📊 statistics

Semi-parametric Functional Classification via Path Signatures Logistic Regression with Adaptive Order Selection

This paper introduces Path Signatures Logistic Regression (PSLR), a semi-parametric framework that combines interpretable linear effects with a data-driven, theoretically guaranteed adaptive order selection mechanism to achieve robust and accurate classification of irregularly sampled vector-valued functional data.

Original authors: Pengcheng Zeng, Siyuan Jiang

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pengcheng Zeng, Siyuan Jiang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to tell the difference between two types of people walking down a street. One group has a specific medical condition (like Parkinson's), and the other is healthy.

To do this, you have two types of information:

The Walk (Functional Data): A complex, multi-dimensional recording of how their feet hit the ground over time. This data is messy; sometimes the sensors miss a step, sometimes the steps are uneven, and the data comes from multiple sensors at once.
The Person (Scalar Data): Simple facts like their age, height, or walking speed.

The Old Way: Trying to Fit a Square Peg in a Round Hole

Traditionally, statisticians tried to solve this by forcing the messy walking data into neat, pre-made boxes (called "basis expansions," like Fourier series or splines).

The Problem: It's like trying to describe a swirling, chaotic dance by forcing it into a rigid grid. If the dancer moves irregularly (which real people do), the grid fails. The model gets confused, makes mistakes, and struggles to understand how the left foot and right foot interact with each other. It also assumes the relationship between the walk and the disease is a straight line, which is rarely true in biology.

The New Way: Path Signatures Logistic Regression (PSLR)

The authors of this paper propose a new tool called PSLR. Think of it as a "geometric translator" that doesn't care about grids or neat boxes.

1. The "Signature" (The Fingerprint of the Path)
Instead of forcing the walk into a grid, PSLR looks at the shape of the walk itself. It treats the walking path as a continuous line drawn in space.

The Analogy: Imagine tracing a path with your finger. A "signature" captures not just where you started and ended, but the twists, turns, loops, and how your finger moved relative to itself. It captures the "personality" of the movement.
The Magic: This method is basis-free. It doesn't need to guess which "box" to put the data in. It naturally understands that the left foot and right foot are connected (cross-channel dependency) and handles missing steps or uneven timing without getting confused.

2. The "Semi-Parametric" Mix
The model is smart enough to know that some things are simple and some are complex.

Scalar Covariates (The Simple Stuff): For facts like "Age" or "Speed," it uses a simple, straight-line logic (e.g., "Older people are slightly more likely to have the condition"). This keeps the results easy to understand.
The Walk (The Complex Stuff): For the messy walking data, it uses the complex "Signature" translator to find hidden, non-linear patterns that simple lines would miss.

3. The "Adaptive Order" (The Goldilocks Rule)
Here is the biggest innovation. The "Signature" can be calculated at different levels of detail (called "orders").

Low Order: Only looks at the big picture (start and end points). Too simple? It misses the details.
High Order: Looks at every tiny wiggle. Too complex? It starts memorizing the noise (like a student memorizing a specific test question instead of learning the concept).
The Innovation: Instead of a human guessing the right level of detail, this paper introduces a self-correcting rule. The computer tries different levels of detail, checks how well it predicts the data, and automatically picks the "Goldilocks" level—not too simple, not too complex, but just right. It does this mathematically, ensuring it doesn't overthink the data.

Why This Matters (The Results)

The authors tested this on both fake data and real-world medical data (Parkinson's gait analysis and smartphone motion sensors).

Robustness: When the data was messy (missing steps or uneven timing), the old methods (the rigid grids) fell apart. PSLR kept working because it looks at the shape of the path, not the specific timing of the points.
Accuracy: PSLR consistently beat the old methods and even beat a version of itself that only looked at the walk (ignoring age/speed) or only looked at the age/speed (ignoring the walk). It proved that combining the "shape" of the movement with simple facts gives the best result.
Interpretability: The model didn't just give a "yes/no" answer; it told the researchers why. For example, it identified that specific interactions between the left and right foot (captured by the signature) were strong indicators of Parkinson's, matching what doctors know about the disease.

In a Nutshell

This paper presents a new way to classify complex, messy movement data. Instead of forcing the data into rigid, pre-made boxes, it uses a flexible mathematical tool (Path Signatures) to read the "shape" of the movement. It automatically figures out how much detail is needed to be accurate without getting confused, and it successfully combines this complex movement data with simple facts like age to make better predictions than ever before.

Technical Summary: Semi-parametric Functional Classification via Path Signatures Logistic Regression with Adaptive Order Selection

Problem Statement

The paper addresses the challenge of classifying vector-valued functional data (e.g., multi-dimensional time-series trajectories) in the presence of scalar covariates. Traditional functional logistic regression models rely on linear assumptions and fixed basis expansions (e.g., Fourier, B-splines, FPCA). These classical approaches face three primary limitations:

Linearity Assumption: They assume a linear relationship between functional predictors and the log-odds of the response, which may fail to capture complex nonlinear dynamics.
Basis Sensitivity: Performance is highly dependent on the choice of basis functions and knot placement, leading to potential model misspecification and obscured interpretability.
Fragility to Irregular Sampling: Basis projections assume complete, uniformly sampled trajectories. Irregular sampling or missing data introduces representation bias and requires interpolation that can lead to over-smoothing or structural bias.

