Structural Information Research Articles

A Research Approach to Port Information Security Link Prediction Based on HWA Algorithm

For the protection of information security, link prediction, as a basic problem of network science, has important application significance. However, most of the existing link prediction algorithms rely on the node information of the graph structure, which is not applicable in some graph structure data involving privacy. At the same time, most of the algorithms only consider the general graph structure and do not fully consider the high-order information in the graph. Because of this, this paper proposes an algorithm called hypergraph-based link prediction with self-attention (HWA) to solve the above problems. The algorithm can obtain hypergraphs without knowing the attribute information of hypergraph nodes and combines the graph convolutional network (GCN) framework to capture node feature information for link prediction. Experiments show that the HWA algorithm proposed in this paper, combined with the GCN framework, shows better link prediction performance than other graph-based neural network benchmark algorithms on eight real networks. This further verifies the validity and reliability of the model in this paper and provides new protection ideas and technical means for information security.

Hypercube Pooling for Visual Semantic Embedding

Visual Semantic Embedding ( VSE ) is a primary model for cross-modal retrieval, wherein the global feature aggregator is a crucial component of the VSE model. In recent research, the General Pooling Operator ( GPO ) aggregator, which weighs the features reconstructed from the local feature set to aggregate, facilitates the related models to achieve good retrieval performance. However, the reason for the effectiveness remains to be explored. To enhance the rationality of aggregator designs, we analyze the reason from the perspective of feature space. Indeed, for each data, the local feature set forms a hypercube containing abundant data information, and the feature learned by GPO measures the hypercube, thereby representing the data. The geometric structure of the hypercube implies that the set containing all points within the hypercube is a convex set, so the feature learned by weighted aggregation is an interior point of the hypercube. However, using the interior point to measure the hypercube leads to some problems in feature representation and model optimization, as well as the reduction of retrieval efficiency caused by weight computation. For example, the related pair’s features may be far, while the unrelated ones may be close. To measure the hypercube more clearly and alleviate the problems mentioned above, we propose Hypercube Pooling ( HCP ) aggregator. Specifically, HCP concatenates the Max and Min Pooling features as the global features. This aggregation method has multiple advantages, e.g., the learned global feature represents all hyperplanes of the hypercube that contain critical information and hypercube geometric structure. Moreover, HCP adds normalization-before-concatenation and reduces the usual setting of margin in the loss function by half to avoid gradient loss caused by the difference in the feature value and dimensionality doubling. The experimental results on the Flickr30K and MSCOCO datasets show that the HCP model has excellent performance with high efficiency, confirming the correctness of the spatial analysis and the effectiveness of the HCP aggregator.

Deep reinforcement learning-based spatio-temporal graph neural network for solving job shop scheduling problem

AbstractThe job shop scheduling problem (JSSP) is a well-known NP-hard combinatorial optimization problem that focuses on assigning tasks to limited resources while adhering to certain constraints. Currently, deep reinforcement learning (DRL)-based solutions are being widely used to solve the JSSP by defining the problem structure on disjunctive graphs. Some of the proposed approaches attempt to leverage the structural information of the JSSP to capture the dynamics of the environment without considering the time dependency within the JSSP. However, learning graph representations only from the structural relationship of nodes results in a weak and incomplete representation of these graphs which does not provide an expressive representation of the dynamics in the environment. In this study, unlike existing frameworks, we defined the JSSP as a dynamic graph to explicitly consider the time-varying aspect of the JSSP environment. To this end, we propose a novel DRL framework that captures both the spatial and temporal attributes of the JSSP to construct rich and complete graph representations. Our DRL framework introduces a novel attentive graph isomorphism network (Attentive-GIN)-based spatial block to learn the structural relationship and a temporal block to capture the time dependency. Additionally, we designed a gated fusion block that selectively combines the learned representations from the two blocks. We trained the model using the proximal policy optimization algorithm of reinforcement learning. Experimental results show that our trained model exhibits significant performance enhancement compared to heuristic dispatching rules and learning-based solutions for both randomly generated datasets and public benchmarks.

Markerless tracking of tumor and tissues: A motion model approach.

