Attention-based Method Research Articles

ADNNet: Attention-based deep neural network for Air Quality Index prediction

The Air Quality Index (AQI) is a crucial indicator for assessing the degree of atmospheric pollution. Accurately forecasting AQI is notably challenging due to the unpredictable weather conditions and the intricate interactions among various pollutants. To this end, we propose a AQI prediction framework, entitled ADNNet, to develop a streamlined attention-based method for AQI prediction. Specifically, it includes a deep neural network architecture that exclusively integrates multi-layer perceptron (MLP) and attention mechanism modules. This approach involves a novel design paradigm, where data information is first fed into the MLP layer to learn global features, which are then refined by the attention module to capture local features such as time delays and periodicity. Additionally, we incorporate a Bayesian hyperparameter optimization algorithm to make the trade-off between computational time and prediction accuracy efficient. To further enhance model performance, we apply exponential smoothing (ETS) and its inverse, named inverse-ETS, boosting the model’s predictive accuracy. We leverage air pollution datasets from various cities to verify the effectiveness of the proposed ADNNet. The results demonstrate that the proposed method significantly outperforms state-of-the-art (SOTA) models like LSTM, N-BEATS, Informer, Autoformer and VMD-TCN. The compressed model with reducing the mean absolute error (MAE) by 4.17%–16.12%, decreasing the root mean squared error (RMSE) by 3.52%–12.72%, and reducing the mean absolute scaled error (MASE) by 5.26%–17.28%.The source code of our model is available at https://github.com/xiankuiwu/AQI-Attention-based-DNN.

Attention-based approach to predict drug-target interactions across seven target superfamilies.

Drug-target interactions (DTIs) hold a pivotal role in drug repurposing and elucidation of drug mechanisms of action. While single-targeted drugs have demonstrated clinical success, they often exhibit limited efficacy against complex diseases, such as cancers, whose development and treatment is dependent on several biological processes. Therefore, a comprehensive understanding of primary, secondary and even inactive targets becomes essential in the quest for effective and safe treatments for cancer and other indications. The human proteome offers over a thousand druggable targets, yet most FDA-approved drugs bind to only a small fraction of these targets. This study introduces an attention-based method (called as MMAtt-DTA) to predict drug-target bioactivities across human proteins within seven superfamilies. We meticulously examined nine different descriptor sets to identify optimal signature descriptors for predicting novel DTIs. Our testing results demonstrated Spearman correlations exceeding 0.72 (P < 0.001) for six out of seven superfamilies. The proposed method outperformed fourteen state-of-the-art machine learning, deep learning and graph-based methods and maintained relatively high performance for most target superfamilies when tested with independent bioactivity data sources. We computationally validated 185676 drug-target pairs from ChEMBL-V33 that were not available during model training, achieving a reasonable performance with Spearman correlation >0.57 (P < 0.001) for most superfamilies. This underscores the robustness of the proposed method for predicting novel DTIs. Finally, we applied our method to predict missing bioactivities among 3492 approved molecules in ChEMBL-V33, offering a valuable tool for advancing drug mechanism discovery and repurposing existing drugs for new indications. https://github.com/AronSchulman/MMAtt-DTA.

Multi-filed data fusion through attention-based networks for readiness prediction in aircraft maintenance: natural language processing (NLP) approach

ABSTRACT Military aircraft data is analyzed from a readiness perspective to pursue sustainability. Aircraft readiness can be described as the percentage of fighting force available to perform a mission given a fixed period. It is critical to predict the readiness length of Non-Mission Capable (NMC) to prepare for alternative strategies to achieve mission success before a failure occurs. NMC also affects the maintenance process of an aircraft. In existing readiness state analysis, domain experts must manually assess significant amounts of data and identify the frequency and severity of failure modes, which is time-consuming, subjective, and merely descriptive analytics. This paper proposes a multi-filed data fusion framework through an attention-based network to predict aircraft mission capability. The model employs and investigates structured categorical information and manually inputs textual notes. The attention-based method is applied to retain and identify critical textual details, integrated with the dense representation of various categorical features. We demonstrate that the proposed model framework can contribute to capturing and analyzing essential features related to mission capability. The proposed method’s detailed performance is compared with existing approaches.

