DISC: Domain-Integrated Similarity Cost for Time-Series Dataset

DISC. Domain-Integrated Similarity Cost for Time-Series Dataset. By Yen-Ku Liu and three others, from National Yang Ming Chiao Tung University, Dolby Laboratories, and one other institution. Published on the twenty-second of October. No code repository is linked to it.

With more time-series datasets and more foundation models, comparing datasets has become a central part of the workflow. Before adapting a model, you would like to know whether a source dataset will transfer to a target. Checking that by evaluating every candidate source against every target is exhaustive, so what is wanted instead is a similarity estimate that summarizes the relationships that matter for transfer.

Plenty of similarity metrics for time-series datasets already exist. The trouble is that they capture different relationships, and they vary in how well they line up with downstream usefulness from one task to another. The authors measure that gap directly, by correlating source-target similarity scores with downstream forecasting errors, in out-of-distribution settings and in transfer-learning settings.

Rather than choosing one metric by hand, the proposal is to learn how to combine several. Domain-Integrated Similarity Cost, or DISC, is a learnable aggregation framework. Different metrics see different differences, and integrating their complementary values lets DISC better separate the source-target differences that actually matter for downstream behavior.

DISC is fitted on calibration transfer observations: cases where transfer has already been observed. From those it learns to select, weight and combine similarity metrics, and that combination can differ for different downstream tasks. Once calibrated, it no longer needs any transfer run. It scores a new source-target pair using only that pair's similarity-metric values.

After calibration, DISC aligns more closely with downstream outcomes than any single-metric reference, on the split the authors report. From that result they draw two conclusions. Time-series dataset similarity is task-dependent. And a learned integration of metrics is a stronger proxy for downstream transfer than manually selecting one similarity metric.

The alternative being argued against is picking one similarity measure by hand and relying on it for every task. DISC treats the existing measures as raw material, and lets observed transfers decide, task by task, how much each one should count.