Respiratory motion management is essential in order to achieve high-precision radiotherapy. Markerless motion tracking of tumor can provide a non-invasive way to manage respiratory motion, thereby enhancing treatment accuracy. However, the low contrast in real-time x-ray images for image guidance limits the application of markerless tracking. We present a novel approach based on a motion model to perform markerless tracking of tumor and surrounding tissues even when they have low contrast in real-time x-ray images. A proof-of-concept validation of the method has been performed using digital and physical phantoms at breathing conditions that are significantly different than the planning stage. A motion model is first constructed by performing principal component analysis (PCA) on the planning 4DCT. During treatment, the motion of a surrogate is tracked and used as the input of the motion model, which generates a 3D real-time volume estimation. Such 3D estimation is then projected to 2D to create digitally reconstructed radiographs (DRRs). The relationships between the real-time DRRs, reference DRRs, and reference x-ray images are first established to simulate 2D real-time images from the real-time volume. The registration between the simulated 2D real-time images and real-time x-ray images corrects the initial motion model estimation to ensure the estimated volume matches the real-time condition. In digital phantom, the Dice index of pancreas was improved from 0.74 to 0.78 after correction using real-time DRRs in fully inhaled phase. Validation on lung and pancreas is performed in physical phantom with two motion traces. The surrogate-tumor relationships were intentionally altered to generate large target localization errors due to the differences in body condition between treatment planning stage and during treatment. The real-time correction for the estimated 3D real-time volume was performed using a pair of 2D x-ray images. For the deep breathing motion trace, the tumor localization mean absolute error (MAE) throughout the tracking decreases from around 3mm to less than 1mm after correction. For the shallow breathing motion trace with a 1.7mm baseline shift, the tumor localization MAE throughout the tracking decreases from around 1.5mm to less than 1mm after correction. The method combines the detailed structural information from planning 4DCT and real-time information from real-time x-ray images through a motion model. The matching between the real-time model estimation and 2D real-time images is performed in the same modality so that it can be applied to regions with low contrast in the images. The real-time images successfully corrected the initial motion model estimations in our proof-of-concept validation. This suggests the potential to perform markerless tracking in low-contrast region using a motion model.

Structure of scavenger receptor SCARF1 and its interaction with lipoproteins.

SCARF1 (scavenger receptor class F member 1, SREC-1 or SR-F1) is a type I transmembrane protein that recognizes multiple endogenous and exogenous ligands such as modified low-density lipoproteins (LDLs) and is important for maintaining homeostasis and immunity. But the structural information and the mechanisms of ligand recognition of SCARF1 are largely unavailable. Here, we solve the crystal structures of the N-terminal fragments of human SCARF1, which show that SCARF1 forms homodimers and its epidermal growth factor (EGF)-like domains adopt a long-curved conformation. Then, we examine the interactions of SCARF1 with lipoproteins and are able to identify a region on SCARF1 for recognizing modified LDLs. The mutagenesis data show that the positively charged residues in the region are crucial for the interaction of SCARF1 with modified LDLs, which is confirmed by making chimeric molecules of SCARF1 and SCARF2. In addition, teichoic acids, a cell wall polymer expressed on the surface of gram-positive bacteria, are able to inhibit the interactions of modified LDLs with SCARF1, suggesting the ligand binding sites of SCARF1 might be shared for some of its scavenging targets. Overall, these results provide mechanistic insights into SCARF1 and its interactions with the ligands, which are important for understanding its physiological roles in homeostasis and the related diseases.

Exploring Structural Anisotropy in Amorphous Tb-Co via Changes in Medium-Range Ordering.

Amorphous thin films grown by magnetron co-sputtering exhibit changes in atomic structure with varying growth and annealing temperatures. Structural variations influence the bulk properties of the films. Scanning nanodiffraction performed in a transmission electron microscope (TEM) is applied to amorphous Tb17Co83 (a-Tb-Co) films deposited over a range of temperatures to measure relative changes in medium-range ordering (MRO). These measurements reveal an increase in MRO with higher growth temperatures and a decrease in MRO with higher annealing temperatures. The trend in MRO indicates a relationship between the growth conditions and local atomic ordering. By tilting select films, the TEM measures variations in the local atomic structure as a function of orientation within the films. The findings support claims that preferential ordering along the growth direction results from temperature-mediated adatom configurations during deposition, and that oriented MRO correlates with increased structural anisotropy, explaining the strong growth-induced perpendicular magnetic anisotropy found in rare earth-transition metal films. Beyond magnetic films, we propose the tilted FEM workflow as a method of extracting anisotropic structural information in a variety of amorphous materials with directionally dependent bulk properties, such as films with inherent bonding asymmetry grown by physical vapor deposition.

