The lernaean hydra of data series similarity search

Abstract

Increasingly large data series collections are becoming commonplace across many different domains and applications.A key operation in the analysis of data series collections is similarity search, which has attracted lots of attention and effort over the past two decades. Even though several relevant approaches have been proposed in the literature, none of the existing studies provides a detailed evaluation against the available alternatives. The lack of comparative results is further exacerbated by the non-standard use of terminology, which has led to confusion and misconceptions. In this paper, we provide definitions for the different flavors of similarity search that have been studied in the past, and present the first systematic experimental evaluation of the efficiency of data series similarity search techniques. Based on the experimental results, we describe the strengths and weaknesses of each approach and give recommendations for the best approach to use under typical use cases. Finally, by identifying the shortcomings of each method, our findings lay the ground for solid further developments in the field. PVLDB Reference Format:the whole (not a sub-) sequence. This problem represents a common use case across many domains [1, 2,38,29]. This work is the most extensive experimental comparison of the efficiency of similarity search methods ever conducted. Contributions. We make the following contributions:1. We present a thorough discussion of the data series similarity search problem, formally defining its different variations that have been studied in the literature under diverse and conflicting names. Thus, establishing a common language that will facilitate further work in this area.2. We include a brief survey of data series similarity search approaches, bringing together studies presented in different communities that have been treated in isolation from each other. These approaches range from smart serial scan methods to the use of indexing, and are based on a variety of classic and specialized data summarization techniques.3. We make sure that all approaches are evaluated under the same conditions, so as to guard against implementation bias. To this effect, we used implementations in C/C++ for all approaches, and reimplemented in C the ones that were only available in other programming languages. Moreover, we conducted a careful inspection of the code bases, and applied to all of them the same set of optimizations (e.g., with respect to memory management, Euclidean distance calculation, etc.), leading to considerably faster performance.4. We conduct the first comprehensive experimental evaluation of the efficiency of data series similarity search approaches, using several synthetic and 4 real datasets from diverse domains. In addition, we report the first large scale experiments with carefully crafted query workloads that include queries of varying difficulty, which can effectively stress-test all the approaches. Our results reveal characteristics that have not been reported in the literature, and lead...

Return of the Lernaean Hydra: Experimental Evaluation of Data Series Approximate Similarity Search

“…Approximate query answering with DSTree and iSAX2+ outperfom SRS and QALSH (state-of-the-art LSH-based methods) both in space and time, while supporting better theoretical guarantees. This surprising result opens up exciting research opportunities, that is, devising efficient disk-based techniques that support both ng-approximate and δ--approximate search with top performance [53]. Note that data series indexes developed for distributed platforms [157,162] also have the potential of outperforming LSH techniques [21,143] if extended following the ideas discussed in Section 2.…”

Approximate Similarity Joins over Dense Vector Embeddings

“…Among the mentioned algorithms, the HNSW stands out for offering significant improvements in search time with minimal reduction in recall [Echihabi et al 2019, Aumüller et al 2020. It has been widely used in various applications, ranging from realtime searches to data discovery in data lakes [Fan et al 2023].…”