Re-Ranking and Representations for Time Series Retrieval: A Comparative Study
Bionda Rozin, Daniel Carlos Guimarães Pedronette, Ricardo da Silva Torres · IEEE Access · 2025
Properly understanding the trends and patterns of multiple variables over time is important for decision-making in several applications, ranging from ecological monitoring based on vegetation index time series to healthy condition assessment based on physiological indicator temporal profiles. In this context, effective time series retrieval systems, which involve time series pattern representations and similarity ranking, are vital. This paper presents a comparative study that includes different techniques for time series representations and ranking. Special attention is given to the comparison of the retrieval effectiveness performance of re-ranking methods based on unsupervised distance learning for time series. Conducted experiments included the comparison of 10 distinct representations and four different re-ranking approaches on six diverse time series datasets. Experimental results demonstrate that the proper combination of representation and re-ranking methods can often enhance the overall quality of ranking results, leading to gains on mAP up to 31.78%. Also, the proposed approach is flexible, allowing the use of other feature extractors, distance measures, and re-ranking methods The results suggest that the choice of time series representation and re-ranking method depends on the specific application context.