Temporal graph networks (TGNs) have gained prominence as models for embedding dynamic interactions, but little is known about their theoretical underpinnings. We establish fundamental results about the representational power and limits of the two main categories of TGNs: those that aggregate temporal walks (WA-TGNs), and those that augment local message passing with recurrent memory modules (MP-TGNs). Specifically, novel constructions reveal the inadequacy of MP-TGNs and WA-TGNs, proving that neither category subsumes the other. We extend the 1-WL (Weisfeiler-Leman) test to temporal graphs, and show that the most powerful MP-TGNs should use injective updates, as in this case they become as expressive as the temporal WL. Also, we show that sufficiently deep MP-TGNs cannot benefit from memory, and MP/WA-TGNs fail to compute graph properties such as girth. These theoretical insights lead us to PINT -- a novel architecture that leverages injective temporal message passing and relative positional features. Importantly, PINT is provably more expressive than both MP-TGNs and WA-TGNs. PINT significantly outperforms existing TGNs on several real-world benchmarks.

In this setting, we explore two different data-augmentation (DA) policies, one aligned with the data distribution (rotation+translation+scaling) and one largely misaligned with the data distribution (aggressive Gaussian blur). Because our goal is to understand how much DA impacts the SSL kernel compared to a fully supervised benchmark, we consider two (supervised) benchmarks: one that employs the labels of the sampled training set and all the augmented samples and one that only employs the sampled training set and no augmented samples. We explore a small training set size going from N = 16 to N = 256 and for each case we produce a number of augmented samples for each datapoint so that the total number of samples does not exceed 50, 000 which is a standard threshold for kernel methods. We note that our implementation is based on the common Python scientific library Numpy/Scipy (Harris et al., 2020) and runs on CPU. We observed in Figure 3 that the SSL kernel is able to match and even outperform the fully supervised case when employing the correct data-augmentation, while with the incorrect data-augmentation, the performance is not even able to match the supervised case that did not see the augmented samples. To better understand the impact of different hyper-parameters onto the two kernels, we also study in Figure 4 the MNIST test set performances when varying the representation dimension K, the SVM's l

The presence of snow and ice on runway surfaces reduces the available tire-pavement friction needed for retardation and directional control and causes potential economic and safety threats for the aviation industry during the winter seasons. To activate appropriate safety procedures, pilots need accurate and timely information on the actual runway surface conditions. In this study, XGBoost is used to create a combined runway assessment system, which includes a classification model to identify slippery conditions and a regression model to predict the level of slipperiness. The models are trained on weather data and runway reports. The runway surface conditions are represented by the tire-pavement friction coefficient, which is estimated from flight sensor data from landing aircrafts. The XGBoost models are combined with SHAP approximations to provide a reliable decision support system for airport operators, which can contribute to safer and more economic operations of airport runways. To evaluate the performance of the prediction models, they are compared to several state-of-the-art runway assessment methods. The XGBoost models identify slippery runway conditions with a ROC AUC of 0.95, predict the friction coefficient with a MAE of 0.0254, and outperforms all the previous methods. The results show the strong abilities of machine learning methods to model complex, physical phenomena with a good accuracy. Published version: https://doi.org/10.1016/j.coldregions.2022.103556.

Molecular pretraining, which learns molecular representations over massive unlabeled data, has become a prominent paradigm to solve a variety of tasks in computational chemistry and drug discovery. Recently, prosperous progress has been made in molecular pretraining with different molecular featurizations, including 1D SMILES strings, 2D graphs, and 3D geometries. However, the role of molecular featurizations with their corresponding neural architectures in molecular pretraining remains largely unexamined. In this paper, through two case studies -- chirality classification and aromatic ring counting -- we first demonstrate that different featurization techniques convey chemical information differently. In light of this observation, we propose a simple and effective MOlecular pretraining framework with COmplementary featurizations (MOCO). MOCO comprehensively leverages multiple featurizations that complement each other and outperforms existing state-of-the-art models that solely relies on one or two featurizations on a wide range of molecular property prediction tasks.

