Joint Registration and Conformal Prediction for Partially Observed Functional Data

Fangyi Wang, Sebastian Kurtek, Yuan Zhang · 2025

Read Joint Registration and Conformal Prediction for Partially Observed Functional Data by Fangyi Wang, Sebastian Kurtek, Yuan Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-15000.txt.

Introduction

Fangyi Wang, Sebastian Kurtek, and Yuan Zhang tackle prediction for partly seen curves in a 2025 paper. Such curves arise when a measure is tracked over time but only in pieces. Two kinds of variation make the task hard: shifts up and down, and shifts in timing along the axis. Most older methods ignore the timing shifts.

The authors join two steps that are often split: aligning the curves, and predicting the missing parts. They work inside conformal prediction, a frame that gives bands with firm coverage claims under weak assumptions. A smoothing step over near neighbors builds the bands point by point. The method is simple to run and can be split across machines. Tests on made-up and real data show the bands work well in practice.

Catalog record

CREATORFangyi Wang, Sebastian Kurtek, Yuan Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-15000

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Joint Registration and Conformal Prediction for Partially Observed Functional Data Fangyi Wang, Sebastian Kurtek, Yuan Zhang Abstract Predicting missing segments in partially observed functions is challenging due to infinite-dimensionality, complex dependence within and across observations, and irregular noise. These challenges are further exacerbated by the existence of two distinct sources of variation in functional data, termed amplitude (variation along the y y -axis) and phase (variation along the x x -axis). While registration can disentangle them from complete functional data, the process is more difficult for partial observations. Thus, existing methods for functional data prediction often treat phase variation as negligible. Furthermore, they typically require precise model specifications and/or rely on computationally intensive tools, such as bootstrapping, to construct prediction intervals. We propose a unified registration and prediction approach for partially observed functions using conformal prediction. Our method integrates registration and prediction in one algorithm while ensuring exchangeability through carefully constructed predictor-response pairs. Using a neighborhood smoothing algorithm, the framework produces pointwise prediction bands with finite-sample marginal coverage guarantees under weak assumptions. The method is easy to implement, computationally efficient, and permits simple parallelization. Numerical studies and real-world data examples demonstrate the effectiveness and practical utility of our method. Keywords: conformal prediction, elastic functional data analysis, neighborhood smoothing, phase variation 1 Introduction Accurate prediction of future trajectories given historical functional data is an important question in many applications. For instance, given previous (complete) daily observations of traffic flow rate, and a partial observation (up to some time) for a new day, prediction of traffic flow rate for the rest of that day can help optimize transportation networks and reduce congestion during rush hour ( Chiou, 2012 ) . As another example, given complete historical daily maximum temperatures at a particular location, one may be interested in a daily maximum temperature forecast for the rest of the current year, which can help mitigate societal risks related to extremely cold weather. See Section 6 for more detailed analyses. Other examples include growth rate prediction for people or other natural objects, forecasting of pollutant density, and so on. These predictions can be used to anticipate future trends and help inform biomedical, environmental and social decisions. Functional data is challenging to analyze, because it is inherently infinite-dimensional, exhibits complex dependencies within and potentially across observations, and is often observed at discrete time points with heteroscedastic noise. Functional observations encompass two sources of variation, one along the x -axis called phase and one along the y -axis called amplitude . A common practice is to first separate amplitude and phase variations. This is done via a process called registration , which “warps” the functions’ domains to align their key features with respect to time ( Kneip and Gasser, 1992 ; Ramsay and Silverman, 2005 ; Srivastava et al., 2011 ) . However, registration is often treated as a pre-processing step, with subsequent statistical modeling and inference applied to “well-aligned” data. Meanwhile, registration uncertainty is often ignored in downstream tasks. Furthermore, when the data