Multivariate Attention-Based Orbit Uncertainty Propagation and Orbit Determination Method for Earth–Jupiter Transfer

Current orbit uncertainty propagation (OUP) and orbit determination (OD) methods suffer from drawbacks related to high computational burden, limiting their applications in deep space missions. To this end, this paper proposes a multivariate attention-based method for efficient OUP and OD of Earth–Jupiter transfer. First, a neural network-based OD framework is utilized, in which the orbit propagation process in a traditional unscented transform (UT) and unscented Kalman filter (UKF) is replaced by the neural network. Then, the sample structure of training the neural network for the Earth–Jupiter transfer is discussed and designed. In addition, a method for efficiently generating a large number of samples for the Earth–Jupiter transfer is presented. Next, a multivariate attention-based neural network (MANN) is designed for orbit propagation, which shows better capacity in terms of accuracy and generalization than the deep neural network. Finally, the proposed method is successfully applied to solve the OD problem in an Earth–Jupiter transfer. Simulations show that the proposed method can obtain a similar estimation to the UKF while saving more than 90% of the computational cost.

ProDiv: Prototype-driven consistent pseudo-bag division for whole-slide image classification

Background and ObjectivePathology image classification is one of the most essential auxiliary processes in cancer diagnosis. To overcome the problem of inadequate Whole-Slide Image (WSI) samples with weak labels, pseudo-bag-based multiple instance learning (MIL) methods have attracted wide attention in pathology image classification. In this type of method, the division scheme of pseudo-bags is usually a primary factor affecting classification performance. In order to improve the division of WSI pseudo-bags on existing random/clustering approaches, this paper proposes a new Prototype-driven Division (ProDiv) scheme for the pseudo-bag-based MIL classification framework on pathology images. MethodsThis scheme first designs an attention-based method to generate a bag prototype for each slide. On this basis, it further groups WSI patch instances into a series of instance clusters according to the feature similarities between the prototype and patches. Finally, pseudo-bags are obtained by randomly combining the non-overlapping patch instances of different instance clusters. Moreover, the design scheme of our ProDiv considers practicality, and it could be smoothly assembled with almost all the MIL-based WSI classification methods in recent years. ResultsEmpirical results show that our ProDiv, when integrated with several existing methods, can deliver classification AUC improvements of up to 7.3% and 10.3%, respectively on two public WSI datasets. ConclusionsProDiv could almost always bring obvious performance improvements to compared MIL models on typical metrics, which suggests the effectiveness of our scheme. Experimental visualization also visually interprets the correctness of the proposed ProDiv.

Heterogeneous Graph Reasoning for Fact Checking over Texts and Tables

Fact checking aims to predict claim veracity by reasoning over multiple evidence pieces. It usually involves evidence retrieval and veracity reasoning. In this paper, we focus on the latter, reasoning over unstructured text and structured table information. Previous works have primarily relied on fine-tuning pretrained language models or training homogeneous-graph-based models. Despite their effectiveness, we argue that they fail to explore the rich semantic information underlying the evidence with different structures. To address this, we propose a novel word-level Heterogeneous-graph-based model for Fact Checking over unstructured and structured information, namely HeterFC. Our approach leverages a heterogeneous evidence graph, with words as nodes and thoughtfully designed edges representing different evidence properties. We perform information propagation via a relational graph neural network, facilitating interactions between claims and evidence. An attention-based method is utilized to integrate information, combined with a language model for generating predictions. We introduce a multitask loss function to account for potential inaccuracies in evidence retrieval. Comprehensive experiments on the large fact checking dataset FEVEROUS demonstrate the effectiveness of HeterFC. Code will be released at: https://github.com/Deno-V/HeterFC.

An Attention-Based Method for the Minimum Vertex Cover Problem on Complex Networks

Solving combinatorial problems on complex networks represents a primary issue which, on a large scale, requires the use of heuristics and approximate algorithms. Recently, neural methods have been proposed in this context to find feasible solutions for relevant computational problems over graphs. However, such methods have some drawbacks: (1) they use the same neural architecture for different combinatorial problems without introducing customizations that reflects the specificity of each problem; (2) they only use a nodes local information to compute the solution; (3) they do not take advantage of common heuristics or exact algorithms. Following this interest, in this research we address these three main points by designing a customized attention-based mechanism that uses both local and global information from the adjacency matrix to find approximate solutions for the Minimum Vertex Cover Problem. We evaluate our proposal with respect to a fast two-factor approximation algorithm and a widely adopted state-of-the-art heuristic both on synthetically generated instances and on benchmark graphs with different scales. Experimental results demonstrate that, on the one hand, the proposed methodology is able to outperform both the two-factor approximation algorithm and the heuristic on the test datasets, scaling even better than the heuristic with harder instances and, on the other hand, is able to provide a representation of the nodes which reflects the combinatorial structure of the problem.

