Abstract
We propose a similarity measure for sparsely sampled time course data in the form of a log-likelihood ratio of Gaussian processes (GP). The proposed GP similarity is similar to a Bayes factor and provides a robust way to cluster "messy" data where traditional measures like Euclidean distance often fail.
Introduction
In biological and economic research, time-course data is frequently "short and sparse." For instance, gene expression might be measured at only a handful of time points. Such data is often plagued by noise and irregular sampling intervals, making traditional similarity measures like Euclidean distance or Dynamic Time Warping (DTW) ineffective. Euclidean distance ignores temporal trends and requires synchronous sampling, while DTW can over-fit noise, creating artificial similarities.
Methodology: Gaussian Process Similarity
The proposed method models time series as continuous functions using Gaussian Processes (GPs). Instead of treating data as discrete vectors, a GP provides a probability distribution over possible functions that could have generated the observations. Similarity is calculated as a log-likelihood ratio between two hypotheses:
- Hypothesis "Same": Both series were generated by the same underlying function.
- Hypothesis "Different": The series were generated by two independent functions from the same GP.
A significant theoretical finding is that when noise is negligible, this GP similarity simplifies to the squared Euclidean distance, establishing it as a principled generalization for noisy, temporal data. It naturally incorporates uncertainty, preventing outliers from skewing the similarity score.
Experimental Evaluation
The measure was tested on synthetic datasets with varying noise and sampling intervals, as well as a real-world "omics" dataset involving cellular reprogramming. The real-world data consisted of 1,912 genes measured at five time points with two replicates each, representing a classic "short and sparse" scenario.
Results and Discussion
The GP-based measure demonstrated superior robustness to noise. In synthetic tests, it consistently achieved higher Normalized Mutual Information (NMI) scores than Euclidean distance, DTW, and Bregman divergence as noise levels increased from 0.08 to 0.12.
On biological data, the method yielded clusters with higher Biological Homogeneity Index (BHI) scores, indicating that the resulting groups were more functionally coherent according to Gene Ontology. The GP measure correctly identifies shared trends (e.g., "upward-then-downward") that match human visual intuition, even when Euclidean distances might suggest otherwise.
Conclusion
By treating time series as continuous processes and explicitly modeling uncertainty, the GP similarity measure provides a powerful tool for analyzing sparse and noisy data. While it is not a formal distance metric and has higher computational costs for asynchronous data, its ability to capture underlying trends makes it highly effective for complex biological datasets. Future work may involve scaling the method using sparse GP techniques to handle larger datasets efficiently.
Summary updated