., Shang, J., Li, X., & Chen, F. (2024). ATPF: An Adaptive Temporal Perturbation Framework for Adversarial Attacks on Temporal Knowledge Graph. IEEE Transactions on Knowledge and Data Engineering.
Download Paper

Decomposition with feature attention and graph convolution network for traffic forecasting

Published in Knowledge-Based Systems, 2024

Traffic forecasting is a crucial task for enhancing the quality and efficiency of intelligent transportation systems. In recent years, several neural networks have been proposed to tackle this challenge. However, traffic data itself has complex composition and intertwined temporal patterns. Currently, the absence of effective methods to capture the inherent properties and overall profile of traffic data has become a bottleneck in improving the traffic prediction capabilities of models. To address these issues, we introduce a novel model, Decomposition with Feature Attention and Graph Convolution Network (DFAGCN), for Traffic Forecasting. To highlight the inherent properties of complicated traffic data, we present a unique Decomposition with Feature Attention (DFA) method to decompose the traffic data into seasonality, trend, and remainder. On the one hand, we introduce a new method to filter out noise and possible outliers from time series data. On the other hand, we employ an attention mechanism to make our decomposition method adaptively extract the trend. Subsequently, we utilize spatial–temporal attention to simulate the dynamic spatial–temporal correlations of traffic data. Finally, we use spatial–temporal convolution to extract local dependencies of traffic data. Experiments conducted on four real-world traffic forecasting datasets demonstrate that DFAGCN achieves state-of-the-art performance. The results show that the application of time series decomposition to traffic data significantly enhances the accuracy of traffic prediction.

Recommended citation: Liu, Y., Wu, X., Tang, Y., Li, X., Sun, D., & Zheng, L. (2024). Decomposition with feature attention and graph convolution network for traffic forecasting. Knowledge-Based Systems, 112193.
Download Paper

IMTCN: An Interpretable Flight Safety Analysis and Prediction Model Based on Multi-Scale Temporal Convolutional Networks

Published in IEEE Transactions on Intelligent Transportation Systems, 2023

Flight safety is a key issue in the aviation industry. Recently, with the prevalence of flight data recording systems, some deep learning-based studies have been devoted to predicting safety incidents based on flight data. However, these studies, although they exhibit higher prediction accuracy, have largely neglected the interpretability analysis of safety incidents which is of great concern to airlines and pilots. To address this issue, we define flight safety prediction as a multiscale time series classification problem and propose an interpretable model named IMTCN to provide both accurate predictions and high interpretability of flight safety. First, multiple temporal convolutional networks (TCNs) are utilized to capture local representations and long effective histories from multivariate flight data. Because different flight parameters are collected with diverse sampling frequencies, multiple TCNs are used to handle these parameters separately. Then, we creatively adapt the class activation mapping (CAM) method, which has been used for interpretation in image classification, and combine it with the TCN to provide flight data interpretability. The established model can pinpoint key flight parameters and corresponding moments that contribute most to safety incidents. Experimental results on a real-world dataset with 37,943 Airbus A320 aircraft flights show that our model outperforms the baselines on the task of exceedance classification and prediction 2 seconds and 4 seconds in advance, and case studies demonstrate its superb interpretability for flight safety analysis.

Recommended citation: Li, X., Shang, J., Zheng, L., Wang, Q., Liu, D., Liu, X., ... & Sun, H. (2024). IMTCN: An Interpretable Flight Safety Analysis and Prediction Model Based on Multi-Scale Temporal Convolutional Networks. IEEE Transactions on Intelligent Transportation Systems.
Download Paper

SDTAN: Scalable Deep Time-Aware Attention Network for Interpretable Hard Landing Prediction

Published in IEEE Transactions on Intelligent Transportation Systems, 2023

Hard landing, as one of the most frequent flight safety incidents during the landing stage, is highly concerned by the aviation industry. Recently, the popularization of Quick Access Recorder (QAR), a modern flight data recording system, has made it possible to collect large volume of flight parameters and incorporate state-of-the-art AI technologies to improve flight safety. However, due to the complex, multivariate, and highly specialized nature of QAR data, most existing studies either suffer from information loss caused by rough feature extraction methods, or rely solely on black-box models with no interpretations, making themselves difficult to achieve satisfactory performance in terms of prediction and explainability. To address this issue, we propose a novel attention-driven model named SDTAN (Scalable Deep Time-Aware Attention Network), which can accurately predict hard landing events and provide interpretable insights to help reveal the possible reasons leading to the events. Specifically, SDTAN fully captures information to learn the local representations of parameters, and leverages the time-interval attention mechanism to focus on the entire temporal pattern of flight over the relevant time intervals. It further re-encodes the representations of parameters in a global view and learns the global effect of parameters on the predicted output to uncover the ones which strongly indicate the flight safety status, enabling both high prediction accuracy and qualitative interpretability. We conduct experiments on real-world QAR datasets of 37,920 Airbus A320 flight samples. Experimental results demonstrate that SDTAN outperforms other state-of-the-art baselines and provides effective interpretability by visualizing the importance of parameters.

