🤖 machine learning

Learning Hyperspherical Time-Frequency Representations for Time-Series Out-of-Distribution Detection

This paper proposes a hyperspherical time-frequency representation learning framework using von Mises-Fisher objectives to improve out-of-distribution detection in time-series data, demonstrating consistent performance gains over strong baselines across the UCR and UEA archives.

Original authors: Willian T. Lunardi, Samridha Shrestha, Martin Andreoni

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Willian T. Lunardi, Samridha Shrestha, Martin Andreoni

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 a security guard at a very exclusive club. Your job is to let in only the people who belong (the "In-Distribution" or ID guests) and turn away anyone who doesn't fit the vibe (the "Out-of-Distribution" or OOD strangers).

In the world of computer science, most security guards are great at spotting strangers in photos (vision) or text (language). But when it comes to time-series data—which is just a stream of numbers changing over time, like a heartbeat monitor, stock prices, or a robot's movement—these guards often get confused. They might let a stranger in because they look similar on the surface, or they might kick out a real guest because they are wearing a slightly different hat.

This paper introduces a new security system called HyperTF designed specifically to catch these time-series strangers. Here is how it works, broken down into simple concepts:

1. The Problem: Why Current Guards Fail

Time-series data is tricky. A heartbeat from a healthy person in a hospital looks different from a heartbeat from a healthy person at home, even though both are "normal." Current systems often fail because they don't understand the deep structure of these signals. They treat the data like a flat list of numbers rather than a complex story with a beginning, middle, and end.

2. The Solution: A "Globe" Instead of a Flat Map

The authors decided to stop looking at the data on a flat map and instead put it on a globe (a hypersphere).

The Old Way: Imagine trying to sort different fruits on a long, flat table. Similar fruits might end up far apart if the table is messy.
The HyperTF Way: Imagine rolling all the fruits onto a giant globe. You group all the "apples" together in one spot on the globe, all the "oranges" in another, and so on.
The Magic: By forcing the data onto this globe, the computer learns that "apples" are always close to other "apples," no matter how the data was recorded. If a new piece of data lands in the middle of the ocean between the fruit groups, the system knows immediately: "This isn't an apple or an orange. This is a stranger!"

3. The "Two-Eyed" Approach

To make sure the security guard sees everything clearly, HyperTF looks at the data through two different lenses at the same time:

The Time Lens: It watches the signal as it happens second-by-second (like watching a movie).
The Frequency Lens: It listens to the "pitch" and "rhythm" of the signal (like looking at the musical notes in a song).

Think of it like listening to a song. The Time Lens hears the lyrics and the melody as they play out. The Frequency Lens hears the bass, the treble, and the specific instruments being used. By combining both views, the system gets a much richer understanding of what a "normal" signal looks like.

4. The Training: Learning with "Outlier" Friends

To get really good at spotting strangers, the system is trained with a special trick. The authors sometimes show the system a pile of "fake" or "strange" data (called auxiliary outliers) during training.

The Analogy: Imagine training a guard by showing them photos of people who don't belong at the club (like people in clown costumes or people wearing construction gear).
The Result: The guard learns exactly where the "club boundary" is. When a real stranger shows up later, the guard doesn't hesitate. The paper found that using a special mixing technique (called MixOE) to create even more diverse "stranger" examples made the guard even sharper.

5. The Results: A Better Guard

The authors tested their new system on a massive collection of time-series data (the UCR and UEA archives), which includes everything from heart monitors to robot movements.

The Scorecard: They compared HyperTF against the best existing security guards.
The Outcome: HyperTF consistently caught more strangers and let fewer strangers slip through the cracks. It was particularly good at spotting strangers who were very different from the normal guests (the "far-OOD" cases).
No Trade-offs: Usually, making a system better at spotting strangers makes it worse at recognizing real guests. But HyperTF managed to get better at spotting strangers without getting worse at recognizing the real guests.

Summary

In short, this paper built a smarter security guard for time-based data. Instead of just looking at the data in a straight line, it rolls the data onto a globe, looks at it through two different lenses (time and frequency), and trains it with a diverse mix of "strangers." The result is a system that is much better at knowing when something doesn't belong, which is crucial for safety-critical systems like medical monitors or autonomous machines.

Technical Summary: Learning Hyperspherical Time–Frequency Representations for Time-Series Out-of-Distribution Detection

Problem Statement
Out-of-distribution (OOD) detection for time-series data remains significantly underexplored compared to vision and language domains. Existing approaches, often adapted from image-based benchmarks, frequently fail to address the unique nature of distributional shifts in time-series data. Unlike exchangeable or spatially indexed data, time-series observations involve structured temporal dependencies, frequency-domain characteristics, and measurement processes tightly coupled to specific sensing hardware and acquisition protocols. Consequently, distribution shifts in time-series often reflect combined changes in the underlying process and the measurement pipeline, leading to systematic differences in temporal or spectral structure. Current evaluation protocols, which often rely on intra-dataset, label-disjoint splits, fail to capture these cross-environment shifts, blurring the distinction between open-set recognition and true OOD detection.

