Abstract
Traditional time-series analysis often struggles with data containing multiple asynchronous event markers across trials. This paper introduces a novel Gaussian Process (GP) framework that treats time relative to each marker as a separate input dimension. By enforcing causality and leveraging multidimensional covariance functions, the proposed model outperforms standard alignment and windowing techniques in diverse applications, including neuroscience and traffic analysis.
The Alignment Problem in Time-Series Data
In many scientific disciplines, researchers collect "trials" of data—repeated measurements of the same underlying process. However, these trials are rarely perfectly synchronized. For instance, in a reaching experiment, a subject might initiate movement at varying intervals after a stimulus.
Standard analysis typically involves aligning all trials to a single "start" signal. While this keeps the data clean at the beginning, the signal often becomes a "blurry mess" by the time the actual movement occurs. Conversely, aligning to the movement onset makes the initial phase appear disorganized. The common workaround, "clipping" small windows around events, is problematic as it discards temporal relationships and requires arbitrary parameter choices.
Proposed Methodology: Time-Marked Gaussian Processes
The authors propose a Time-Marked Gaussian Process that redefines the input space. Instead of a simple one-dimensional time input ($t$), the model uses a $K$-dimensional vector where each dimension represents the time elapsed since a specific marker $m_k$.
Key Innovations
- Multidimensional Input Space: The model learns a covariance function over relative time offsets, allowing the GP to simultaneously account for multiple events.
- Causal Warping: To prevent future events from influencing past predictions, the authors introduce a warping function $h(t) = \max(0, t)$. This ensures the signal remains stationary with respect to a marker until that marker occurs.
- Trial-Specific Noise: The framework incorporates temporally colored noise unique to each trial, accounting for individual fluctuations while sharing an underlying dynamical structure.
Experimental Evaluation
The model was evaluated across three distinct domains:
- Arm Reach Data: Modeling horizontal velocity during point-to-point reaching tasks.
- Gameday Traffic Data: Predicting automobile traffic counts near a stadium, using game start/end times as markers.
- Neural Firing Rate Data: Analyzing motor cortex electrophysiology during reaching tasks.
Key Results
Using Root Mean Squared Error (RMSE) via Leave-One-Out Cross-Validation, the Causal Time-Marked GP consistently achieved the highest accuracy. In neural data, it reduced RMSE from 20.8 (standard GP) to 15.7. In traffic prediction, it successfully captured both daily oscillations and irregular game-day spikes, reducing RMSE from 59.6 to 43.2.
Conclusion and Future Work
By projecting 1D time into an event-relative space, this framework allows for Automatic Relevance Determination (ARD) to identify which markers most strongly influence the signal. While the current model assumes known marker timings, future work aims to treat marker locations as latent parameters, potentially combining this approach with automated change-point detection. Although the $O(N^3)$ scaling of GPs remains a challenge, the foundational framework provides a robust alternative to traditional epoch-based analysis.
Summary updated