Why is there no negativity bias in evaluative conditioning? A cognitive-ecological answer.

Evaluative conditioning (EC) is the change of a conditioned stimulus's evaluation due to its pairing with an unconditioned stimulus (US). While learning typically shows negativity biases, we found no such biases in a reanalysis of meta-analytic EC data. We provide and test a cognitive-ecological answer for this lack of negativity bias. We assume that negativity effects follow from ecological differences in evaluative information's distributions (i.e., differential frequency). Accordingly, no negativity bias emerges because positive and negative information is equally frequent in most EC experiments. However, if negative (or positive) information is rare, we predict a negativity (positivity) bias. We tested this prediction in five preregistered experiments (three laboratory-based, N = 394, two online, N = 391). As predicted, if negative USs were rare, a negativity bias followed. However, if positive USs were rare, we also observed positivity biases in participants' conditioned stimulus evaluations. These data support a cognitive-ecological explanation of valence asymmetries and partially explain why EC experiments show no negativity bias: Typical EC designs do not reflect the ecological information structure that contributes to a negativity bias in the first place. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Building Change Detection Network Based on Multilevel Geometric Representation Optimization Using Frame Fields

To address the challenges of accurately segmenting irregular building boundaries in complex urban environments faced by existing remote sensing change detection methods, this paper proposes a building change detection network based on multilevel geometric representation optimization using frame fields called BuildingCDNet. The proposed method employs a multi-scale feature aggregation encoder–decoder architecture, leveraging contextual information to capture the characteristics of buildings of varying sizes in the imagery. Cross-attention mechanisms are incorporated to enhance the feature correlations between the change pairs. Additionally, the frame field is introduced into the network to model the complex geometric structure of the building target. By learning the local orientation information of the building structure, the frame field can effectively capture the geometric features of complex building features. During the training process, a multi-task learning strategy is used to align the predicted frame field with the real building outline, while learning the overall segmentation, edge outline, and corner point features of the building. This improves the accuracy of the building polygon representation. Furthermore, a discriminative loss function is constructed through multi-task learning to optimize the polygonal structured information of the building targets. The proposed method achieves state-of-the-art results on two commonly used datasets.

Developing Hydration Maps of Polymer Latex Film Formation Using Terahertz Time-Domain Spectroscopy.

The dynamics of the drying process of polymer latexes after casting as a wet film onto a substrate are important to track as they influence the physical and mechanical properties and performance of the dried polymer films. Current methods used to follow this drying process include gravimetric analysis, coupled with advanced techniques like GARField-NMR or optical coherence tomography. The latter two methods provide height and spatial information in the z-direction, normal to the substrate, and occasionally in the xz- or yz-planes. Terahertz time-domain spectroscopy (THz-TDS) is a welcome addition as it provides both the structural and spectroscopic information in the parallel xy-plane, filling the geometric gap. Herein, we utilize THz-TDS to study the drying and film formation process of various polymer latexes with a broad range of glass transition temperatures. We showcase the applicability of this technique in obtaining 2D parallel hydration maps of the drying dispersions, in the form of droplets, using latex-dependent calibration lines. Our findings display known phenomena arising from the drying of the colloidal dispersions.

Image enhancement with art design: a visual feature approach with a CNN-transformer fusion model

Graphic design, as a product of the burgeoning new media era, has seen its users’ requirements for images continuously evolve. However, external factors such as light and noise often cause graphic design images to become distorted during acquisition. To enhance the definition of these images, this paper introduces a novel image enhancement model based on visual features. Initially, a histogram equalization (HE) algorithm is applied to enhance the graphic design images. Subsequently, image feature extraction is performed using a dual-flow network comprising convolutional neural network (CNN) and Transformer architectures. The CNN employs a residual dense block (RDB) to embed spatial local structure information with varying receptive fields. An improved attention mechanism module, attention feature fusion (AFF), is then introduced to integrate the image features extracted from the dual-flow network. Finally, through image perception quality guided adversarial learning, the model adjusts the initial enhanced image’s color and recovers more details. Experimental results demonstrate that the proposed algorithm model achieves enhancement effects exceeding 90% on two large image datasets, which represents a 5%–10% improvement over other models. Furthermore, the algorithm exhibits superior performance in terms of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) image quality evaluation metrics. Our findings indicate that the fusion model significantly enhances image quality, thereby advancing the field of graphic design and showcasing its potential in cultural and creative product design.

