Supervised mixture of experts models for population health.

We propose a machine learning driven approach to derive insights from observational healthcare data to improve public health outcomes. Our goal is to simultaneously identify patient subpopulations with differing health risks and to find those risk factors within each subpopulation. We develop two supervised mixture of experts models: a Supervised Gaussian Mixture model (SGMM) for general features and a Supervised Bernoulli Mixture model (SBMM) tailored to binary features. We demonstrate the two approaches on an analysis of high cost drivers of Medicaid expenditures for inpatient stays. We focus on the three diagnostic categories that accounted for the highest percentage of inpatient expenditures in New York State (NYS) in 2016. When compared with state-of-the-art learning methods (random forests, boosting, neural networks), our approaches provide comparable prediction performance while also extracting insightful subpopulation structure and risk factors. For problems with binary features the proposed SBMM provides as good or better performance than alternative methods while offering insightful explanations. Our results indicate the promise of such approaches for extracting population health insights from electronic health care records.

Efficient parameter inference in networked dynamical systems via steady states: A surrogate objective function approach integrating mean-field and nonlinear least squares.

In networked dynamical systems, inferring governing parameters is crucial for predicting nodal dynamics, such as gene expression levels, species abundance, or population density. While many parameter estimation techniques rely on time-series data, particularly systems that converge over extreme time ranges, only noisy steady-state data is available, requiring a new approach to infer dynamical parameters from noisy observations of steady states. However, the traditional optimization process is computationally demanding, requiring repeated simulation of coupled ordinary differential equations. To overcome these limitations, we introduce a surrogate objective function that leverages decoupled equations to compute steady states, significantly reducing computational complexity. Furthermore, by optimizing the surrogate objective function, we obtain steady states that more accurately approximate the ground truth than noisy observations and predict future equilibria when topology changes. We empirically demonstrate the effectiveness of the proposed method across ecological, gene regulatory, and epidemic networks. Our approach provides an efficient and effective way to estimate parameters from steady-state data and has the potential to improve predictions in networked dynamical systems.

Reverse engineering a social agent-based hidden markov model--visage.

We present a machine learning approach to discover the agent dynamics that drives the evolution of the social groups in a community. We set up the problem by introducing an agent-based hidden Markov model for the agent dynamics: an agent's actions are determined by micro-laws. Nonetheless, We learn the agent dynamics from the observed communications without knowing state transitions. Our approach is to identify the appropriate micro-laws corresponding to an identification of the appropriate parameters in the model. The model identification problem is then formulated as a mixed optimization problem. To solve the problem, we develop a multistage learning process for determining the group structure, the group evolution, and the micro-laws of a community based on the observed set of communications among actors, without knowing the semantic contents. Finally, to test the quality of our approximations and the feasibility of the approach, we present the results of extensive experiments on synthetic data as well as the results on real communities, such as Enron email and Movie newsgroups. Insight into agent dynamics helps us understand the driving forces behind social evolution.

Efficient optimal linear boosting of a pair of classifiers.

Boosting is a meta-learning algorithm which takes as input a set of classifiers and combines these classifiers to obtain a better classifier. We consider the combinatorial problem of efficiently and optimally boosting a pair of classifiers by reducing this problem to that of constructing the optimal linear separator for two sets of points in two dimensions. Specifically, let each point x element of R be assigned a weight W(x) > 0, where the weighting function can be an arbitrary positive function. We give efficient (low-order polynomial time) algorithms for constructing an optimal linear "separator" l defined as follows. Let Q be the set of points misclassified by l. Then, the weight of Q, defined as the sum of the weights of the points in Q, is minimized. If W(z) = 1 for all points, then the resulting separator minimizes (exactly) the misclassification error. Without an increase in computational complexity, our algorithm can be extended to output the leave-one-out error, an unbiased estimate of the expected performance of the resulting boosted classifier.