GADNN: A graph attention-based method for drug-drug association prediction considering the contribution rate of different types of drug-related features

The simultaneous use of multiple drugs, known as drug combinations, is increasingly common in the treatment of complex diseases such as cancer. However, drug-drug interactions (DDIs) caused by the simultaneous use of multiple drugs can lead to unexpected side effects and even death, resulting in high medical costs. Therefore, understanding potential DDI is an essential step in reducing the risk of adverse drug reactions before administering clinical drug compounds. Recently, many machine learning-based methods for DDI prediction have been introduced, but in these methods, no attention is paid to determining the contribution rate of each type of drug feature. Here, a Graph Attention-based Deep Neural Network method called GADNN is proposed to predict DDIs based on considering the influence or contribution rate of different drug-related features. The introduced method consists of two stages. In the first stage, four basic drug datasets, including target, enzyme, infrastructure, and pathway, are used as types of drug features. Using a graph neural network-based method, the embedding vector is generated for each drug based on each dataset separately. In the second stage, the contribution coefficient of each dataset is calculated dynamically through a new graph attention mechanism. Then, the weighted combination of the separate vectors is used to predict the probability of drug-drug interactions using a dense neural network. The experimental results confirmed that the presented innovation has been able to improve the accuracy compared to state-of-the-art methods.

An efficient loss function and deep learning approach for ranking stock returns in the absence of prior knowledge

To pursue profit from the dynamic, complex, and noisy stock markets, various efforts utilizing deep learning methods to forecast asset price movements have sprung up. We observe that there are two issues in the current work. Firstly, there exists a discrepancy between the forecasting target and actual profitability, as achieving better forecasting results does not necessarily guarantee higher profits. Secondly, many existing methods heavily rely on prior knowledge during the forecasting process, which entails the need for information gathering and may not adapt well to dynamic and complex market conditions. For the first issue, we design a novel reward learning loss function as an optimization object to better model the bridge between the forecasting target and the profit from the trading process. To solve the second issue, we propose a structure based on multi-head attention to model the inter-stock relations directly from trading data without relying on prior knowledge. Also, we present a simple time-asynchronous attention-based method to model the lead–lag phenomenon in the market. We conduct experiments using over 600 stocks of the CSI100, CSI300, and CSI500 indexes from 2010 to 2020 with five strong baselines. The experimental results demonstrate that our methods achieve annualized returns of 5%, 10%, and 13% for long and 5%, 6%, and 8% for short above the optimal baseline results on the three indexes. Further analysis shows that our RL-Loss is better than classic PR-Loss, and the inter-stock relation modeling methods proposed without prior knowledge are effective.

DeepSA: a deep-learning driven predictor of compound synthesis accessibility

With the continuous development of artificial intelligence technology, more and more computational models for generating new molecules are being developed. However, we are often confronted with the question of whether these compounds are easy or difficult to synthesize, which refers to synthetic accessibility of compounds. In this study, a deep learning based computational model called DeepSA, was proposed to predict the synthesis accessibility of compounds, which provides a useful tool to choose molecules. DeepSA is a chemical language model that was developed by training on a dataset of 3,593,053 molecules using various natural language processing (NLP) algorithms, offering advantages over state-of-the-art methods and having a much higher area under the receiver operating characteristic curve (AUROC), i.e., 89.6%, in discriminating those molecules that are difficult to synthesize. This helps users select less expensive molecules for synthesis, reducing the time and cost required for drug discovery and development. Interestingly, a comparison of DeepSA with a Graph Attention-based method shows that using SMILES alone can also efficiently visualize and extract compound’s informative features. DeepSA is available online on the below web server (https://bailab.siais.shanghaitech.edu.cn/services/deepsa/) of our group, and the code is available at https://github.com/Shihang-Wang-58/DeepSA.

Automated Small Tumor Segmentation by a Template-Based Global Hierarchical Attention Method.

