Background: The methods with which prediction models are usually developed mean that neither the parameters nor the predictions should be interpreted causally. For many applications this is perfectly acceptable. However, when prediction models are used to support decision making, there is often a need for predicting outcomes under hypothetical interventions. Aims: We aimed to identify and compare published methods for developing and validating prediction models that enable risk estimation of outcomes under hypothetical interventions, utilizing causal inference. We aimed to identify the main methodological approaches, their underlying assumptions, targeted estimands, and possible sources of bias. Finally, we aimed to highlight unresolved methodological challenges. Methods: We systematically reviewed literature published by December 2019, considering papers in the health domain that used causal considerations to enable prediction models to be used to evaluate predictions under hypothetical interventions. We included both methodology development studies and applied studies. Results: We identified 4919 papers through database searches and a further 115 papers through manual searches. Of these, 87 papers were retained for full text screening, of which 12 were selected for inclusion. We found papers from both the statistical and the machine learning literature. Most of the identified methods for causal inference from observational data were based on marginal structural models and g-estimation.

This study introduces a framework for quality control of measured weather data, including anomaly detection, and infilling missing values. Weather data is a fundamental input to building performance simulations, in which anomalous values defect the results while missing data lead to an unexpected termination of the simulation process. Traditionally, infilling missing values in weather data is performed through periodic or linear interpolations. However, when missing values exceed many consecutive hours, the accuracy of traditional methods is subject to debate. This study demonstrates how Neural Networks can increase the accuracy of data imputation when compared to other supervised learning methods. The framework is validated by predicting missing temperature and relative humidity data for an observation site, through a network of nearby weather stations in Milan, Italy. Results show that the proposed method can facilitate real-time building simulations with accurate and rapid quality control.

Adversarial examples are a widely studied phenomenon in machine learning models. While most of the attention has been focused on neural networks, other practical models also suffer from this issue. In this work, we propose an algorithm for evaluating the adversarial robustness of $k$-nearest neighbor classification, i.e., finding a minimum-norm adversarial example. Diverging from previous proposals, we take a geometric approach by performing a search that expands outwards from a given input point. On a high level, the search radius expands to the nearby Voronoi cells until we find a cell that classifies differently from the input point. To scale the algorithm to a large $k$, we introduce approximation steps that find perturbations with smaller norm, compared to the baselines, in a variety of datasets. Furthermore, we analyze the structural properties of a dataset where our approach outperforms the competition.

Adjusting the learning rate schedule in stochastic gradient methods is an important unresolved problem which requires tuning in practice. If certain parameters of the loss function such as smoothness or strong convexity constants are known, theoretical learning rate schedules can be applied. However, in practice, such parameters are not known, and the loss function of interest is not convex in any case. The recently proposed batch normalization reparametrization is widely adopted in most neural network architectures today because, among other advantages, it is robust to the choice of Lipschitz constant of the gradient in loss function, allowing one to set a large learning rate without worry. Inspired by batch normalization, we propose a general nonlinear update rule for the learning rate in batch and stochastic gradient descent so that the learning rate can be initialized at a high value, and is subsequently decreased according to gradient observations along the way. The proposed method is shown to achieve robustness to the relationship between the learning rate and the Lipschitz constant, and near-optimal convergence rates in both the batch and stochastic settings ($O(1/T)$ for smooth loss in the batch setting, and $O(1/\sqrt{T})$ for convex loss in the stochastic setting). We also show through numerical evidence that such robustness of the proposed method extends to highly nonconvex and possibly non-smooth loss function in deep learning problems.Our analysis establishes some first theoretical understanding into the observed robustness for batch normalization and weight normalization.

BEGIN ARTICLE PREVIEW: Ensemble methods combine several machine learning models to improve results This article is part of Demystifying AI, a series of posts that (try to) disambiguate the jargon and myths surrounding AI. The principle of “the wisdom of the crowd” shows that a large group of people with average knowledge on a topic can provide reliable answers to questions such as predicting quantities, spatial reasoning, and general knowledge. The aggregate results cancel out the noise and can often be superior to those of highly knowledgeable experts. The same rule can apply to artificial intelligence applications that rely on machine learning, the branch of AI that predicts outcomes based on mathematical models. In machine learning,

Nontechnical stakeholders struggle to define business requirements. Crossfunctional teams face an uphill battle to set up robust pipelines for replicable data delivery. Machine learning models can take on a life of their own. If you've been ignoring these critical elements in the past, you may find your deployment rate skyrockets. Your data products may depend on correctly deploying the tips from this article.