Recommended citation: Chen, H., Shang, J., Zheng, L., Li, X., Liu, X., Sun, H., ... & Yu, L. (2023). SDTAN: Scalable Deep Time-Aware Attention Network for Interpretable Hard Landing Prediction. IEEE Transactions on Intelligent Transportation Systems, 24(9), 10211-10223.
Download Paper

PathSAGE: Spatial Graph Attention Neural Networks with Random Path Sampling

Published in Neural Information Processing, 2021

Graph Convolutional Networks (GCNs) achieve great success in non-Euclidean structure data processing recently. In existing studies, deeper layers are used in CCNs to extract deeper features of Euclidean structure data. However, for non-Euclidean structure data, too deep GCNs will confront with problems like “neighbor explosion” and “over-smoothing”, it also cannot be applied to large datasets. To address these problems, we propose a model called PathSAGE, which can learn high-order topological information and improve the model’s performance by expanding the receptive field. The model randomly samples paths starting from the central node and aggregates them by Transformer encoder. PathSAGE has only one layer of structure to aggregate nodes which avoid those problems above. The results of evaluation shows that our model achieves comparable performance with the state-of-the-art models in inductive learning tasks.

Recommended citation: Ma, J., Li, J., Li, X., & Li, X. (2021). PathSAGE: Spatial Graph Attention Neural Networks with Random Path Sampling. In Neural Information Processing: 28th International Conference, ICONIP 2021, Sanur, Bali, Indonesia, December 8–12, 2021, Proceedings, Part II 28 (pp. 111-120). Springer International Publishing.
Download Paper

CurveCluster+: Curve Clustering for Hard Landing Pattern Recognition and Risk Evaluation Based on Flight Data

Published in IEEE Transactions on Intelligent Transportation Systems, 2021

Hard landing is a typical flight safety incident, and interpretability plays an important role in flight safety research. However, existing studies failed to provide good interpretability of the reasons for hard landing incidents and suffer from low prediction accuracy. To address the above problems, in this paper we propose CurveCluster+, a curve clustering method based on quick access recorder (QAR) data for hard landing risk evaluation. Specifically, we first conduct an in-depth analysis on hard landing flights by comparing key QAR parameter curves with the group behavior, based on which we establish a two-level hierarchical classification of hard landing incidents according to the hard landing patterns. Then we extract curve-level features from key QAR parameters through interpolation and resampling. After that we turn the classic K-means clustering into a semi-supervised algorithm by incorporating some expert experience and apply it on the curve-level features to automatically recognize the hard landing patterns. Finally, we propose a risk evaluation model based on the clustering results to discover high-risk flights from normal ones. We evaluate our method on a QAR dataset of 37,943 Airbus 320 aircraft flights. The results show that compared with other state-of-the-art data-driven methods, CurveCluster+ provides strong interpretability of hard landing incidents and exhibits good performance in recognizing hard landing patterns (the overall accuracy of our method reaches up to 92.99%). Moreover, it only requires a handful of hard landing samples to discover high-risk flights from tremendous normal landing flights, which is critical for flight safety warnings.

Recommended citation: Li, X., Shang, J., Zheng, L., Wang, Q., Sun, H., & Qi, L. (2021). Curvecluster+: Curve clustering for hard landing pattern recognition and risk evaluation based on flight data. IEEE Transactions on Intelligent Transportation Systems, 23(8), 12811-12821.
Download Paper

A Relation-Guided Attention Mechanism for Relational Triple Extraction

Published in 2021 International Joint Conference on Neural Networks (IJCNN), 2021

Relational triples are the essential parts of knowledge graphs, which can be usually found in natural language sentences. Relational triple extraction aims to extract all entity pairs with semantic relations from sentences. Recent studies on triple extraction focus on the triple overlap problem where multiple relational triples share single entities or entity pairs in a sentence. Besides, we find sentences may contain implicit relations, and it is challenging for most existing methods to extract implicit relational triples whose relations are implicit in the sentence. In this paper, we propose a relation-guided attention mechanism (RGAM) for relational triple extraction. Firstly, we extract subjects of all possible triples from the sentence, and then identify the corresponding objects under target relations with relation guidance. We utilize relations as prior knowledge instead of regarding relations as classification labels, and apply attention mechanism to obtain fine-grained relation representations, which guide extracted subjects to find the corresponding objects. Our approach (RGAM) can not only learn multiple dependencies in each triple, but also be suitable for extracting implicit relational triples and handling the overlapping triple problem. Extensive experiments show that our model achieves state-of-the-art performance on two public datasets NYT and WebNLG, which demonstrates the effectiveness of our approach.

