Along with the increase of digital healthcare providers, the interest in diagnostic aids for remote diagnosis has increased as well. As patients write about their symptoms themselves, we have access to a type of data which previously was rarely recorded, and which has not been filtered by a healthcare professional. Knowledge of similar patients and similar symptoms is beneficial for doctors to arrive at a diagnosis. Therefore, the remote diagnostic process could be aided by presenting patient cases together with information about similar patients and their self-reported symptom descriptions. Apart from online diagnosis, such an aid could be beneficial in many healthcare settings, such as long-distance visits and knowledge gain from patient diaries. In this paper, we present the impact of aiding remote diagnosis by presenting clusters of similar symptoms, using symptom descriptions collected from a virtual visit application by the Swedish telemedicine provider KRY. Symptom descriptions were represented using the bag-of-words model and were then clustered using the k-means algorithm. An experiment was then conducted with 13 doctors, where patient cases were presented together with the most representative words of the associated cluster, to measure how their work was impacted. Results indicated that it was useful in more complicated cases, but also that future experiments will require further instructions on how the information is to be interpreted.

The accuracy of pattern-classification methods depends on how well the measured characteristics (i.e., features) represent the object to be classified. When using pre-designed features as it is the case of many pattern classifiers, one can try to enhance the features' discriminative power by inserting high-level semantic information into the feature vectors. In this paper, we propose a method that increases the discriminative power of features by augmenting them with high-level semantic information learned from training data. Our method combines the advantages of dimensionality reduction techniques and feature-selection techniques. Instead of augmenting feature vectors, we map them using a modified neural network that has been trained to categorize the data into target groups. This neural network embeds categorization information. We tested the method on classification tasks for pollen species, human action, and acoustic signals. In all these tasks, our feature-enhancing method improved classification rates.

Categorical regressor variables are usually handled by introducing a set of indicator variables, and imposing a linear constraint to ensure identifiability in the presence of an intercept, or equivalently, using one of various coding schemes. As proposed in Yuan and Lin [J. R. Statist. Soc. B, 68 (2006), 49-67], the group lasso is a natural and computationally convenient approach to perform variable selection in settings with categorical covariates. As pointed out by Simon and Tibshirani [Stat. Sin., 22 (2011), 983-1001], "standardization" by means of block-wise orthonormalization of column submatrices each corresponding to one group of variables can substantially boost performance. In this note, we study the aspect of standardization for the special case of categorical predictors in detail. The main result is that orthonormalization is not required; column-wise scaling of the design matrix followed by re-scaling and centering of the coefficients is shown to have exactly the same effect. Similar reductions can be achieved in the case of interactions. The extension to the so-called sparse group lasso, which additionally promotes within-group sparsity, is considered as well. The importance of proper standardization is illustrated via extensive simulations.

The K-nearest neighbors algorithm (k-NN) is widely used in instance-based learning and case-based reasoning. The basic k-NN approach has been refined and augmented in many ways, including the use of local weighting, asymmetric metrics, and class-specific weighting, which enables the use of different similarity criteria for each class. This paper extends class-specific weighting with a method we call class-to-class (C2C) weighting. Beyond class-specific weighting, which learns feature weightings to identify the most similar cases to a class, C2C weighting also focuses on learning differences between classes to potentially apply those differences to classification. Once C2C weighting has learned how class C_1 is different from class C_2, given a new case is different from a C_1 case in a way similar to the way C_2 cases are different from C_1 cases, then the new case is assigned to class C_2. C2C offers two potential advantages: First, unlike global weighting, it is robust to deletion of the cases in a given class, because non-native class weightings can still make relatively good predictions. We demonstrate experimentally that this can be true even when a whole class of cases is dropped. Additionally, C2C might provide a new potential form of explainability, in explaining classifications based on pattern of differences. Preliminary results suggest that in normal settings C2C offers accuracy comparable to standard methods, though slightly lower. However, with our initial learning method, the native class weightings of C2C weighting are easily skewed and can lead to worse performance than traditional global weightings. We argue this is not an intrinsic flaw in C2C weighting, but rather an issue in the combination of C2C weighting with global weighting, and propose an approach to address this issue.

