Background: The current gold standard for measuring sleep is polysomnography (PSG), but it can be obtrusive and costly. Actigraphy is a relatively low-cost and unobtrusive alternative to PSG. Of particular interest in measuring sleep from actigraphy is prediction of sleep-wake states. Current literature on prediction of sleep-wake states from actigraphy consists of methods that use population data, which we call generalized models. However, accounting for variability of sleep patterns across individuals calls for personalized models of sleep-wake states prediction that could be potentially better suited to individual-level data and yield more accurate estimation of sleep.

Low-rank tensor decomposition and completion have attracted significant interest from academia given the ubiquity of tensor data. However, the low-rank structure is a global property, which will not be fulfilled when the data presents complex and weak dependencies given specific graph structures. One particular application that motivates this study is the spatiotemporal data analysis. As shown in the preliminary study, weakly dependencies can worsen the low-rank tensor completion performance. In this paper, we propose a novel low-rank CANDECOMP / PARAFAC (CP) tensor decomposition and completion framework by introducing the $L_{1}$-norm penalty and Graph Laplacian penalty to model the weakly dependency on graph. We further propose an efficient optimization algorithm based on the Block Coordinate Descent for efficient estimation. A case study based on the metro passenger flow data in Hong Kong is conducted to demonstrate improved performance over the regular tensor completion methods.

Research has found that many occupational accidents are foreseeable, being the result of people's unsafe behaviour from a retrospective point of view. The prediction of workers' safety behaviour will enable the prior insights into each worker's behavioural tendency and will be useful in the design of management practices prior to the occurrence of accidents and contribute to the reduction of injury rates. In recent years, researchers have found that people do have stable predispositions to engage in certain safety behavioural patterns which vary among individuals as a function of personality features. In this study, an innovative forecasting model, which employs machine learning algorithms, is developed to estimate construction workers' behavioural tendency based on the Big Five personality taxonomy. The data-driven nature of machine learning technique enabled a reliable estimate of the personality-safety behaviour relationship, which allowed this study to provide novel insight that nonlinearity may exist in the relationship between construction workers' personality traits and safety behaviour. The developed model is found to be sufficient to have satisfactory accuracy in explaining and predicting workers' safety behaviour. This finding provides the empirical evidence to support the usefulness of personality traits as effective predictors of people's safety behaviour at work. In addition, this study could have practical implications. The machine learning model developed can help identify vulnerable workers who are more prone to undertake unsafe behaviours, which is proven to have good prediction accuracy and is thereby potentially useful for decision making and safety management on construction sites.

A key requirement for supervised machine learning is labeled training data, which is created by annotating unlabeled data with the appropriate class. Because this process can in many cases not be done by machines, labeling needs to be performed by human domain experts. This process tends to be expensive both in time and money, and is prone to errors. Additionally, reviewing an entire labeled dataset manually is often prohibitively costly, so many real world datasets contain mislabeled instances. To address this issue, we present in this paper a non-parametric end-to-end pipeline to find mislabeled instances in numerical, image and natural language datasets. We evaluate our system quantitatively by adding a small number of label noise to 29 datasets, and show that we find mislabeled instances with an average precision of more than 0.84 when reviewing our system's top 1\% recommendation. We then apply our system to publicly available datasets and find mislabeled instances in CIFAR-100, Fashion-MNIST, and others. Finally, we publish the code and an applicable implementation of our approach.

In these online tools, patients input their initial symptoms and then proceed to answer a series of questions that the system deems relevant to those symptoms. The output of these online tools is a differential diagnosis (ranked list of diseases) that helps educate patients on possible relevant health conditions. Online symptom checkers are powered by underlying diagnosis models or engines similar to those used for advising physicians in "clinical decision support tools"; the main difference in this scenario being that the resulting differential diagnosis is not directly shared with the patient, but rather used by a physician for professional evaluation. Diagnosis models must have high accuracy while covering a large space of symptoms and diseases to be useful to patients and physicians. Accuracy is critically important, as incorrect diagnoses can give patients unnecessary cause for concern.