Target Structure Learning Framework for Unsupervised Multi-Class Domain Adaptation

Unsupervised multi-class domain adaptation (multi-class UDA) has recently been proposed to fill the gap between empirically practical methods for multi-class classification and well-founded theory with the setting of binary classification. Nevertheless, the multi-class UDA methods use model predictions to characterize the disagreement of multi-class scoring hypotheses, which is used to optimize the divergence between domain distributions. Such self-training manner may bring inaccurate model predictions, which would damage the target structure due to the absence of labels of the target domain, leading to sub-optimal performance. On the other hand, this disagreement between multi-class scoring hypotheses does not involve the relationships among all of the multiple classes. It causes that multi-class UDA cannot properly connect the advanced practical UDA methods that consider class-conditional distribution alignment. Thus, we propose to exploit the target structure information and then incorporate it into multi-class UDA to achieve class-conditional distribution alignment. We theoretically and experimentally explain the importance of accurate target structure information to reduce the expected error on the target domain. Notably, our method achieves state-of-the-art results on three commonly-used benchmarks with different scales. In addition, using the target structure information, we propose a variant to cope with noisy open-world source domains such as noisy labels and out-of-distribution samples, enhancing the robustness of our method. The source code is available at https://github.com/jingzhengli/Multi_Class_UDA .

DGTNet：dynamic graph attention transformer network for traffic flow forecasting

Abstract Graph-based traffic flow prediction plays a crucial role in urban traffic management and planning. In this paper, we propose a novel Dynamic Graph Attention Transformer Network (DGTnet), which is designed to address the issue of inadequately integrating of temporal and spatial dimensions in traditional models. DGTnet maintains temporal continuity while revealing the complex dynamic relationships between key nodes in the urban traffic system, capturing the periodic changes in the rhythm of city life. Specifically, this study adopts adaptive signal decomposition technology to decompose traffic data into multiple intrinsic mode functions (IMFs), effectively capturing the dynamic changes in traffic flow. This decomposition method is key to the implementation of DGTnet dynamic graph construction, enabling the analysis of traffic flow at different time scales, thereby providing a new perspective for traffic flow prediction research. In the traffic prediction module, we comprehensively consider node, edge, and graph structural information, adopting a multi-head self-attention mechanism to achieve direct cross-modeling in both temporal and spatial dimensions. Finally, we introduce a position-wise feed forward network layer to integrate different types of data and capture nonlinear relationships. The experimental results, based on the public transportation network data sets METR_LA, PEMS_BAY, PEMS03 and PEMS07, demonstrate that DGTNet exhibits notable enhancements in three evaluation indicators, namely the Mean Absolute Percentage Error (MAPE), the Root Mean Square Error (RMSE) and the Mean Absolute Error (MAE). The pertinent code has been made available for public access at https://github.com/chenjing0616/DGTNet.

Examining the persuasiveness of text and voice agents: prosody aligned with information structure increases human-likeness, perceived personalisation and brand attitude

ABSTRACT To give product and brand recommendations, marketers make use of conversational agents which increasingly communicate via voice rather than text. Existing research comparing the persuasiveness of text and voice agents showed mixed results. The quality of the speech synthesis employed may strongly influence consumers’ responses. This study investigates to what extent a voice agent with pragmatically aligned prosody is more persuasive (i.e. yields a more positive brand attitude) than an agent with a standard voice or text, and whether perceived human-likeness and perceived personalisation provide an underlying mechanism to explain these differences. In an experiment (n = 212), participants interacted with a conversational agent that recommended a camera. Results showed that a voice agent using prosody aligned to the information state of the user is more persuasive than a text agent. This effect is mediated by perceived human-likeness and perceived personalisation. Hence, aligned prosody can make synthetic speech meet a certain quality threshold to be perceived as more human-like. Theoretically, this study helps to unravel why conversational agents with human-like features are more persuasive.