Accurate segmentation of tumors is significant for radiation therapy treatment planning and clinical decision-making. While deep convolutional neural network-based methods have found valuable applications in automatic medical segmentation, tumor segmentation, especially small tumor segmentation, remains challenging due to deficiencies of current deep learning in convolutional and pooling operations, which often results in the loss of small object information. This research proposes a global hierarchical attention-based method for accurate and automated segmentation of small tumors by exploiting the associations between small tumors and the feature maps of large tumors. This study included 131 patients with liver cancer. The in-plane resolution of the patients' CTs is from 0.55 mm to 1.0 mm and slice spacing from 0.45 mm to 6.0 mm. We randomly selected 100 CT scans as the training set and others as the testing set. Each CT slice of the testing set was separated into groups according to tumor size as follows: 0.1-2.0, 2.1-5.0, 5.1-10.0, and 10.1-20.0 cm. The CT slice without tumor or tumor size > 20 cm were excluded. This work presents a tumor template-based hierarchical attention method to quantify the relation between small and large tumors by computing their feature maps. The relation of small-large tumors can compensate for the information loss of small tumors during the convolutional and pooling operations and improve the performance of small tumor segmentation. Among 20,693 CT slices of the 31 testing patients, 3.0% CT slices with tumors ≤2 cm, 6.7% ≤5 cm, 10.6% ≤10 cm, and 13.4%≤20 cm. We compared our method with six widely used segmentation models. The results show our model outperforms other methods on all sizes of liver tumors, especially for small size tumors: For the 0.1-2.0 cm liver tumor, it achieved 8.4%, 10.0%, 11.3%, 9.1%, 10.9%, and 9.6% improvement compared to Unet, PAN, DeepLabV3, FPN, LinkNet, and PSPNet, respectively. We found that the small-large tumors relation can significantly improve small tumor segmentation, which is valuable for treatment planning, and clinical decision-making. Our experimental results show that our method can significantly improve the accuracy of segmenting small liver tumors compared to existing deep-learning-based models. The method is quite general and can be extended to other types of tumor detection and segmentation. We discovered that the relationship between small and large tumors can significantly enhance the segmentation of small tumors, which has significant value for treatment planning and clinical decision-making. Our experiments demonstrate that our approach significantly improves the accuracy of small liver tumor segmentation compared to existing deep learning-based models. Our method is quite versatile and can be extended to other types of tumor detection and segmentation.

A spectral-temporal constrained deep learning method for tree species mapping of plantation forests using time series Sentinel-2 imagery

Plantation forests provide critical ecosystem services and have experienced worldwide expansion during the past few decades. Accurate mapping of tree species through remote sensing is critical for managing plantation forests. The typical temporal behaviors and traits of tree species in satellite image time series (SITS) generate temporal and spectral features in multiple phenological stages that are critical to improve tree species mapping. However, the diverse input features, sequential relations and complex structures in SITS drastically increase the dimension and difficulty of spectral-temporal feature extraction, which challenges the capacity of many general classifiers not explicitly adapted for spectral-temporal learning. As a result, there is still a lack of a method that could automatically extract spectral-temporal features with high separability and regional adaptability from high-dimensional SITS for tree species mapping of plantation forests. Moreover, the effects of varying temporal resolution and feature combination on the plantation tree species mapping are under-explored. Here, we developed a multi-head attention-based method for automatically extracting spectral-temporal features with high separability based on a modified Transformer network (Transformer4SITS) for improved plantation tree species mapping. The end-to-end network model consists of a feature extraction module to learn deep spectral-temporal features from SITS and a fusion module to combine multiple features for improving mapping accuracy. We applied this method to two representative plantation forests in southern and northern China for tree species mapping. The results show that: (1) Transformer4SITS method could self-adaptively extract typical spectral-temporal features of key phenological stages (e.g., greenness rising and falling) from SITS, and achieved significantly improved accuracies by at most 15% in comparison with all four baseline methods (i.e., long short-term memory, harmonic analysis, time-weighted dynamic time warping, linear discriminant analysis); (2) time series with higher temporal resolution tended to produce more accurate species maps consistently across two sites, with their overall accuracies (OA) respectively increasing from 91.05% and 84.33% (60-day) to 94.88% and 88.72% (5-day), but the effect of high temporal resolution respectively leveled off around 90-day and 50-day resolution across two sites; (3) the mapping results using all available bands and two-band spectral indices outperformed the results using a subset of them, but with only modest increase in the accuracy (i.e., OA increased from 93.63% and 86.01% to 94.88% and 88.72%. This study thus provides a state-of-the-art deep learning-based method for improved tree species mapping, which is critical for sustainable management and biodiversity monitoring of plantation forests across large scales.

