Stephen Adams

Feature Selection for Hidden Markov Models and Hidden Semi-Markov Models

In this paper, a joint feature selection and parameter estimation algorithm is presented for hidden Markov models (HMMs) and hidden semi-Markov models (HSMMs). New parameters, feature saliencies, are introduced to the model and used to select features that distinguish between states. The feature saliencies represent the probability that a feature is relevant by distinguishing between state-dependent and state-independent distributions. An expectation maximization algorithm is used to calculate maximum a posteriori estimates for model parameters. An exponential prior on the feature saliencies is compared with a beta prior. These priors can be used to include cost in the model estimation and feature selection process. This algorithm is tested against maximum likelihood estimates and a variational Bayesian method. For the HMM, four formulations are compared on a synthetic data set generated by models with known parameters, a tool wear data set, and data collected during a painting process. For the HSMM, two formulations, maximum likelihood and maximum a posteriori, are tested on the latter two data sets, demonstrating that the feature saliency method of feature selection can be extended to semi-Markov processes. The literature on feature selection specifically for HMMs is sparse, and non-existent for HSMMs. This paper fills a gap in the literature concerning simultaneous feature selection and parameter estimation for HMMs using the EM algorithm, and introduces the notion of selecting features with respect to cost for HMMs.

Horse race analysis in credit card fraud—deep learning, logistic regression, and Gradient Boosted Tree

Fraud detection is an industry where incremental gains in predictive accuracy can have large benefits for banks and customers. Banks adapt models to the novel ways in which “fraudsters” commit credit card fraud. They collect data and engineer new features in order to increase predictive power. This research compares the algorithmic impact on the predictive power across three supervised classification models: logistic regression, gradient boosted trees, and deep learning. This research also explores the benefits of creating features using domain expertise and feature engineering using an autoencoder—an unsupervised feature engineering method. These two methods of feature engineering combined with the direct mapping of the original variables create six different feature sets. Across these feature sets this research compares the aforementioned models. This research concludes that creating features using domain expertise offers a notable improvement in predictive power. Additionally, the autoencoder offers a way to reduce the dimensionality of the data and slightly boost predictive power.

Adversarial learning in credit card fraud detection

Credit card fraud is an expensive problem for many financial institutions, costing billions of dollars to companies annually. Many adversaries still evade fraud detection systems because these systems often do not include information about the adversarys knowledge of the fraud detection mechanism. This project aims to include information about the “fraudsters” motivations and knowledge base into an adaptive fraud detection system. In this project, we use a game theoretical adversarial learning approach in order to model the fraudsters best strategy and pre-emptively adapt the fraud detection system to better classify these future fraudulent transactions. Using a logistic regression classifier as the fraud detection mechanism, we initially identify the best strategy for the adversary based on the number of fraudulent transactions that go undetected, and assume that the adversary uses this strategy for future transactions in order to improve our classifier. Prior research has used game theoretic models for adversarial learning in the domains of credit card fraud and email spam, but this project adds to the literature by extending these frameworks to a practical, real-world data set. Test results show that our adversarial framework produces an increasing AUC score on validation sets over several iterations in comparison to the static model usually employed by credit card companies.

Systems thinking and predictive analytics to improve veteran healthcare scheduling

As a culture, the United States acknowledges that the delivery of veteran services to those who risked their lives and suffered to protect our nation is a top national priority. Over the past several years, however, the media have reported stark examples of how these services are lacking, particularly in the case of medical appointment scheduling. At the same time, the Veterans Health Administration is plagued by strikingly high no-show rates at its medical outpatient clinics and a resulting handicap in resource allocation. We bring to bear systems thinking to address these issues. As a result, we developed a model for a dynamic overbooking system that receives the probability of a patient arriving on-time for their appointment from the patients phone and couples this real-time probability with prior probability derived from existing VA data. Note that the system protects patient privacy by never transmitting nor sharing location data. When the arrival probability of a patient falls below a given threshold, an algorithm can automatically cancel a patients appointment and re-assign it to another patient drawn from a pool of wait-list and other patients with high arrival probabilities given their current location. In this presentation, we share the progress to date on our approach, and our proposals for future work and implementation.