Furthermore, existing methods often neglect cross-channel correlations or require perfect temporal alignment, which is impractical in real-world scenarios like gait analysis or sensor-based activity recognition.

Methodology: Path Signatures Logistic Regression (PSLR)

The authors propose Path Signatures Logistic Regression (PSLR), a semi-parametric framework that replaces the linear functional term with a nonlinear transformation based on path signatures.

Model Formulation

The model assumes a semi-parametric additive structure for the conditional log-odds:
$\text{Logit}(P(y=1|X, z)) = F(X) + z^\top \gamma$
where $X$ is a $d$ -dimensional functional predictor, $z$ is a vector of scalar covariates, and $F$ is a smooth, nonparametric functional mapping.

To approximate $F(X)$ , the method treats the time-augmented signal $\tilde{X} = (X, t)$ as a continuous path. Leveraging the universality property of path signatures (Stone–Weierstrass-type theorem), $F(X)$ is approximated by a linear form of the truncated signature:
$F(X) \approx S_p(\tilde{X})^\top \beta_p$
where $S_p(\tilde{X})$ is the truncated path signature of order $p$ , and $\beta_p$ is a coefficient vector. The full model becomes a generalized linear model:
$\text{Logit}(P(y=1|X, z)) = \tilde{S}_p^\top \theta_p$
where $\tilde{S}_p$ concatenates the signature features and scalar covariates.

Key Innovations

Basis-Free Representation: Signatures capture geometric and temporal features (displacement, curvature, signed area) without requiring basis selection.
Robustness to Irregularity: The signature transform is stable under small perturbations of the path. The method uses local piecewise linear interpolation, which theoretically vanishes in error as the sampling grid refines, unlike global basis projections that suffer from irreducible structural bias.
Adaptive Order Selection: A critical contribution is a fully data-driven procedure to select the truncation order $p$ . Instead of heuristically fixing $p$ , the authors propose minimizing a penalized empirical risk criterion:
$\hat{p} = \arg\min_{p} \left( \hat{L}_n(p) + \text{penn}(p, q) \right)$
where $\hat{L}_n(p)$ is the regularized empirical risk and the penalty term $\text{penn}(p, q)$ scales with the complexity of the signature features ( $\sqrt{s_d(p)}$ ) and the dimension of scalar covariates ( $\sqrt{q}$ ).

Theoretical Contributions

The paper provides rigorous non-asymptotic guarantees for the proposed framework:

Existence of Optimal Order: It proves the existence of an optimal truncation order $p^*$ and coefficients $\theta^*_{p^*}$ that approximate the population risk arbitrarily well.
Consistency: The data-driven estimator $\hat{p}$ is shown to be consistent, converging to the true optimal order $p^*$ with exponential probability as the sample size increases.
Risk Convergence: The classifier risk converges at the standard parametric rate of $O(n^{-1/2})$ , provided the correct order is selected.
Finite Search Bound: The authors derive a computable upper bound $P$ for the search range of $p$ , ensuring the optimization is finite and practical ( $P = O(\log n)$ ).
Error Propagation: A framework is established showing that PSLR's reconstruction error vanishes with grid refinement, contrasting with the irreducible bias of basis-expansion methods under irregular sampling.

Experimental Results

The method was evaluated on synthetic data and three real-world datasets:

Synthetic Data: PSLR consistently outperformed classical functional classifiers (B-spline, Fourier, FPCA) and ablated versions (Signature-only, Scalar-only) across varying dimensions, scalar covariate counts, and irregular sampling scenarios (missing data and uneven time grids).
Parkinson's Disease Gait Analysis: Using Vertical Ground Reaction Force (VGRF) data, PSLR achieved superior accuracy and F1 scores compared to baselines. The model successfully identified interpretable coefficients, such as the negative impact of reduced variability in specific foot regions on PD probability, aligning with clinical observations of rigid gait.
Human Activity Recognition (MotionSense): In classifying walking vs. jogging using smartphone sensor data, PSLR demonstrated higher accuracy and stability than baselines, effectively leveraging both functional motion signals and scalar attributes (e.g., weight).
Multi-Cohort Clinical Gait: The method was applied to distinguish neurological patients from healthy controls using wearable sensor data, further validating its robustness in complex, high-dimensional settings.

Significance and Claims

The paper claims that PSLR offers a practical and theoretical advancement in functional data analysis by:

Integrating Rough Path Theory with Adaptive Complexity: It bridges the gap between the theoretical power of path signatures and practical model selection, moving beyond fixed-order heuristics.
Handling Real-World Data Imperfections: By avoiding basis expansions, the method inherently handles irregular sampling and inter-channel dependencies without requiring temporal alignment or complex imputation.
Interpretability: The framework maintains interpretability for scalar covariates while providing a geometric interpretation of functional effects through signature terms (e.g., displacement, area, interactions).
Theoretical Rigor: It fills a critical gap in the literature by providing the first non-asymptotic guarantees for adaptive order selection in semi-parametric signature-based classification, ensuring that model complexity grows only as necessary to capture intrinsic data structure.

The authors emphasize that while the signature transform itself is known, its integration with adaptive complexity control and scalar covariates within a unified semi-parametric logistic regression framework is a novel contribution supported by rigorous finite-sample guarantees.