Training-serving skew: The offline numbers may look great, but what if your model will be evaluated on a different or broader set of examples than those found in the training set? This phenomenon, more generally referred to as "dataset shift" or "distribution shift", happens all the time in real-world situations. Models are trained on a curated set of examples, or clicks on top-ranked recommendations, or a specific geographical region, and then applied to every user or use case. Curiosities and anomalies in your training and testing data become genuine and sustained loss patterns. Bad individual errors: Models are often judged by their worst behavior --- a single egregious outcome can damage the faith that important stakeholders have in the model and even cause serious reputational harm to your business or institution.

Estimation models from connected vehicles often assume low level parameters such as arrival rates and market penetration rates as known or estimate them in real-time. At low market penetration rates, such parameter estimators produce large errors making estimated queue lengths inefficient for control or operations applications. In order to improve accuracy of low level parameter estimations, this study investigates the impact of connected vehicles information filtering on queue length estimation models. Filters are used as multilevel real-time estimators. Accuracy is tested against known arrival rate and market penetration rate scenarios using microsimulations. To understand the effectiveness for short-term or for dynamic processes, arrival rates, and market penetration rates are changed every 15 minutes. The results show that with Kalman and Particle filters, parameter estimators are able to find the true values within 15 minutes and meet and surpass the accuracy of known parameter scenarios especially for low market penetration rates. In addition, using last known estimated queue lengths when no connected vehicle is present performs better than inputting average estimated values. Moreover, the study shows that both filtering algorithms are suitable for real-time applications that require less than 0.1 second computational time.

Recent advances in training of neural networks make high-dimensional numerical studies feasible for decision problems in uncertain environments. Although reinforcement learning has been widely used in optimal control for several decades [6], only recently Han and E [18], Han et al. [20] combine it with Monte Carlo type regression for the off-line construction of optimal feedback actions. In these problems, the randomness and the state are observable and a training set based on historical or simulated data is readily available. One then approximates the objective functions of these problems by the empirical averages over this training data, constructing a loss function which is minimized over the network parameters. The minimizer or a near-minimizer is the trained network and it is an approximation of the optimal feedback action.

In transportation applications such as real-time route guidance, ramp metering, congestion pricing and special events traffic management, accurate short-term traffic flow prediction is needed. For this purpose, this paper proposes several novel \textit{online} Grey system models (GM): GM(1,1$|cos(\omega t)$), GM(1,1$|sin(\omega t)$, $cos(\omega t)$), and GM(1,1$|e^{-at}$,$sin(\omega t)$,$cos(\omega t)$). To evaluate the performance of the proposed models, they are compared against a set of benchmark models: GM(1,1) model, Grey Verhulst models with and without Fourier error corrections, linear time series model, and nonlinear time series model. The evaluation is performed using loop detector and probe vehicle data from California, Virginia, and Oregon. Among the benchmark models, the error corrected Grey Verhulst model with Fourier outperformed the GM(1,1) model, linear time series, and non-linear time series models. In turn, the three proposed models, GM(1,1$|cos(\omega t)$), GM(1,1$|sin(\omega t)$,$cos(\omega t)$), and GM(1,1$|e^{-at}$,$sin(\omega t)$,$cos(\omega t)$), outperformed the Grey Verhulst model in prediction by at least $65\%$, $16\%$, and $11\%$, in terms of Root Mean Squared Error, and by $82\%$, $58\%$, and $42\%$, in terms of Mean Absolute Percentage Error, respectively. It is observed that the proposed Grey system models are more adaptive to location (e.g., perform well for all roadway types) and traffic parameters (e.g., speed, travel time, occupancy, and volume), and they do not require as many data points for training (4 observations are found to be sufficient).