Recommended citation: Yang, Y., Li, X., & Li, X. (2021, July). A relation-guided attention mechanism for relational triple extraction. In 2021 International Joint Conference on Neural Networks (IJCNN) (pp. 1-8). IEEE.
Download Paper

Active instance segmentation with fractional-order network and reinforcement learning

Published in The Visual Computer, 2021

In this paper, a novel model is proposed to segment image instance based on fractional-order chaotic synchronization network and reinforcement learning method. In the proposed model, fractional-order network is used for the preliminary image segmentation, which can obtain fine-grained information to provide a guiding strategy for the exploration of reinforcement learning; afterward, reinforcement learning method is committed to generate high-quality bounding contour curves for the object instances, which can combine the pixel features with local information in the image to improve the overall accuracy. Compared with other fractional-order models, the experimental results show that our proposed model achieves higher accuracy on the datasets of Pascal VOC2007 and Pascal VOC2012.

Recommended citation: Li, X., Wu, G., Zhou, S., Lin, X., & Li, X. (2022). Active instance segmentation with fractional-order network and reinforcement learning. The Visual Computer, 1-14.
Download Paper

A Deep learning Method for Landing Pitch Prediction based on Flight Data

Published in 2020 IEEE 2nd International Conference on Civil Aviation Safety and Information Technology (ICCASIT), 2021

With the development of the aviation industry, aircraft has increasingly become one of the most preferred longdistance transportation tools, and aviation safety incidents have attracted extensive attention. The key to dealing with aviation safety incidents is to accurately predict anomalies and potential hazards in advance and instruct pilots to perform corrective operations. As one of the safety incidents, tail strike may cause damage to the aircraft fuselage which may bring financial losses, or even threaten lives. However, there are few studies on tail strike in depth at present. In order to fill this gap, this paper mainly focuses on the tail strike risk, which is defined as the incident that the maximum pitch angle of the aircraft one second after and before touchdown exceeds a certain threshold. Specifically, we employ the LSTM model to make predictions of the maximum pitch angle with 22 parameters from QAR data. Extensive experiments based on a large-scale data show that the prediction model in this paper achieves the lowest MSE, MAE and the highest fitting coefficient R2-score, as compared to 9 traditional machine learning algorithms, which validates the effectiveness of our model in finding high risk flights.

Recommended citation: Chen, H., Shang, J., Zhao, X., Li, X., Zheng, L., & Chen, F. (2020, October). A deep learning method for landing pitch prediction based on flight data. In 2020 IEEE 2nd International Conference on Civil Aviation Safety and Information Technology (ICCASIT (pp. 199-204). IEEE.
Download Paper

CurveCluster: Automated Recognition of Hard Landing Patterns Based on QAR Curve Clustering

Published in 2019 IEEE Ubiquitous Intelligence & Computing (UIC), 2020

Flight safety is of vital importance to the aviation industry. As one of the most typical security events, hard landing is extremely concerned by airlines and related studies have received extensive attention in recent years. However, existing regression or risk based models either suffers from low prediction accuracy, or cannot provide good interpretability, making themselves impractical in real applications. To solve these problems, in this paper we propose CurveCluster: a curve clustering-based approach which is able to automatically recognize hard landing patterns from quick access recorder (QAR) data. We first provide a two-level hierarchical classification of hard landing events based on different hard landing patterns. Then we extract curve features from several key QAR parameters through interpolation and resampling. Finally, we apply K-means clustering algorithm on the curve features to automatically recognize the hard landing patterns. We test our approach on a dataset of 9,203 A320 flight QAR data samples and the overall recognition accuracy reaches up to 93.1%. Moreover, our results directly reflect the reasons of different types of hard landing events, which show strong interpretability.

Recommended citation: Li, X., Shang, J., Zheng, L., Liu, D., Qi, L., & Liu, L. (2019, August). CurveCluster: Automated recognition of hard landing patterns based on QAR curve clustering. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI) (pp. 602-609). IEEE.
Download Paper