Analyzing healthcare data poses several challenges including the limited number of samples, missing measurements, noisy labels, and heterogeneous data types. Tree-based boosting is well-suited for modeling such data as it is insensitive to data types and missingness. Moreover, Stochastic Gradient TreeBoost is often found in many winning solutions in public data science challenges. Unfortunately, the best performance requires extensive hyperparameter tuning and can be prone to overfitting. We propose PaloBoost, a Stochastic Gradient TreeBoost model that uses novel regularization techniques to guard against overfitting and is robust to hyperparameter settings. PaloBoost uses the out-of-bag samples to perform gradient-aware pruning and estimate adaptive learning rates. Unlike other Stochastic Gradient TreeBoost models that use the out-of-bag samples to estimate test errors, PaloBoost treats the samples as a second batch of training samples to prune the trees and adjust the learning rates. As a result, PaloBoost can dynamically adjust tree depths and learning rates to achieve faster learning at the start and slower learning as the algorithm converges. Experimental results on four datasets demonstrate that PaloBoost is robust to overfitting and is less sensitive to the hyperparameters.

Latent vector autoregressive models for categorical time series have a wide range of potential applications from marketing research to healthcare analytics. However, a brute-force particle filter implementation of the Expectation-Maximization (EM) algorithm often fails to estimate the maximum likelihood parameters due to the Monte Carlo approximation of the E-step and multiple local optima of the log-likelihood function. This paper proposes two auxiliary techniques that help stabilize and calibrate the estimated parameters. These two techniques, namely \textitasymptotic mean regularization and \textitlow-resolution augmentation, do not require any additional parameter tuning, and can be implemented by modifying the brute-force EM algorithm. Experiments with simulated data show that the proposed techniques effectively stabilize the parameter estimation process. Also, experimental results using Medicare and MIMIC-II datasets illustrate various potential applications of the proposed model and methods.

While temporal behavioral patterns can be discerned to underlie real crowd work, prior studies have typically modeled worker performance under a simplified i.i.d. assumption. To better model such temporal worker behavior, we propose a time-series label prediction model for crowd work. This latent variable model captures and summarizes past worker behavior, enabling us to better predict the quality of each worker's next label. Given inherent uncertainty in prediction, we also investigate a decision reject option to balance the tradeoff between prediction accuracy vs. coverage. Results show our model improves accuracy of both label prediction on real crowd worker data, as well as data quality overall.

In the past few years, the government and other agencies have publicly released a prodigious amount of data that can be potentially mined to benefit the society at large. However, data such as health records are typically only provided at aggregated levels (e.g. per State, per Hospital Referral Region, etc.) to protect privacy. Unfortunately aggregation can severely diminish the utility of such data when modeling or analysis is desired at a per-individual basis. So, not surprisingly, despite the increasing abundance of aggregate data, there have been very few successful attempts in exploiting them for individual-level analyses. This paper introduces LUDIA, a novel low-rank approximation algorithm that utilizes aggregation constraints in addition to auxiliary information in order to estimate or "reconstruct" the original individual-level values from aggregate data. If the reconstructed data are statistically similar to the original individual-level data, off-the-shelf individual-level models can be readily and reliably applied for subsequent predictive or descriptive analytics. LUDIA is more robust to nonlinear estimates and random effects than other reconstruction algorithms. It solves a Sylvester equation and leverages multi-level (also known as hierarchical or mixed-effect) modeling approaches efficiently. A novel graphical model is also introduced to provide a probabilistic viewpoint of LUDIA. Experimental results using a Texas inpatient dataset show that individual-level data can be reasonably reconstructed from county-, hospital-, and zip code-level aggregate data. Several factors affecting the reconstruction quality are discussed, along with the implications of this work for current aggregation guidelines.

This paper introduces two kinds of decision tree ensembles for imbalanced classification problems, extensively utilizing properties of α-divergence. First, a novel splitting criterion based on α-divergence is shown to generalize several well-known splitting criteria such as those used in C4.5 and CART. When the α-divergence splitting criterion is applied to imbalanced data, one can obtain decision trees that tend to be less correlated (α-diversification) by varying the value of α. This increased diversity in an ensemble of such trees improves AUROC values across a range of minority class priors. The second ensemble uses the same alpha trees as base classifiers, but uses a lift-aware stopping criterion during tree growth. The resultant ensemble produces a set of interpretable rules that provide higher lift values for a given coverage, a property that is much desirable in applications such as direct marketing. Experimental results across many class-imbalanced data sets, including BRFSS, and MIMIC data sets from the medical community and several sets from UCI and KEEL are provided to highlight the effectiveness of the proposed ensembles over a wide range of data distributions and of class imbalance.

Cardiac arrest is a deadly condition caused by a sudden failure of the heart with an in-hospital mortality rate of ∼80%. Therefore, the ability to accurately estimate patients at high risk of cardiac arrest is crucial for improving the survival rate. Existing research generally fails to utilize a patient’s temporal dynamics. In this paper, we present two dynamic cardiac risk estimation models, focusing on different temporal signatures in a patient’s risk trajectory. These models can track a patient’s risk trajectory in real time, allow interpretability and predictability of a cardiac arrest event, provide an intuitive visualization to medical professionals, offer a personalized dynamic hazard function, and estimate the risk for a new patient.

In healthcare-related studies, individual patient or hospital data are not often publicly available due to privacy restrictions, legal issues, or reporting norms. However, such measures may be provided at a higher or more aggregated level, such as state-level, county-level summaries or averages over health zones, such as hospital referral regions (HRR) or hospital service areas (HSA). Such levels constitute partitions over the underlying individual level data, which may not match the groupings that would have been obtained if one clustered the data based on individual-level attributes. Moreover, treating aggregated values as representatives for the individuals can result in the ecological fallacy. How can one run data mining procedures on such data where different variables are available at different levels of aggregation or granularity? In this article, we seek a better utilization of variably aggregated datasets, which are possibly assembled from different sources. We propose a novel cross-level imputation technique that models the generative process of such datasets using a Bayesian directed graphical model. The imputation is based on the underlying data distribution and is shown to be unbiased. This imputation can be further utilized in a subsequent predictive modeling, yielding improved accuracies. The experimental results using a simulated dataset and the Behavioral Risk Factor Surveillance System (BRFSS) dataset are provided to illustrate the generality and capabilities of the proposed framework.

Search plays an important role in online social networks as it provides an essential mechanism for discovering members and content on the network. Related search recommendation is one of several mechanisms used for improving members' search experience in finding relevant results to their queries. This paper describes the design, implementation, and deployment of Metaphor, the related search recommendation system on LinkedIn, a professional social networking site with over 175~million members worldwide. Metaphor builds on a number of signals and filters that capture several dimensions of relatedness across member search activity. The system, which has been in live operation for over a year, has gone through multiple iterations and evaluation cycles. This paper makes three contributions. First, we provide a discussion of a large-scale related search recommendation system. Second, we describe a mechanism for effectively combining several signals in building a unified dataset for related search recommendations. Third, we introduce a query length model for capturing bias in recommendation click behavior. We also discuss some of the practical concerns in deploying related search recommendations.