As digital twins (DTs) to physical communication systems, network simulators can aid the design and deployment of communication networks. However, time-consuming simulations must be run for every new set of network configurations. Learnable digital twins (LDTs), in contrast, can be trained offline to emulate simulation outcomes and serve as a more efficient alternative to simulation-based DTs at runtime. In this work, we propose GLANCE, a communication LDT that learns from the simulator ns-3. It can evaluate network key performance indicators (KPIs) and assist in network management with exceptional efficiency. Leveraging graph learning, we exploit network data characteristics and devise a specialized architecture to embed sequential and topological features of traffic flows within the network. In addition, multi-task learning (MTL) and transfer learning (TL) are leveraged to enhance GLANCE's generalizability to unseen inputs and efficacy across different tasks. Beyond end-to-end KPI prediction, GLANCE can be deployed within an optimization framework for network management. It serves as an efficient or differentiable evaluator in optimizing network configurations such as traffic loads and flow destinations. Through numerical experiments and benchmarking, we verify the effectiveness of the proposed LDT architecture, demonstrate its robust generalization to various inputs, and showcase its efficacy in network management applications.

We propose a novel data-driven approach to allocate transmit power for federated learning (FL) over interference-limited wireless networks. The proposed method is useful in challenging scenarios where the wireless channel is changing during the FL training process and when the training data are not independent and identically distributed (non-i.i.d.) on the local devices. Intuitively, the power policy is designed to optimize the information received at the server end during the FL process under communication constraints. Ultimately, our goal is to improve the accuracy and efficiency of the global FL model being trained. The proposed power allocation policy is parameterized using graph convolutional networks (GCNs), and the associated constrained optimization problem is solved through a primal-dual (PD) algorithm. Theoretically, we show that the formulated problem has a zero duality gap and, once the power policy is parameterized, optimality depends on how expressive this parameterization is. Numerically, we demonstrate that the proposed method outperforms existing baselines under different wireless channel settings and varying degrees of data heterogeneity.

Network digital twins (NDTs) facilitate the estimation of key performance indicators (KPIs) before physically implementing a network, thereby enabling efficient optimization of the network configuration. In this paper, we propose a learning-based NDT for network simulators. The proposed method offers a holistic representation of information flow in a wireless network by integrating node, edge, and path embeddings. Through this approach, the model is trained to map the network configuration to KPIs in a single forward pass. Hence, it offers a more efficient alternative to traditional simulation-based methods, thus allowing for rapid experimentation and optimization. Our proposed method has been extensively tested through comprehensive experimentation in various scenarios, including wired and wireless networks. Results show that it outperforms baseline learning models in terms of accuracy and robustness. Moreover, our approach achieves comparable performance to simulators but with significantly higher computational efficiency.

We develop a novel graph-based trainable framework to maximize the weighted sum energy efficiency (WSEE) for power allocation in wireless communication networks. To address the non-convex nature of the problem, the proposed method consists of modular structures inspired by a classical iterative suboptimal approach and enhanced with learnable components. More precisely, we propose a deep unfolding of the successive concave approximation (SCA) method. In our unfolded SCA (USCA) framework, the originally preset parameters are now learnable via graph convolutional neural networks (GCNs) that directly exploit multi-user channel state information as the underlying graph adjacency matrix. We show the permutation equivariance of the proposed architecture, which is a desirable property for models applied to wireless network data. The USCA framework is trained through a stochastic gradient descent approach using a progressive training strategy. The unsupervised loss is carefully devised to feature the monotonic property of the objective under maximum power constraints. Comprehensive numerical results demonstrate its generalizability across different network topologies of varying size, density, and channel distribution. Thorough comparisons illustrate the improved performance and robustness of USCA over state-of-the-art benchmarks.