contains partial observations, registration is much more challenging since the domains of partial and complete functions are fundamentally different ( Bauer et al., 2021 ; Bryner and Srivastava, 2021 ) . To tackle the aforementioned challenges, we propose a framework for prediction of partially observed functions wherein registration is directly incorporated. Comparing to a sequential approach of “registration then prediction”, our approach bypasses the need for registration of partial observations to complete ones, and propagates registration uncertainty to the prediction step. We adapt conformal prediction , a distribution-free uncertainty quantification method that provides valid finite-sample coverage validity ( Vovk et al., 2005 ; Lei et al., 2018 ) . This requires a careful construction of exchangeable predictor-response data pairs, which are not naturally defined in our context. In addition, we employ a nonparametric, neighborhood smoothing prediction algorithm ( Zhang et al., 2017 ) , which offers flexibility without strong modeling assumptions on the relationship between predictors and responses. Thus, our main contributions are as follows. 1. We propose a novel framework that integrates registration into conformal prediction of the amplitude component for partial functional data. This results in more accurate prediction intervals, as compared to procedures that do not utilize registration, when functional data contains phase variation. We additionally define a separate conformal prediction procedure for the phase component, which incorporates a monotonicity constraint. Our approach builds on the elastic functional data analysis (EFDA) framework for registration ( Srivastava et al., 2011 ; Srivastava and Klassen, 2016 ) . 2. Our method offers distribution-free uncertainty quantification with a finite-sample guarantee. For conformal prediction of partially observed functions, we construct predictors and responses from raw data that maintain exchangeability throughout, and leverage ideas from split conformal methods to integrate registration. Unlike traditional parametric approaches, our method imposes minimal assumptions on the data generating process. We employ a nonparametric neighborhood smoothing prediction algorithm that provides more flexibility than competitors such as functional linear regression. 3. Compared to nonparametric bootstrap or Bayesian uncertainty quantification, our approach is much faster, making it well-suited for big data. Related work. There exist many approaches for functional data prediction, but most of them do not consider the challenges posed by phase variation. They include functional regression ( Chiou, 2012 ; Maity, 2017 ) , time series models ( Leroux et al., 2018 ; Jiao et al., 2023 ) and neural networks ( Yin et al., 2021 ) . For uncertainty quantification, these approaches generally use closed-form confidence intervals based on variance estimates ( Leroux et al., 2018 ) , or nonparametric bootstrap residual sampling ( Chiou, 2012 ; Jiao et al., 2023 ) . However, these methods require strong parametric assumptions, and the resulting prediction intervals’ coverage rate converges to the desired nominal level asymptotically with no finite-sample guarantee. Matuk et al. (2022) and Earls and Hooker (2017) define Bayesian models for joint registration and prediction, where prediction intervals are constructed from posterior samples. Nonetheless, the high computational cost of these approaches hinders their application to large datasets. Conformal prediction provides prediction intervals that have a finite-sample coverage guarantee without imposing assumptions on the data generating process ( Vovk et al., 2005 ) . It is generally computationally more efficient when compared to Bayesian methods, which utilize Markov chain Monte Carlo to sample from the posterior distribution. Since being introduced, conformal prediction has been widely applied in a variety of scenarios, ranging from regression and classification for multivariate data ( Lei et al., 2018 ; Cauchois et al., 2021 ) to prediction tasks involving more complex data structures, e.g., survival analysis ( Candès et al., 2023 ; Gui et al., 2024 ) , time series ( Zaffran et al., 2022 ; Angelopoulos et al., 2023 ) , and matrices ( Gui et al., 2023 ; Shao and Zhang, 2023 ) . However, relatively few works have focused on the application of conformal prediction to functional data. Lei et al. (2015) used a basis projection for functional observations to obtain prediction sets for basis coefficients; they did not consider partially observed functions. Diquigiovanni et al. (2021) ; Diquigiovanni et al. (2022) and Ajroldi et al. (2023) applied conformal prediction to functional data and functional time series using a domain-adaptive nonconformity

Read the complete text