Graph neural link predictor based on cycle structure

AbstractCurrently, the link prediction algorithms primarily focus on studying the interaction between nodes based on chain structure and star structure, which predominantly rely on low‐order structural information and do not explore the multivariate interactions between nodes from the perspective of higher‐order structural information present in the network. The cycle structure is a higher‐order structure that lies between the star and clique structures, where all nodes within the same cycle can interact with each other, even in the absence of direct edges. If a node is encompassed by multiple cycles, it indicates that the node interacts and associates with a greater number of nodes in the network, and it means the node is more important in the network to some extent. Furthermore, if two nodes are included in multiple cycles, it signifies the two nodes are more likely to be connected. Therefore, firstly, a multi‐information fusion node importance algorithm based on the cycle structure information is proposed, which integrates both high‐order and low‐order structural information. Secondly, the obtained integrated structure information and node feature information is regarded as the input features, a two‐channel graph neural network model is designed to learn the cycle structure information. Then, the cycle structure information is utilised for the task of link prediction, and a graph neural link predictor with multi‐information interactions based on the cycle structure is developed. Finally, extensive experimental validation and analysis show that the node ranking result of the proposed node importance index is more consistent with the actual situation, the proposed graph neural network model can effectively learn the cycle structure information, and using higher‐order structural information—cycle information proves to significantly enhance the overall link prediction performance.

Abstract 4138722: Large Language Model based multi-agent Transcatheter Aortic Valve Implantation team to augment multidisciplinary meetings - proof of concept.

Introduction: Multidisciplinary team (MDT) discussions are integral to Transcatheter Aortic Valve Implantation (TAVI) decision making. Large language model (LLM) ubiquity and low-code no-code platforms have enabled clinician lead solution development. Specialised chatbots or ‘agents’ have evolved into multi-agent systems that can personify human collaboration. We assess the performance of an artificial intelligence (AI) multi-agent TAVI MDT. Methods: Four de-identified TAVI cases from two metropolitan Australian hospitals were assessed by a mock human TAVI MDT (h-MDT) and an AI multi-agent TAVI MDT (ai-MDT). The ai-MDT was created with Agentflow within Flowise AITM and had a hierarchical multi-agent architecture suited to complex reasoning required for TAVI MDT simulation (figure). LLM limitations necessitated the ai-MDT rely on imaging reports rather than clinical images. The h-MDT and ai-MDT consisted of similar team members. Outputs from the h-MDT and ai-MDT was adjudicated by a panel of four blinded TAVI doctors that determined if output was human vs AI and assigned a SMIC score (4-12, 4=good, 12=poor) that assessed structure, missing information, incorrect information and clinical utility. Time durations for h-MDT and ai-MDT were recorded. Results: Adjudicators differentiated human vs. AI output 100% of the time and ai-MDT output had better SMIC scores than h-MDT (U-stat 213, p=0.0011). ai-MDT outperformed h-MDT in the domains of structure, missing information and clinical utility but was not statistically different in the incorrect information domain (U-stat 132, p=0.88). The average time for each case in h-MDT was 15 minutes and 45 seconds compared to 97 seconds for ai-MDT. Conclusion: This demonstrates the potential of using LLM based multi-agent systems as a clinical adjunct in highly specialized multidisciplinary clinical meetings. AI responses were superior for structure, clinical utility and missing information and non-inferior for incorrect information compared to humans, which highlights that hallucinations remain an issue with generative AI. Time was saved but image interpretation still requires human input, for now. Cognitive AI continues to require human supervision for implementation.

EvoAI enables extreme compression and reconstruction of the protein sequence space.

Designing proteins with improved functions requires a deep understanding of how sequence and function are related, a vast space that is hard to explore. The ability to efficiently compress this space by identifying functionally important features is extremely valuable. Here we establish a method called EvoScan to comprehensively segment and scan the high-fitness sequence space to obtain anchor points that capture its essential features, especially in high dimensions. Our approach is compatible with any biomolecular function that can be coupled to a transcriptional output. We then develop deep learning and large language models to accurately reconstruct the space from these anchors, allowing computational prediction of novel, highly fit sequences without prior homology-derived or structural information. We apply this hybrid experimental-computational method, which we call EvoAI, to a repressor protein and find that only 82 anchors are sufficient to compress the high-fitness sequence space with a compression ratio of 1048. The extreme compressibility of the space informs both applied biomolecular design and understanding of natural evolution.

