Unassigned · 2026-10-15
TEX-MEX: Similarity-Guided Temporal Contrastive Learning for Time Series
Aditya Radhakrishnan, Maxwell A. Xu, Wanting Mao, Santosh Kumar, Harish Haresamudram, James M. Rehg
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
TEX-MEX is a time-series SSL contrastive objective that unifies structural (morphology) and temporal (time proximity) similarity. For each anchor window, it grades batch members as structurally similar/dissimilar in encoder representation space, then pulls/pushes the anchor toward/away from temporally shifted versions of those graded similar/dissimilar windows. It also contrasts temporal shifts against unshifted frames to separate temporal content.
Why it matters
The authors target a key issue in time-series contrastive learning: few augmentations reliably preserve meaning, so positives are often defined using disjoint similarity notions. TEX-MEX combines these notions in one objective while using encoder-based similarity (via cosine similarity) that improves during training, and they report strong cross-domain SSL results plus a small open foundation model for PPG.
Method
- Randomly sample a batch of windows; sample a shared temporal offset δ (with min/max constraints) and create shifted windows.
- Use encoder representations to compute cosine similarities between the anchor and unshifted batch members; convert to soft positive/negative weights.
- Train with graded soft contrastive terms that attract shifted representations of structurally-similar windows, repel shifted representations of dissimilar windows, and include a frame-separation term contrasting shifted vs unshifted representations.
Limitation
The authors suspect TEX-MEX underperforms on sleep because the MOODS pretraining dataset “contains no sleep recordings,” and note expanding MOODS to include sleep recordings as future work.
Abstract (from arXiv)
Contrastive learning for time series depends on deciding which pairs of inputs should be treated as similar and dissimilar. Because few augmentations reliably preserve meaning in time series, prior methods have relied on two disjoint notions of similarity between windows: structural similarity, which relates windows with similar morphology, and temporal similarity, which relates windows that occur close together in time. We introduce TEX-MEX, to our knowledge the first approach to unify the two into a single relationship. Given a random batch of windows drawn across the dataset, TEX-MEX uses the encoder to identify which windows are structurally similar or dissimilar to each other. It then pulls each window toward the temporally shifted versions of similar windows and pushes it away from those of dissimilar ones, learning what typically accompanies similar windows in time. Across PPG, EEG, ECG, accelerometry, and continuous glucose monitoring, TEX-MEX broadly achieves state-of-the-art performance against contrastive, reconstructive, JEPA-based, and modality-specific approaches, establishing it as a strong general SSL objective. We further train a 560K-parameter PPG foundation model that achieves state-of-the-art performance across a broad range of downstream benchmarks while being roughly 50x smaller than the next-best unimodally pretrained PPG foundation model in our evaluation. We will release this foundation model open-source and open-weight to the community to support future work and practical applications.
Related papers
- Similar Predictive Fit but Different Latent Dynamics: Characterizing Learned Dynamical Structure in Personalized Models of Brain Disorders
- LogSig-SSM: Time-Series Modelling with Multi-Scale Log-Signature Compression for State-Space Models
- KITE: A Knowledge-Injected Temporal Encoder for Time-Series Event Forecasting
- NeuroLens: Learning Latent Embeddings of Neural Semantics from Chronic Recordings
- Pythia: Toward Foundation World Models for Multimodal Time Series
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-14 · Indexed by Pub-AI: 2026-10-15 UTC
TEX-MEX: Similarity-Guided Temporal Contrastive Learning for Time Series
Preprint posted: 2026-10-14 · Original record ↗
- arxiv
- 2610.19557
Explicit source versions (1)
- v1 · 2026-10-14
- Preprint available
Source date: 2026-10-14 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Contrastive learning for time series depends on deciding which pairs of inputs should be treated as similar and dissimilar. Because few augmentations reliably preserve meaning in time series, prior methods have relied on two disjoint notions of similarity between windows: structural similarity, which relates windows with similar morphology, and temporal similarity, which relates windows that occur close together in time. We introduce TEX-MEX, to our knowledge the first approach to unify the two into a single relationship. Given a random batch of windows drawn across the dataset, TEX-MEX uses the encoder to identify which windows are structurally similar or dissimilar to each other. It then pulls each window toward the temporally shifted versions of similar windows and pushes it away from those of dissimilar ones, learning what typically accompanies similar windows in time. Across PPG, EEG, ECG, accelerometry, and continuous glucose monitoring, TEX-MEX broadly achieves state-of-the-art performance against contrastive, reconstructive, JEPA-based, and modality-specific approaches, establishing it as a strong general SSL objective. We further train a 560K-parameter PPG foundation model that achieves state-of-the-art performance across a broad range of downstream benchmarks while being roughly 50x smaller than the next-best unimodally pretrained PPG foundation model in our evaluation. We will release this foundation model open-source and open-weight to the community to support future work and practical applications.