Author Correction to: Dynamic Scheduling for Veterans Health Administration Patients Using Geospatial Dynamic Overbooking

The original version of this article unfortunately contained a mistake. The name of Matthew Gerber was incorrectly spelled as Mathew Gerber. The correct spelling is now presented correctly in this correction article.

Health-aware hierarchical control for smart manufacturing using reinforcement learning

Manufacturing facilities are laid out in a natural hierarchy of assembly lines, work cells, machines, and components. Currently, prognostics and health management (PHM) information is confined to the lowest levels of this hierarchy and finds primary use in decisions and control policies for machine maintenance and replacement. For the smart manufacturing systems of the future, however, PHM information should be passed to all levels of the hierarchy and incorporated into high level decision making about production quantities, rates, and locations. This paper proposes a hierarchical control methodology that passes PHM health estimates up the hierarchy and optimization objectives down the hierarchy. Individual nodes in the hierarchy are modeled as Markov decision processes (MDPs) and reinforcement learning is used to estimate optimal policies. This work makes several contributions to the PHM community. First, we propose a novel model of a control system that makes uses of health information throughout the manufacturing hierarchy. Second, we define a reinforcement learning based approach to solving the MDPs for optimal or near-optimal policies. Third, we illustrate the method on a numerical example based on a simulation of a real-world manufacturing environment.

Dynamic Scheduling for Veterans Health Administration Patients using Geospatial Dynamic Overbooking

The Veterans Health Administration (VHA) is plagued by abnormally high no-show and cancellation rates that reduce the productivity and efficiency of its medical outpatient clinics. We address this issue by developing a dynamic scheduling system that utilizes mobile computing via geo-location data to estimate the likelihood of a patient arriving on time for a scheduled appointment. These likelihoods are used to update the clinic’s schedule in real time. When a patient’s arrival probability falls below a given threshold, the patient’s appointment is canceled. This appointment is immediately reassigned to another patient drawn from a pool of patients who are actively seeking an appointment. The replacement patients are prioritized using their arrival probability. Real-world data were not available for this study, so synthetic patient data were generated to test the feasibility of the design. The method for predicting the arrival probability was verified on a real set of taxicab data. This study demonstrates that dynamic scheduling using geo-location data can reduce the number of unused appointments with minimal risk of double booking resulting from incorrect predictions. We acknowledge that there could be privacy concerns with regards to government possession of one’s location and offer strategies for alleviating these concerns in our conclusion.

A survey of feature selection methods for Gaussian mixture models and hidden Markov models

Feature selection is the process of reducing the number of collected features to a relevant subset of features and is often used to combat the curse of dimensionality. This paper provides a review of the literature on feature selection techniques specifically designed for Gaussian mixture models (GMMs) and hidden Markov models (HMMs), two common parametric latent variable models. The primary contribution of this work is the collection and grouping of feature selection methods specifically designed for GMMs and for HMMs. An additional contribution lies in outlining the connections between these two groups of feature selection methods. Often, feature selection methods for GMMs and HMMs are treated as separate topics. In this survey, we propose that methods developed for one model can be adapted to the other model. Further, we find that the number of feature selection methods for GMMs outweighs the number of methods for HMMs and that the proportion of methods for HMMs that require supervised data is larger than the proportion of GMM methods that require supervised data. We conclude that further research into unsupervised feature selection methods for HMMs is required and that established methods for GMMs could be adapted to HMMs. It should be noted that feature selection can also be referred to as dimensionality reduction, variable selection, attribute selection, and variable subset reduction. In this paper, we make a distinction between dimensionality reduction and feature selection. Dimensionality reduction, which we do not consider, is any process that reduces the number of features used in a model and can include methods that transform features in order to reduce the dimensionality. Feature selection, by contrast, is a specific form of dimensionality reduction that eliminates feature as inputs into the model. The primary difference is that dimensionality reduction can still require the collection of all the data sources in order to transform and reduce the feature set, while feature selection eliminates the need to collect the irrelevant data sources.