Enabling robots to perform complex dynamic tasks such as picking up an object in one sweeping motion or pushing off a wall to quickly turn a corner is a challenging problem. The dynamic interactions implicit in these tasks are critical towards the successful execution of such tasks. Graph neural networks (GNNs) provide a principled way of learning the dynamics of interactive systems but can suffer from scaling issues as the number of interactions increases. Furthermore, the problem of using learned GNN-based models for optimal control is insufficiently explored. In this work, we present a method for efficiently learning the dynamics of interacting systems by simultaneously learning a dynamic graph structure and a stable and locally linear forward model of the system. The dynamic graph structure encodes evolving contact modes along a trajectory by making probabilistic predictions over the edges of the graph. Additionally, we introduce a temporal dependence in the learned graph structure which allows us to incorporate contact measurement updates during execution thus enabling more accurate forward predictions. The learned stable and locally linear dynamics enable the use of optimal control algorithms such as iLQR for long-horizon planning and control for complex interactive tasks. Through experiments in simulation and in the real world, we evaluate the performance of our method by using the learned interaction dynamics for control and demonstrate generalization to more objects and interactions not seen during training. We introduce a control scheme that takes advantage of contact measurement updates and hence is robust to prediction inaccuracies during execution.

Air pollution and carbon emissions caused by modern transportation are closely related to global climate change. With the help of next-generation information technology such as Internet of Things (IoT) and Artificial Intelligence (AI), accurate traffic flow prediction can effectively solve problems such as traffic congestion and mitigate environmental pollution and climate change. It further promotes the development of Intelligent Transportation Systems (ITS) and smart cities. However, the strong spatial and temporal correlation of traffic data makes the task of accurate traffic forecasting a significant challenge. Existing methods are usually based on graph neural networks using predefined spatial adjacency graphs of traffic networks to model spatial dependencies, ignoring the dynamic correlation of relationships between road nodes. In addition, they usually use independent Spatio-temporal components to capture Spatio-temporal dependencies and do not effectively model global Spatio-temporal dependencies. This paper proposes a new Spatio-temporal Causal Graph Attention Network (STCGAT) for traffic prediction to address the above challenges. In STCGAT, we use a node embedding approach that can adaptively generate spatial adjacency subgraphs at each time step without a priori geographic knowledge and fine-grained modeling of the topology of dynamically generated graphs for different time steps. Meanwhile, we propose an efficient causal temporal correlation component that contains node adaptive learning, graph convolution, and local and global causal temporal convolution modules to learn local and global Spatio-temporal dependencies jointly. Extensive experiments on four real, large traffic datasets show that our model consistently outperforms all baseline models.

This paper is concerned with forecast error, particularly in relation to loss reserving. This is generally regarded as consisting of three components, namely parameter, process and model errors. The first two of these components, and their estimation, are well understood, but less so model error. Model error itself is considered in two parts: one part that is capable of estimation from past data (internal model error), and another part that is not (external model error). Attention is focused here on internal model error. Estimation of this error component is approached by means of Bayesian model averaging, using the Bayesian interpretation of the LASSO. This is used to generate a set of admissible models, each with its prior probability and the likelihood of observed data. A posterior on the model set, conditional on the data, results, and an estimate of model error (contained in a loss reserve) is obtained as the variance of the loss reserve according to this posterior. The population of models entering materially into the support of the posterior may turn out to be thinner than desired, and bootstrapping of the LASSO is used to gain bulk. This provides the bonus of an estimate of parameter error also. It turns out that the estimates of parameter and model errors are entangled, and dissociation of them is at least difficult, and possibly not even meaningful. These matters are discussed. The majority of the discussion applies to forecasting generally, but numerical illustration of the concepts is given in relation to insurance data and the problem of insurance loss reserving.