Methodology: HyperTF
The authors propose HyperTF, a hyperspherical representation learning framework designed to induce class-conditional structure on the unit sphere for robust OOD detection. The methodology integrates three core components:

Hyperspherical Embeddings: The framework models class-conditional features as lying on a unit sphere ( $S^{d-1}$ ), utilizing a von Mises–Fisher (vMF) likelihood-based objective. Under this model, the conditional distribution of a normalized feature vector $z$ given a class $c$ is defined by a prototype direction $\mu_c$ and a concentration parameter $\kappa$ . Classification is driven by angular similarity, and OOD detection arises naturally from the geometry of these embeddings, where OOD samples are expected to lie far from the learned class prototypes.
Dual-View Time–Frequency Architecture: To capture complementary aspects of time-series data, HyperTF employs a dual-stream architecture.
- Time-Domain View: The input sequence is projected and processed directly.
- Frequency-Domain View: The input is transformed via a real-valued Fast Fourier Transform (rFFT) followed by logarithmic magnitude scaling.
- Alignment: Both views are processed by domain-specific encoders (sharing architecture but with separate parameters) and mapped to a shared hyperspherical space via supervised projection heads. A cross-domain consistency regularizer ( $L_{CD}$ ) enforces angular alignment between the time and frequency embeddings, ensuring the model encodes both temporal dynamics and spectral characteristics within a unified geometric space.
Training Objective: The model minimizes a composite loss function:
$\mathcal{L} = \mathcal{L}_{TF} + \lambda_{CD} \mathcal{L}_{CD} + \lambda_{AX} \mathcal{L}_{AX}$
- $\mathcal{L}_{TF}$ : A supervised hyperspherical classification loss (cross-entropy on the sphere) applied independently to time and frequency branches.
- $\mathcal{L}_{CD}$ : A cross-domain consistency term minimizing the angular distance between time and frequency embeddings.
- $\mathcal{L}_{AX}$ : An optional auxiliary contrastive regularizer that leverages unlabeled auxiliary outlier data to shape the representation geometry near the in-distribution (ID) boundary, preventing collapse and enhancing separation.

Key Contributions

Hyperspherical Framework: The paper introduces a principled approach to time-series OOD detection that induces class-conditional structure on the unit sphere via a vMF formulation, moving beyond standard Euclidean embedding spaces.
Dual-View Architecture: A novel architecture that learns aligned representations from temporal and spectral views, coupled through a shared hyperspherical decision geometry and a cross-domain consistency regularizer.
Large-Scale Evaluation: The authors conduct extensive experiments on the complete UCR (univariate) and UEA (multivariate) time-series archives under a rigorous cross-dataset protocol. This protocol treats each dataset as ID in turn, with all others serving as OOD, distinguishing between "near-OOD" (same modality) and "far-OOD" (different modality) scenarios.

Experimental Results
Evaluated against strong baselines including post-hoc methods (MSP, Mahalanobis, Center Loss) and training-time methods (SimCLR, SSD+, KNN+, CIDER), HyperTF demonstrates consistent improvements:

Performance: HyperTF achieves state-of-the-art results in both near-OOD and far-OOD settings across the UCR and UEA archives. Notably, it significantly outperforms baselines under Mahalanobis++ (Maha++) scoring, achieving an FPR@95 of 1.90% on UCR near-OOD and 4.21% on UEA near-OOD (with auxiliary data).
ID Classification: The method improves OOD robustness without sacrificing in-distribution classification accuracy, achieving F1-scores of 82.53% (UCR) and 81.74% (UEA), surpassing the CIDER baseline.
Auxiliary Data: The inclusion of auxiliary outlier data, particularly when augmented with MixOE, further enhances detection performance in limited-coverage regimes, reducing FPR@95 by significant margins in far-OOD scenarios.
Ablation: Step-wise ablation studies confirm that the combination of hyperspherical loss, dual-view processing, and FFT consistency yields the strongest performance, with the hyperspherical formulation providing the most significant gain over standard Euclidean softmax baselines.

Significance and Claims
The paper claims that jointly modeling complementary temporal and spectral structures within a shared geometric embedding space provides a robust foundation for time-series OOD detection under realistic distribution shifts. The authors emphasize that their findings highlight the necessity of evaluating time-series OOD methods beyond simple intra-dataset label splits. Instead, they advocate for benchmark settings that reflect deployment-time variations across sensing conditions, domains, and data sources. The work suggests that hyperspherical representations, which naturally enforce angular separation and compactness, are particularly well-suited for handling the complex, multi-modal distributional shifts inherent in time-series data.