Abstract
In this paper, we build an organization of high-dimensional datasets that cannot be cleanly embedded into a low-dimensional representation due to missing entries and a subset of the features being irrelevant to modeling functions of interest. Our algorithm begins by defining coarse neighborhoods of the points and defining an expected empirical function value on these neighborhoods. We then generate new non-linear features with deep net representations tuned to model the approximate function, and re-organize the geometry of the points with respect to the new representation. Finally, the points are locally z-scored to create an intrinsic geometric organization which is independent of the parameters of the deep net, a geometry designed to assure smoothness with respect to the empirical function. We examine this approach on data from the Center for Medicare and Medicaid Services Hospital Quality Initiative, and generate an intrinsic low-dimensional organization of the hospitals that is smooth with respect to an expert driven function of quality.
Introduction
High-dimensional data, such as medical records or hospital performance metrics, often presents significant challenges for traditional unsupervised learning. These datasets frequently suffer from missing entries, irrelevant features that act as noise, and non-Euclidean geometries where standard distance metrics fail to capture meaningful similarities. Standard techniques like PCA or Diffusion Maps assume that Euclidean proximity implies similarity, but in sparse, noisy data, these neighborhoods are often spurious.
This research introduces a "bigeometric" organization—a method that simultaneously organizes features and data points to create a representation where an underlying function of interest (e.g., "quality") varies smoothly. By combining deep learning with expert-driven guidance, the proposed framework filters out noise and reveals intrinsic structures hidden in messy data.
Methodology: Expert Driven Functional Discovery
The proposed pipeline, termed Expert Driven Functional Discovery, consists of four primary stages:
- Initial Organization & Imputation: Using coupled partition trees, the method organizes both features and data points. This structure allows for the estimation of missing values by identifying similarities between how different "subjects" respond to similar "questions."
- Pseudopoints and Expert Input: To minimize human effort, the algorithm groups data into "pseudopoints" (cluster centroids). Experts provide a rough ranking for these representative groups rather than the entire dataset.
- Deep Learning with Back-Propagation of Rankings: A stacked autoencoder (SDAE) is trained with a dual-purpose loss function: reconstructing the input and predicting the expert's ranking. This forces the network to learn "meta-features" relevant to the goal, effectively filtering out irrelevant noise.
- Local Z-Scoring: To ensure a homogeneous map where distances are consistent across the manifold, a local Mahalanobis distance is applied. This local normalization makes the final embedding robust to variations in local density.
Case Study: CMS Hospital Quality Initiative
The framework was evaluated using the 2014 Center for Medicare and Medicaid Services (CMS) Hospital Quality Initiative dataset. This dataset includes approximately 100 measures for over 5,000 hospitals. The data is extremely sparse; only about 30% of hospitals reported at least 90% of the measures. Furthermore, different features often provide conflicting signals, making simple averaging an unreliable metric for quality.
Key Results and Insights
The bigeometric approach yielded several significant improvements over standard manifold learning techniques:
- Dimensionality Reduction: The intrinsic dimension of the hospital data was reduced from 14 (using standard Euclidean metrics) to 7 using the DNN-derived metric. This suggests the new metric captures a more concise representation of the underlying truth.
- Tighter Clustering and Smoothness: The DNN metric created significantly more tightly clustered neighborhoods. By injecting expert opinion into the training process, the model learned which features to prioritize (e.g., mortality rates over administrative metrics), leading to a more logical reclassification of hospitals.
- Stability: By ensembling results from multiple neural network runs with different random seeds, the authors achieved a stable, invariant distance metric that overcomes the inherent instability of individual deep nets.
Conclusion
The bigeometric organization of deep nets provides a powerful tool for navigating complex, high-dimensional datasets. By leveraging local normalization and expert-driven supervision, the method successfully identifies intrinsic structures that are smooth with respect to functional goals. While the approach is computationally intensive and dependent on the quality of initial expert input, it offers a more nuanced alternative to simple star ratings, potentially providing a better compass for health policy and performance evaluation.
Summary updated