We propose a supervised approach incorporating groupfeature sparsity in multi-class kernel logistic regression(GFR-MKLR). The need for group sparsity arises inseveral practical situations where a subset of a set offactors can explain a predicted variable and each factorconsists of a group of variables. We apply our approachfor predicting human interactions based on bodyparts motion (e.g., hands, legs, head, etc.) where imagefeatures are organised in groups corresponding to bodyparts. Our approach, leads to sparse models by assigningweights to groups of features having the highest discriminationbetween different types of interactions. Experimentsconducted on the UT-Interaction dataset havedemonstrated the performance of our method with regardto stat-of-art methods.

Artificial Intelligence (AI) techniques are essential to the modeling of realistic behaviors for agents in simulation systems. Although Case-Based Reasoning (CBR) and Clustering techniques are being explored in the implementation of such agents in computer games, these techniques are still under-used in the implementation of simulation systems. This work approaches this gap by proposing a new CBR and clustering framework in which clustering algorithms and clustering evaluation techniques are explored in both the construction of adjusted similarity functions and the organization of sub-case bases, which are indexing components to the efficient retrieval of relevant cases from case bases so as to support the solution of new simulation problems. To evaluate this framework, a case-based algorithm was implemented to simulate the choice of military supplies to be used in artillery battery missions in virtual tactical simulations.

In this paper, we provide a novel approach for the Team Formation Problem (TFP) in social networks. With a given social network of experts and communication cost between them, we address the problem of finding a team with a set of required skills necessary to complete a project. An expert of this network is treated as a node and possesses a given set of skills. The basic idea of the Structural Clustering Algorithm for Networks (SCAN) is to detect clusters, hubs, and outliers in networks. To employ SCAN on TFP, we first find the pool of experts with required skills. Then we search for highly connected (core) expert among all experts network. We expand the cluster from core to neighborhood nodes, and it goes from densely connected to loosely connected nodes within a threshold range of communication cost. We solve this TFP by identifying experts while minimizing communication cost for the project with specific skills. We then measure the communication cost with the sum of a distance function. An enhanced variant of SCAN is the weighted structural clustering algorithm (WSCAN) which is implemented in this paper to solve the TFP with minimum communication cost. Our result with WSCAN performed approximately equal to Greedy algorithms while slightly worse than other Genetic, Cultural and Exact Algorithms. The run-time of WSCAN however, was better compared to the others.

In this paper we explore acceleration techniques for large scale nonconvex optimization problems with special focuses on deep neural networks. The extrapolation scheme is a classical approach for accelerating stochastic gradient descent for convex optimization, but it does not work well for nonconvex optimization typically. Alternatively, we propose an interpolation scheme to accelerate nonconvex optimization and call the method Interpolatron. We explain motivation behind Interpolatron and conduct a thorough empirical analysis. Empirical results on DNNs of great depths (e.g., 98-layer ResNet and 200-layer ResNet) on CIFAR-10 and ImageNet show that Interpolatron can converge much faster than the state-of-the-art methods such as the SGD with momentum and Adam. Furthermore, Anderson's acceleration, in which mixing coefficients are computed by least-squares estimation, can also be used to improve the performance. Both Interpolatron and Anderson's acceleration are easy to implement and tune. We also show that Interpolatron has linear convergence rate under certain regularity assumptions.

Causal inference from observation data often assumes "strong ignorability," that all confounders are observed. This assumption is standard yet untestable. However, many scientific studies involve multiple causes, different variables whose effects are simultaneously of interest. We propose the deconfounder, an algorithm that combines unsupervised machine learning and predictive model checking to perform causal inference in multiple-cause settings. The deconfounder infers a latent variable as a substitute for unobserved confounders and then uses that substitute to perform causal inference. We develop theory for when the deconfounder leads to unbiased causal estimates, and show that it requires weaker assumptions than classical causal inference. We analyze its performance in three types of studies: semi-simulated data around smoking and lung cancer, semi-simulated data around genomewide association studies, and a real dataset about actors and movie revenue. The deconfounder provides a checkable approach to estimating close-to-truth causal effects.

We consider the use of no-regret algorithms to compute equilibria for particular classes of convex-concave games. While standard regret bounds would lead to convergence rates on the order of $O(T^{-1/2})$, recent work \citep{RS13,SALS15} has established $O(1/T)$ rates by taking advantage of a particular class of optimistic prediction algorithms. In this work we go further, showing that for a particular class of games one achieves a $O(1/T^2)$ rate, and we show how this applies to the Frank-Wolfe method and recovers a similar bound \citep{D15}. We also show that such no-regret techniques can even achieve a linear rate, $O(\exp(-T))$, for equilibrium computation under additional curvature assumptions.