An Efficient Framework for Finding Similar Datasets Based on Ontology

Governments are embracing an open data philosophy and making their data freely available to the public to encourage innovation and increase transparency. However, the number of available datasets is still limited. Finding relationships between related datasets on different data portals enables users to search the relevant datasets. These datasets are generated from the training data, which need to be curated by the user query. However, relevant dataset retrieval is an expensive operation due to the preparation procedure for each dataset. Moreover, it requires a significant amount of space and time. In this study, we propose a novel framework to identify the relationships between datasets using structural information and semantic information for finding similar datasets. We propose an algorithm to generate the Concept Matrix (CM) and the Dataset Matrix (DM) from the concepts and the datasets, which is then used to curate semantically related datasets in response to the users’ submitted queries. Moreover, we employ the proposed compression, indexing, and caching algorithms in our proposed scheme to reduce the required storage and time while searching the related ranked list of the datasets. Through extensive evaluation, we conclude that the proposed scheme outperforms the existing schemes.

Tactile-GAT: tactile graph attention networks for robot tactile perception classification.

As one of the most important senses in human beings, touch can also help robots better perceive and adapt to complex environmental information, improving their autonomous decision-making and execution capabilities. Compared to other perception methods, tactile perception needs to handle multi-channel tactile signals simultaneously, such as pressure, bending, temperature, and humidity. However, directly transferring deep learning algorithms that work well on temporal signals to tactile signal tasks does not effectively utilize the physical spatial connectivity information of tactile sensors. In this paper, we propose a tactile perception framework based on graph attention networks, which incorporates explicit and latent relation graphs. This framework can effectively utilize the structural information between different tactile signal channels. We constructed a tactile glove and collected a dataset of pressure and bending tactile signals during grasping and holding objects, and our method achieved 89.58% accuracy in object tactile signal classification. Compared to existing time-series signal classification algorithms, our graph-based tactile perception algorithm can better utilize and learn sensor spatial information, making it more suitable for processing multi-channel tactile data. Our method can serve as a general strategy to improve a robot's tactile perception capabilities.

Gas adsorption meets geometric deep learning: points, set and match

Thanks to their unique properties such as ultra high porosity and surface area, metal-organic frameworks (MOFs) are highly regarded materials for gas adsorption applications. However, their combinatorial nature results in a vast chemical space, precluding its exploration with traditional techniques. Recently, machine learning (ML) pipelines have been established as the go-to method for large scale screening by means of predictive models. These are typically built in a descriptor-based manner, meaning that the structure must be first coarse-grained into a 1D fingerprint before it is fed to the ML algorithm. As such, the latter can not fully exploit the 3D structural information, potentially resulting in a model of lower quality. In this work, we propose a descriptor-free framework called “AIdsorb”, which can directly process raw structural information for predicting gas adsorption properties. To accomplish that, the structure is first treated as a point cloud and then passed to a deep learning algorithm suitable for point cloud analysis. As a proof of concept, AIdsorb is applied for predicting CO2\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ ext {CO}_{2}$$\\end{document} uptake in MOFs, outperforming a conventional pipeline that uses geometric descriptors as input. Additionally, to evaluate the transferability of the proposed framework to different host-guest systems, CH4\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ ext {CH}_{4}$$\\end{document} uptake in COFs is examined. Since AIdsorb bases its roots on raw structural information, its applicability extends to all fields of material science.

Travel route recommendation with a trajectory learning model

This study addresses a critical issue in location-based services: travel route recommendation. It leverages historical trajectory data to predict the actual route on a road network from a starting point to a destination, given a specific departure time. However, capturing the latent patterns in complex trajectory data for accurate route planning presents a significant challenge. Existing route recommendation methods commonly face two major problems: first, inadequate integration of multi-source data, which fails to fully consider the potential factors affecting route choice; and second, limited capability to capture road network characteristics, which restricts the effective application of node features and negatively impacts recommendation accuracy. To address these issues, this research introduces a Trajectory Learning Model for Route Recommendation (TLMR) based on deep learning techniques. TLMR enhances the understanding of user route choice behavior in complex environments by integrating multi-source data. Moreover, by incorporating road network features, TLMR more effectively captures and utilizes the structural and dynamic information of the road network. Specifically, TLMR first employs a Position-aware Graph Neural Network to learn features of intersections from the road network, incorporating context features like weather and traffic conditions. Then, it integrates this information through neural networks to predict the next intersection. Finally, a beam search algorithm is applied to generate and recommend multiple candidate routes. Extensive experiments on four large real-world datasets demonstrate that TLMR outperforms existing methods in four key performance metrics. These results prove the effectiveness and superiority of TLMR in route recommendation.