Despite their successes, deep neural networks still make unreliable predictions when faced with test data drawn from a distribution different to that of the training data, constituting a major problem for AI safety. While this motivated a recent surge in interest in developing methods to detect such out-of-distribution (OoD) inputs, a robust solution is still lacking. We propose a new probabilistic, unsupervised approach to this problem based on a Bayesian variational autoencoder model, which estimates a full posterior distribution over the decoder parameters using stochastic gradient Markov chain Monte Carlo, instead of fitting a point estimate. We describe how information-theoretic measures based on this posterior can then be used to detect OoD data both in input space as well as in the model's latent space. The effectiveness of our approach is empirically demonstrated.

As the size and richness of available datasets grow larger, the opportunities for solving increasingly challenging problems with algorithms learning directly from data grow at the same pace. Consequently, the capability of learning algorithms to work with large amounts of data has become a crucial scientific and technological challenge for their practical applicability. Hence, it is no surprise that large-scale learning is currently drawing plenty of research effort in the machine learning research community. In this thesis, we focus on kernel methods, a theoretically sound and effective class of learning algorithms yielding nonparametric estimators. Kernel methods, in their classical formulations, are accurate and efficient on datasets of limited size, but do not scale up in a cost-effective manner. Recent research has shown that approximate learning algorithms, for instance random subsampling methods like Nystr\"om and random features, with time-memory-accuracy trade-off mechanisms are more scalable alternatives. In this thesis, we provide analyses of the generalization properties and computational requirements of several types of such approximation schemes. In particular, we expose the tight relationship between statistics and computations, with the goal of tailoring the accuracy of the learning process to the available computational resources. Our results are supported by experimental evidence on large-scale datasets and numerical simulations. We also study how large-scale learning can be applied to enable accurate, efficient, and reactive lifelong learning for robotics. In particular, we propose algorithms allowing robots to learn continuously from experience and adapt to changes in their operational environment. The proposed methods are validated on the iCub humanoid robot in addition to other benchmarks.

Federated learning (FL) is a distributed learning approach where a set of end-user devices participate in the learning process by acting on their isolated local data sets. Here, we process local data sets of users where worst-case optimization theory is used to reformulate the FL problem where the impact of local data sets in training phase is considered as an uncertain function bounded in a closed uncertainty region. This representation allows us to compare the performance of FL with its centralized counterpart, and to replace the uncertain function with a concept of protection functions leading to more tractable formulation. The latter supports applying a regularization factor in each user cost function in FL to reach a better performance. We evaluated our model using the MNIST data set versus the protection function parameters, e.g., regularization factors.

We introduce a novel framework for analysing stationary time series based on optimal transport distances and spectral embeddings. First, we represent time series by their power spectral density (PSD), which summarises the signal energy spread across the Fourier spectrum. Second, we endow the space of PSDs with the Wasserstein distance, which capitalises its unique ability to preserve the geometric information of a set of distributions. These two steps enable us to define the Wasserstein-Fourier (WF) distance, which allows us to compare stationary time series even when they differ in sampling rate, length, magnitude and phase. We analyse the features of WF by blending the properties of the Wasserstein distance and those of the Fourier transform. The proposed WF distance is then used in three sets of key time series applications considering real-world datasets: (i) interpolation of time series leading to data augmentation, (ii) dimensionality reduction via non-linear PCA, and (iii) parametric and non-parametric classification tasks. Our conceptual and experimental findings validate the general concept of using divergences of distributions, especially the Wasserstein distance, to analyse time series through comparing their spectral representations.

In mixed multi-view data, multiple sets of diverse features are measured on the same set of samples. By integrating all available data sources, we seek to discover common group structure among the samples that may be hidden in individualistic cluster analyses of a single data-view. While several techniques for such integrative clustering have been explored, we propose and develop a convex formalization that will inherit the strong statistical, mathematical and empirical properties of increasingly popular convex clustering methods. Specifically, our Integrative Generalized Convex Clustering Optimization (iGecco) method employs different convex distances, losses, or divergences for each of the different data views with a joint convex fusion penalty that leads to common groups. Additionally, integrating mixed multi-view data is often challenging when each data source is high-dimensional. To perform feature selection in such scenarios, we develop an adaptive shifted group-lasso penalty that selects features by shrinking them towards their loss-specific centers. Our so-called iGecco+ approach selects features from each data-view that are best for determining the groups, often leading to improved integrative clustering. To fit our model, we develop a new type of generalized multi-block ADMM algorithm using sub-problem approximations that more efficiently fits our model for big data sets. Through a series of numerical experiments and real data examples on text mining and genomics, we show that iGecco+ achieves superior empirical performance for high-dimensional mixed multi-view data.