To address the visual detection and positioning challenge of the Automated Guided Vehicle (AGV) terminal, this study proposed a high-precision target recognition and positioning strategy based on cross-layer feature fusion and eccentricity error correction. In the remote coarse positioning stage, small number of receiver domains underwent continuous multi-layer convolution, which broadened the receptive field of a single element in the feature map of a small circular target. Next, a cross-layer connection feature pyramid was constructed, and the deconvolution module was utilized to enlarge the feature map and fuse it with the shallow features, focusing more on the fine-grained image recognition and enhancing the applicability of circular marker detection. In the near-end fine positioning stage, the factors impacting the deviation between the projection point of the circular feature center and the fitting center were analyzed. Based on this analysis, the deviation of the iterative fitting center was corrected to approximate the real center projection sub-pixel coordinates. The experimental results demonstrated that the proposed method achieved dynamic positioning accuracy of better than 2.0 mm, static positioning accuracy of better than 1.5 mm, and a false recognition rate of less than 1.33% in the range of 3 m from AGV to the tray. Consequently, this method has significant application potential in enabling rapid and stable terminal vision positioning.

We propose a data-driven approach for power allocation in the context of federated learning (FL) over interference-limited wireless networks. The power policy is designed to maximize the transmitted information during the FL process under communication constraints, with the ultimate objective of improving the accuracy and efficiency of the global FL model being trained. The proposed power allocation policy is parameterized using a graph convolutional network and the associated constrained optimization problem is solved through a primal-dual algorithm. Numerical experiments show that the proposed method outperforms three baseline methods in both transmission success rate and FL global performance.

We develop a spectral clustering algorithm for general submodular hypergraphs based on their 1-Laplacians. More precisely, we utilize the eigenvector associated with the second smallest eigenvalue of the hypergraph 1-Laplacian to cluster the vertices. The computation of the eigenvector is based on the inverse power method, the key of which is to solve its inner-loop optimization problem. Efficient solutions to this inner problem exist when submodular hypergraphs are equipped with hyperedge splitting functions with special structures such as when they are cardinality based or graph reducible. In this paper, we present a solution to the inner problem for general submodular splitting functions by adopting a random coordinate descent method together with the Fujishige-Wolfe algorithm. Numerical experiments using real-world data demonstrate the effectiveness of the proposed clustering algorithm.

We propose a flexible framework for defining the 1-Laplacian of a hypergraph that incorporates edge-dependent vertex weights. These weights are able to reflect varying importance of vertices within a hyperedge, thus conferring the hypergraph model higher expressivity than homogeneous hypergraphs. We then utilize the eigenvector associated with the second smallest eigenvalue of the hypergraph 1-Laplacian to cluster the vertices. From a theoretical standpoint based on an adequately defined normalized Cheeger cut, this procedure is expected to achieve higher clustering accuracy than that based on the traditional Laplacian. Indeed, we confirm that this is the case using real-world datasets to demonstrate the effectiveness of the proposed spectral clustering approach. Moreover, we show that for a special case within our framework, the corresponding hypergraph 1-Laplacian is equivalent to the 1-Laplacian of a related graph, whose eigenvectors can be computed more efficiently, facilitating the adoption on larger datasets.

We propose a flexible framework for defining the 1-Laplacian of a hypergraph that incorporates edge-dependent vertex weights. These weights are able to reflect varying importance of vertices within a hyperedge, thus conferring the hypergraph model higher expressivity than homogeneous hypergraphs. We then utilize the eigenvector associated with the second smallest eigenvalue of the hypergraph 1-Laplacian to cluster the vertices. From a theoretical standpoint based on an adequately defined normalized Cheeger cut, this procedure is expected to achieve higher clustering accuracy than that based on the traditional Laplacian. Indeed, we confirm that this is the case using real-world datasets to demonstrate the effectiveness of the proposed spectral clustering approach. Moreover, we show that for a special case within our framework, the corresponding hypergraph 1-Laplacian is equivalent to the 1-Laplacian of a related graph, whose eigenvectors can be computed more efficiently, facilitating the adoption on larger datasets.