Cognitive Workload Estimation Using Variational Autoencoder and Attention-Based Deep Model

The estimation of cognitive workload using electroencephalogram (EEG) is an emerging research area. However, due to poor spatial resolution issues, features obtained from EEG signals often lead to poor classification results. As a good generative model, the Variational Autoencoder (VAE) extracts the noise-free robust features from the latent space that lead to better classification performance. The spatial attention-based method (Convolutional Block Attention Module: CBAM) can improve the spatial resolution of EEG signals. In this paper, we propose an effective VAE-CBAM-based deep model for estimating cognitive states from topographical videos. Topographical videos of four different conditions (Baseline: BL, low workload: LW, medium workload: MW, and high workload: HW) of the mental arithmetic task are taken for the experiment. Initially, the VAE extracts localized features from input images (extracted from topographical video), and CBAM infers the spatial-channellevels attention features from those localized features. Finally, the deep CNN-BLSTM model effectively learns those attentionbased spatial features in a timely-distributed manner to classify the cognitive state. For four-class and two-class classification, the proposed model achieves 83.13% and 92.09% classification accuracy, respectively. The proposed model enhances the future research scope of attention-based studies in EEG applications.

Small-Object Sensitive Segmentation Using Across Feature Map Attention.

Semantic segmentation is an important step in understanding the scene for many practical applications such as autonomous driving. Although Deep Convolutional Neural Networks-based methods have significantly improved segmentation accuracy, small/thin objects remain challenging to segment due to convolutional and pooling operations that result in information loss, especially for small objects. This article presents a novel attention-based method called Across Feature Map Attention (AFMA) to address this challenge. It quantifies the inner-relationship between small and large objects belonging to the same category by utilizing the different feature levels of the original image. The AFMA could compensate for the loss of high-level feature information of small objects and improve the small/thin object segmentation. Our method can be used as an efficient plug-in for a wide range of existing architectures and produces much more interpretable feature representation than former studies. Extensive experiments on eight widely used segmentation methods and other existing small-object segmentation models on CamVid and Cityscapes demonstrate that our method substantially and consistently improves the segmentation of small/thin objects.

On the Generalization of Deep Learning Models in Video Deepfake Detection.

The increasing use of deep learning techniques to manipulate images and videos, commonly referred to as "deepfakes", is making it more challenging to differentiate between real and fake content, while various deepfake detection systems have been developed, they often struggle to detect deepfakes in real-world situations. In particular, these methods are often unable to effectively distinguish images or videos when these are modified using novel techniques which have not been used in the training set. In this study, we carry out an analysis of different deep learning architectures in an attempt to understand which is more capable of better generalizing the concept of deepfake. According to our results, it appears that Convolutional Neural Networks (CNNs) seem to be more capable of storing specific anomalies and thus excel in cases of datasets with a limited number of elements and manipulation methodologies. The Vision Transformer, conversely, is more effective when trained with more varied datasets, achieving more outstanding generalization capabilities than the other methods analysed. Finally, the Swin Transformer appears to be a good alternative for using an attention-based method in a more limited data regime and performs very well in cross-dataset scenarios. All the analysed architectures seem to have a different way to look at deepfakes, but since in a real-world environment the generalization capability is essential, based on the experiments carried out, the attention-based architectures seem to provide superior performances.

Learn to Adapt to New Environments From Past Experience and Few Pilot Blocks

In recent years, deep learning has been widely applied in communications and achieved remarkable performance improvement. Most of the existing works are based on data-driven deep learning, which requires a significant amount of training data for the communication model to adapt to new environments and results in huge computing resources for collecting data and retraining the model. In this paper, we will significantly reduce the required amount of training data for new environments by leveraging the learning experience from the known environments. Therefore, we introduce few-shot learning to enable the communication model to generalize to new environments, which is realized by an attention-based method. With the attention network embedded into the deep learning-based communication model, environments with different power delay profiles can be learnt together in the training process, which is called the learning experience. By exploiting the learning experience, the communication model only requires few pilot blocks to perform well in the new environment. Through an example of deep-learning-based channel estimation, we demonstrate that this novel design method achieves better performance than the existing data-driven approach designed for few-shot learning.

A Hierarchical Attention-Based Method for Sleep Staging Using Movement and Cardiopulmonary Signals.

