Abstract
We propose a model for functional data registration that compares favorably to the best methods of functional data registration currently available. It also extends current inferential capabilities for unregistered data by providing a flexible probabilistic framework that 1) allows for functional prediction in the context of registration and 2) can be adapted to include smoothing and registration in one model. The proposed inferential framework is a Bayesian hierarchical model where the registered functions are modeled as Gaussian processes. To address the computational demands of inference in high-dimensional Bayesian models, we propose an adapted form of the variational Bayes algorithm for approximate inference that performs similarly to MCMC sampling methods for well-defined problems. The efficiency of the adapted variational Bayes (AVB) algorithm allows variability in a predicted registered, warping, and unregistered function to be depicted separately via bootstrapping. Temperature data related to the el-niño phenomenon is used to demonstrate the unique inferential capabilities for prediction provided by this model.
Introduction
In Functional Data Analysis (FDA), researchers analyze data that takes the form of continuous curves, such as growth charts or weather patterns. A primary challenge in this field is phase variability: the phenomenon where similar features or "events" (e.g., a growth spurt or a temperature peak) occur at different times across different subjects. To analyze these curves effectively, researchers use "registration" (or alignment) to warp the time axis of each curve so that their features align.
Historically, registration has been treated as a separate pre-processing step. However, this decoupling introduces two significant issues. First, it leads to underestimated uncertainty; by smoothing and aligning data before the primary analysis, researchers ignore the errors introduced during those steps, resulting in overconfident conclusions. Second, it creates a prediction gap, making it difficult to forecast the trajectory of a partially observed curve because its specific "internal clock" (whether it is running early or late) remains unknown. This paper introduces a unified Bayesian hierarchical framework that simultaneously performs smoothing, registration, and prediction.
Methodological Framework
Bayesian Hierarchical Modeling with Gaussian Processes
The proposed model treats both the "true" registered functions and the "warping" functions (the internal clocks) as Gaussian Processes (GPs). This probabilistic approach allows the model to distinguish between amplitude variability (how high or low a curve is) and phase variability (when features occur).
To ensure physical realism, the warping functions must be monotonic—time must only move forward. The authors achieve this by defining the warping functions such that their log-derivatives are GPs. This framework allows the model to handle raw, noisy data directly, integrating smoothing and registration into a single inferential step.
Adapted Variational Bayes (AVB)
High-dimensional Bayesian models typically rely on Markov Chain Monte Carlo (MCMC) sampling for inference. While accurate, MCMC is often computationally prohibitive for complex functional data. Conversely, standard Variational Bayes (VB) is fast but requires "conjugate" priors that are often too restrictive for sophisticated FDA models.
To bridge this gap, the authors developed Adapted Variational Bayes (AVB). This algorithm alternates between a maximization step for non-conjugate parameters and standard VB updates for the remainder of the model. AVB provides MCMC-like accuracy with a fraction of the computational burden, enabling the use of bootstrapping to separately depict variability in registered, warping, and unregistered functions.
Experimental Results
Benchmarking and Efficiency
The model was evaluated using simulated data and the classic Berkeley Boys Growth Velocity dataset (39 subjects). In alignment benchmarks using the Sobolev Least Squares (sls) criterion, the GP-based model outperformed traditional Minimum Eigenvalue methods and performed comparably to the state-of-the-art Fisher-Rao Riemannian metric.
In terms of efficiency, the AVB algorithm demonstrated massive speedups. For the Berkeley Growth data, AVB reached convergence in approximately 10 hours, whereas a comparable MCMC chain required over 75 hours to reach a similar state, and would have required significantly longer to fully characterize the posterior distribution.
Prediction: The El Niño Case Study
The model's unique predictive capabilities were demonstrated using El Niño Sea Surface Temperature data (1950–2004). The task was to predict the remaining five months of a year given only the first seven months. Because the framework estimates the specific warping (the "timing") for a test year, it can align partial data to a global template far more accurately than non-registered methods. This allows for a more nuanced understanding of whether a particular year is following a standard El Niño pattern but on a delayed or accelerated schedule.
Discussion and Limitations
While the AVB framework offers significant advantages, the authors note several areas for further refinement:
- Local Maxima: Like many optimization-based approaches, AVB can occasionally get stuck in local modes. The authors suggest using simple initial warpings or gradually decreasing penalties to guide the algorithm toward the global optimum.
- Identifiability: A "confounding" effect exists where a vertical shift in a curve can look like a horizontal shift in time. Currently, certain registration parameters require manual guidance or short preliminary AVB runs to ensure stability.
- Prior Selection: The study found that "uninformative" Inverse-Gamma priors could sometimes bias noise variance estimates upward, suggesting that uniform priors on the square root of variance components might be more robust in future applications.
Conclusion
The Adapted Variational Bayes framework represents a significant step forward in Functional Data Analysis. By unifying smoothing, registration, and prediction within a computationally efficient Bayesian hierarchy, it provides a rigorous account of uncertainty that traditional pre-processing methods lack. This approach is particularly valuable for complex longitudinal data where understanding the timing of events is as critical as understanding their magnitude.
Summary updated