Effective integration planning for small, distributed solar photovoltaic (PV) arrays into electric power grids requires access to high quality data: the location and power capacity of individual solar PV arrays. Unfortunately, national databases of small-scale solar PV do not exist; those that do are limited in their spatial resolution, typically aggregated up to state or national levels. While several promising approaches for solar PV detection have been published, strategies for evaluating the performance of these models are often highly heterogeneous from study to study. The resulting comparison of these methods for practical applications for energy assessments becomes challenging and may imply that the reported performance evaluations overly optimistic. The heterogeneity comes in many forms, each of which we explore in this work: the degree of diversity of the locations and sensors (e.g. different satellites, aerial photography) from which the training and validation data originate, the validation of ground truth (manual annotation of imagery vs known solar PV locations), the level of spatial aggregation (e.g. array-level vs regional estimates), and inconsistencies in the training and validation datasets (e. g. different datasets are used for each study and those data are not always made accessible). For each, we discuss emerging practices from the literature to address them or suggest directions of future research. As part of our investigation, we evaluate solar PV identification performance in two large regions: the entire state of Connecticut and the city of San Diego, CA. In Connecticut, we also use 33,114 known parcel-level solar PV installations from Berkeley Lab's Tracking the Sun dataset to evaluate parcel-level performance and evaluate capacity estimates using 169 municipalities. We also make our code (which we call SolarMapper), pre-trained models, training data, and predictions publicly available and provide a web portal for interactively inspecting each prediction that was made. Our findings suggest that traditional performance evaluation of the automated identification of solar PV from satellite imagery may be optimistic due to common limitations in the validation process. The takeaways from this work are intended to inform and catalyze the large-scale practical application of automated solar PV assessment techniques by energy researchers and professionals.

We propose a novel method to co-cluster the vertices and hyperedges of hypergraphs with edge-dependent vertex weights (EDVWs). In this hypergraph model, the contribution of every vertex to each of its incident hyperedges is represented through an edge-dependent weight, conferring the model higher expressivity than the classical hypergraph. In our method, we leverage random walks with EDVWs to construct a hypergraph Laplacian and use its spectral properties to embed vertices and hyperedges in a common space. We then cluster these embeddings to obtain our proposed co-clustering method, of particular relevance in applications requiring the simultaneous clustering of data entities and features. Numerical experiments using real-world data demonstrate the effectiveness of our proposed approach in comparison with state-of-the-art alternatives.

We study the temporal reconstruction of epidemics evolving over networks. Given partial or aggregated temporal information of the epidemic, our goal is to estimate the complete evolution of the spread leveraging the topology of the network but being agnostic to the precise epidemic model. We overcome this lack of model awareness through a data-driven solution to the inverse problem at hand. In particular, we propose DDmix, a graph conditional variational autoencoder that can be trained from past epidemic spreads and whose latent space seeks to capture key aspects of the underlying (unknown) spreading dynamics. We illustrate the accuracy and generalizability of DDmix and compare it with non-graph-aware learning algorithms through numerical experiments on epidemic spreads simulated on synthetic and real-world networks.

We develop a novel graph-based trainable framework to maximize the weighted sum energy efficiency (WSEE) for power allocation in wireless communication networks. The framework is designed to unfold a classical iterative algorithm, namely the successive concave approximation (SCA), which suboptimally solves this NP-hard problem. Our unfolded SCA (USCA) framework consists of modular structures, which are enhanced with learnable components parameterized using graph convolutional neural networks (GCNs). Its layers are progressively trained through a stochastic gradient descent approach driven by unsupervised loss and regularization terms. Numerical experiments demonstrate that USCA outperforms state-of-the-art benchmarks across different network topologies of varying size and density.

Predicting mood, health, and stress can sound an early alarm against mental illness. Multi-modal data from wearable sensors provide rigorous and rich insights into one's internal states. Recently, deep learning-based features on continuous high-resolution sensor data have outperformed statistical features in several ubiquitous and affective computing applications including sleep detection and depression diagnosis. Motivated by this, we investigate multi-modal data fusion strategies featuring deep representation learning of skin conductance, skin temperature, and acceleration data to predict self-reported mood, health, and stress scores (0 - 100) of college students (N = 239). Our cross-validated results from the early fusion framework exhibit a significantly higher (p <; 0.05) prediction precision over the late fusion for unseen users. Therefore, our findings call attention to the benefits of fusing physiological data modalities at a low level and corroborate the predictive efficacy of the deeply learned features.