Sleep monitoring typically requires the uncomfortable and expensive polysomnography (PSG) test to determine the sleep stages. Body movement and cardiopulmonary signals provide an alternative way to perform sleep staging. In recent years, long-short term memory (LSTM) networks and convolutional neural networks (CNN) have dominated automatic sleep staging due to their better learning ability than machine learning classifiers. However, LSTM may lose information when dealing with long sequences, while CNN is not good at sequence modeling. As an improvement, we develop a hierarchical attention-based deep learning method for sleep staging using body movement, electrocardiogram (ECG), and abdominal breathing signals. We apply the multi-head self-attention to model the global context of feature sequences and coupled it with CNN to achieve a hierarchical self-attention weight assignment. We evaluate the performance of the method using two public datasets. Our method outperforms other baselines in the three sleep stages, achieving an accuracy of 84.3%, an F1 score of 0.8038, and a Cohen's Kappa coefficient of 0.7036. The result demonstrates the effectiveness of the hierarchical self-attention mechanism when processing feature sequences in the sleep stage classification problem. This paper provides new possibilities for long-term sleep monitoring using movement and cardiopulmonary signals obtained from non-invasive devices. The code can be found at: https://github.com/scutrd/attention-sleep-staging.

An Attention-Based Method for Remaining Useful Life Prediction of Rotating Machinery

Data imbalance and large data probability distribution discrepancies are major factors that reduce the accuracy of remaining useful life (RUL) prediction of high-reliability rotating machinery. In feature extraction, most deep transfer learning models consider the overall features but rarely attend to the local target features that are useful for RUL prediction; insufficient attention paid to local features reduces the accuracy and reliability of prediction. By considering the contribution of input data to the modeling output, a deep learning model that incorporates the attention mechanism in feature selection and extraction is proposed in our work; an unsupervised clustering method for classification of rotating machinery performance state evolution is put forward, and a similarity function is used to calculate the expected attention of input data to build an input data extraction attention module; the module is then fused with a gated recurrent unit (GRU), a variant of a recurrent neural network, to construct an attention-GRU model that combines prediction calculation and weight calculation for RUL prediction. Tests on public datasets show that the attention-GRU model outperforms traditional GRU and LSTM in RUL prediction, achieves less prediction error, and improves the performance and stability of the model.

A hierarchical multilabel graph attention network method to predict the deterioration paths of chronic hepatitis B patients.

Estimating the deterioration paths of chronic hepatitis B (CHB) patients is critical for physicians' decisions and patient management. A novel, hierarchical multilabel graph attention-based method aims to predict patient deterioration paths more effectively. Applied to a CHB patient data set, it offers strong predictive utilities and clinical value. The proposed method incorporates patients' responses to medications, diagnosis event sequences, and outcome dependencies to estimate deterioration paths. From the electronic health records maintained by a major healthcare organization in Taiwan, we collect clinical data about 177 959 patients diagnosed with hepatitis B virus infection. We use this sample to evaluate the proposed method's predictive efficacy relative to 9 existing methods, as measured by precision, recall, F-measure, and area under the curve (AUC). We use 20% of the sample as holdouts to test each method's prediction performance. The results indicate that our method consistently and significantly outperforms all benchmark methods. It attains the highest AUC, with a 4.8% improvement over the best-performing benchmark, as well as 20.9% and 11.4% improvements in precision and F-measures, respectively. The comparative results demonstrate that our method is more effective for predicting CHB patients' deterioration paths than existing predictive methods. The proposed method underscores the value of patient-medication interactions, temporal sequential patterns of distinct diagnosis, and patient outcome dependencies for capturing dynamics that underpin patient deterioration over time. Its efficacious estimates grant physicians a more holistic view of patient progressions and can enhance their clinical decision-making and patient management.

Temporal motif-based attentional graph convolutional network for dynamic link prediction

Dynamic link prediction is an important component of the dynamic network analysis with many real-world applications. Currently, most advancements focus on analyzing link-defined neighborhoods with graph convolutional networks (GCN), while ignoring the influence of higher-order structural and temporal interacting features on link formation. Therefore, based on recent progress in modeling temporal graphs, we propose a novel temporal motif-based attentional graph convolutional network model (TMAGCN) for dynamic link prediction. As dynamic graphs usually contain periodical patterns, we first propose a temporal motif matrix construction method to capture higher-order structural and temporal features, then introduce a spatial convolution operation following a temporal motif-attention mechanism to encode these features into node embeddings. Furthermore, we design two methods to combine multiple temporal motif-based attentions, a dynamic attention-based method and a reinforcement learning-based method, to allow each individual node to make the most of the relevant motif-based neighborhood to propagate and aggregate information in the graph convolutional layers. Experimental results on various real-world datasets demonstrate that the proposed model is superior to state-of-the-art baselines on the dynamic link prediction task. It also reveals that temporal motif can manifest the essential dynamic mechanism of the network.