Concept-based explanations permit to understand the predictions of a deep neural network (DNN) through the lens of concepts specified by users. Existing methods assume that the examples illustrating a concept are mapped in a fixed direction of the DNN's latent space. When this holds true, the concept can be represented by a concept activation vector (CAV) pointing in that direction. In this work, we propose to relax this assumption by allowing concept examples to be scattered across different clusters in the DNN's latent space. Each concept is then represented by a region of the DNN's latent space that includes these clusters and that we call concept activation region (CAR). To formalize this idea, we introduce an extension of the CAV formalism that is based on the kernel trick and support vector classifiers. This CAR formalism yields global concept-based explanations and local concept-based feature importance. We prove that CAR explanations built with radial kernels are invariant under latent space isometries. In this way, CAR assigns the same explanations to latent spaces that have the same geometry. We further demonstrate empirically that CARs offer (1) more accurate descriptions of how concepts are scattered in the DNN's latent space; (2) global explanations that are closer to human concept annotations and (3) concept-based feature importance that meaningfully relate concepts with each other. Finally, we use CARs to show that DNNs can autonomously rediscover known scientific concepts, such as the prostate cancer grading system.

In real-world applications, Federated Learning (FL) meets two challenges: (1) scalability, especially when applied to massive IoT networks, and (2) how to be robust against an environment with heterogeneous data. Realizing the first problem, we aim to design a novel FL framework named Full-stack FL (F2L). Moreover, leveraging the advantages of hierarchical network design, we propose a new labeldriven knowledge distillation (LKD) technique at the global server to address the second problem. As opposed to current knowledge distillation techniques, LKD is capable of training a student model, which consists of good knowledge from all teachers' models. Therefore, our proposed algorithm can effectively extract the knowledge of the regions' data distribution (i.e., the regional aggregated models) to reduce the divergence between clients' models when operating under the FL system with non-independent identically distributed data. Extensive experiment results reveal that: (i) our F2L method can significantly improve the overall FL efficiency in all global distillations, and (ii) F2L rapidly achieves convergence as global distillation stages occur instead of increasing on each communication cycle. Recently, Federated Learning (FL) is known as a novel distributed learning methodology for enhancing communication efficiency and ensuring privacy in traditional centralized one McMahan et al. (2017). However, the most challenge of this method for client models is non-independent and identically distributed (non-IID) data, which leads to divergence into unknown directions. Inspired by this, various works on handling non-IID were proposed in Li et al. (2020); Acar et al. (2021); Dinh et al. (2021a); Karimireddy et al. (2020); Wang et al. (2020); Zhu et al. (2021); Nguyen et al. (2022b). However, these works mainly rely on arbitrary configurations without thoroughly understanding the models' behaviors, yielding low-efficiency results. Aiming to fulfil this gap, in this work, we propose a new hierarchical FL framework using information theory by taking a deeper observation of the model's behaviors, and this framework can be realized for various FL systems with heterogeneous data. In addition, our proposed framework can trigger the FL system to be more scalable, controllable, and accessible through hierarchical architecture.

We study the application of variance reduction (VR) techniques to general non-convex stochastic optimization problems. In this setting, the recent work STORM [Cutkosky-Orabona '19] overcomes the drawback of having to compute gradients of "mega-batches" that earlier VR methods rely on. There, STORM utilizes recursive momentum to achieve the VR effect and is then later made fully adaptive in STORM+ [Levy et al., '21], where full-adaptivity removes the requirement for obtaining certain problem-specific parameters such as the smoothness of the objective and bounds on the variance and norm of the stochastic gradients in order to set the step size. However, STORM+ crucially relies on the assumption that the function values are bounded, excluding a large class of useful functions. In this work, we propose META-STORM, a generalized framework of STORM+ that removes this bounded function values assumption while still attaining the optimal convergence rate for non-convex optimization. META-STORM not only maintains full-adaptivity, removing the need to obtain problem specific parameters, but also improves the convergence rate's dependency on the problem parameters. Furthermore, META-STORM can utilize a large range of parameter settings that subsumes previous methods allowing for more flexibility in a wider range of settings. Finally, we demonstrate the effectiveness of META-STORM through experiments across common deep learning tasks. Our algorithm improves upon the previous work STORM+ and is competitive with widely used algorithms after the addition of per-coordinate update and exponential moving average heuristics.