We sought to characterize local lung complexity in chest computed tomography (CT) and to characterize its impact on the detectability of pulmonary nodules. Forty volumetric chest CT scans were created by embedding between three and five simulated 5-mm lung nodules into one of three volumetric chest CT datasets. Thirteen radiologists evaluated 157 nodules, resulting in 2041 detection opportunities. Analyzing the substrate CT data prior to nodule insertion, 14 image features were measured within a region around each nodule location. A generalized linear mixed-effects statistical model was fit to the data to verify the contribution of each metric on detectability. The model was tuned for simplicity, interpretability, and generalizability using stepwise regression applied to the primary features and their interactions. We found that variables corresponding to each of five categories (local structural distractors, local intensity, global context, local vascularity, and contiguity with structural distractors) were significant (p < 0.01) factors in a standardized model. Moreover, reader-specific models conveyed significant differences among readers with significant distraction (missed detections) influenced by local intensity- versus local-structural characteristics being mutually exclusive. Readers with significant local intensity distraction (n = 10) detected substantially fewer lung nodules than those who were significantly distracted by local structure (n = 2), 46.1% versus 65.3% mean nodules detected, respectively.

Continuous wearable sensor data in high resolution contain physiological and behavioral information that can be utilized to predict human health and wellbeing, establishing the foundation for developing early warning systems to eventually improve human health and wellbeing. We propose a deep neural network framework, the Locally Connected Long Short-Term Memory Denoising AutoEncoder (LC-LSTM-DAE), to automatically extract features from passively collected raw sensor data and perform personalized prediction of self-reported mood, health, and stress scores with high precision. We enabled personalized learning of features by finetuning the general representation model with participant-specific data. The framework was evaluated using wearable sensor data and wellbeing labels collected from college students (total 6391 days from N=239). Sensor data include skin temperature, skin conductance, and acceleration; wellbeing labels include self-reported mood, health and stress scored 0 - 100. Compared to the prediction performance based on hand-crafted features, the proposed framework achieved higher precision with a smaller number of features. We also provide statistical interpretation and visual explanation to the automatically learned features and the prediction models. Our results show the possibility of predicting self-reported mood, health, and stress accurately using an interpretable deep learning framework, ultimately for developing real-time health and wellbeing monitoring and intervention systems that can benefit various populations.

Wearable sensors can capture continuous, high resolution physiological and behavioral data that can be utilized to develop early health and wellbeing detection and lead to early warning, intervention, and recommendation systems to improve health and wellbeing. We have built and evaluated an end-to-end wellbeing prediction framework that pipelines raw wearable sensor data into an unsupervised autoencoder-based representation learning model and a supervised wellbeing regression model. We trained and evaluated the framework using the wearable sensor dataset and wellbeing labels collected from college students (total 6391 days from N=252). Wearable data include skin temperature, skin conductance, and acceleration; the wellbeing labels include self-reported alertness, happiness, energy, health, and calmness scored 0 – 100. We compared the performance of our framework with the performance of wellbeing regression models based on hand-crafted features. Our results showed that the proposed framework can automatically extract features from the current day's 24-hour multi-channel data and predict wellbeing scores for next day with mean absolute errors of 14–16. This result shows the possibility of predicting wellbeing accurately using an end-to-end framework, ultimately for developing real-time health and wellbeing monitoring and intervention systems.

The effective integration of distributed solar photovoltaic (PV) arrays into existing power grids will require access to high quality data; the location, power capacity, and energy generation of individual solar PV installations. Unfortunately, existing methods for obtaining this data are limited in their spatial resolution and completeness. We propose a general framework for accurately and cheaply mapping individual PV arrays, and their capacities, over large geographic areas. At the core of this approach is a deep learning algorithm called SolarMapper - which we make publicly available - that can automatically map PV arrays in high resolution overhead imagery. We estimate the performance of SolarMapper on a large dataset of overhead imagery across three US cities in California. We also describe a procedure for deploying SolarMapper to new geographic regions, so that it can be utilized by others. We demonstrate the effectiveness of the proposed deployment procedure by using it to map solar arrays across the entire US state of Connecticut (CT). Using these results, we demonstrate that we achieve highly accurate estimates of total installed PV capacity within each of CT's 168 municipal regions.