Position: Postdoctoral Fellow

Current Institution: Stanford University

Informative Projection Ensembles: Theory and Algorithms for Interpretable Models

Predictive systems designed to keep human operators in the loop rely on the existence of comprehensible classification models, which facilitate the decision process by presenting interpretable and intuitive views of the data. Often, the domain experts require that the test data be represented using a small number of the original features. This serves to validate the classification outcome by highlighting the similarities to relevant training data. We present Informative Projection Ensembles, a framework designed to extract compact and communicative models, fundamental to decision support systems. Informative Projection Ensembles alternatively use one of several compact submodels that ensure compliance with the stringent requirement on model size, while also attaining high performance through specialization. The decision is presented to the user together with a depiction of how the classification label was assigned. In addition to the complexity bounds for the ensembles, we present case studies of how our framework makes automatic classification transparent by revealing previously unknown patterns in biomedical data.

We provide strong statistical convergence guarantees for our ensembles, when used with specific classes of groupings and predictors quantifying the impact of various parameters to the complexity of the problem. Specifically, we apply compression-based bounds for analyzing the problem of simultaneously learning groups and predictors. Our results illustrate how the complexity of the ensembles scales with parameters of the underlying structure rather than the original dimension of the ambient space. We present a variety of techniques that construct Informative Projections, allowing a more precise recovery of patterns existing in data. The learning procedure is flexible, not only in terms of the hypothesis classes for the local models and for the selection function, but also in terms of model optimization. The current tools allow a trade-off between model fidelity and learning speed. Our experiments show that the methods we introduce can discover and leverage low-dimensional structure in data, if it exists, yielding models that are accurate, compact and interpretable.

One of our case studies targets osteoarthritis, a major chronic disease that we still cannot treat. This is partly due to lack of known biomarkers that can predict disease progression, which are actively sought by the orthopedics research community and the pharmaceutical industry. We analyze public data from the FNIH Osteoarthritis Progression Biomarkers project. Given 144 candidate biomarkers, for 300 patients, the task is to identify the most effective biomarkers indicating whether a patient would progress after four years. Our ensemble pinpoints a set of features based solely on measurements from the baseline visit. This means that we would be able to screen patients at one single time point and predict whether their joint health will worsen over the course of the following few years based on the volumes of particular regions in their meniscus and cartilage. This, in turn, has significant implications for osteoarthritis management and prevention in that the identified anatomical structures may be used as targets to test the effect of novel drugs in clinical trials.


Madalina Fiterau is a Postdoctoral Fellow in the Computer Science Department at Stanford University, working with Professors Chris Re and Scott Delp in the Mobilize Center. Madalina has obtained a PhD in Machine Learning from Carnegie Mellon University in September 2015, advised by Professor Artur Dubrawski. Madalina also holds a B. Eng. from the Politehnica University of Timisoara, Romania.
The focus of her PhD thesis, entitled “Discovering Compact and Informative Structures through Data Partitioning”, was on learning interpretable ensembles, with applicability ranging from image classification to a clinical alert prediction system. Madalina is currently expanding her research on interpretable models, in part by applying deep learning to obtain salient representations from biomedical “deep” data, including time series, text and images. The ultimate goal is to fuse these representations with structured biomedical data to form comprehensive models for clinical instability as well as medical conditions such as cerebral palsy, osteoarthritis, obesity and running injuries.

Madalina is the recipient of the GE Foundation Scholar Leader Award for Central and Eastern
Europe. Her paper, “Deep Neural Decision Forests”, has received the Marr Prize for Best Paper
at ICCV 2015. Also, her presentation entitled “Using expert review to calibrate semi-automated
adjudication of vital sign alerts in Step Down Units”, has won a Star Research Award at the
Annual Congress of the Society of Critical Care Medicine 2016. She has organized two editions of the Machine Learning for Clinical Data Analysis at the Neural Information Processing Systems Conference (NIPS), in 2013 and 2014. She has also published papers and is in the Program Committee for top conferences and journals (NIPS, ICML, AAAI, IJCAI, JBI). She currently holds a fellowship supported by the National Institute of Health (NIH).