Self-Supervised Recurrent Hashing for Efficient Time Series Similarity Search

The vast amount of time series data generated across various domains makes it difficult to search and analyze similar patterns. Traditional methods often struggle with capturing the sequential nature of time series data, leading to inefficiencies with large datasets and a reliance on extensive manual labeling. This paper proposes Self-Supervised Recurrent Hashing (SSRH), a novel approach that addresses these limitations. SSRH leverages Self-Supervised Learning (SSL) to extract meaningful representations from unlabeled time series data, reducing the need for manual labeling. Subsequently, a Recurrent Neural Network (RNN) captures the temporal dependencies within the data. Finally, a Recurrent Hashing Layer compresses the RNN output into compact codes (hashes) while preserving essential information for efficient similarity search. The proposed SSRH framework offers several advantages: it overcomes the limitations of traditional methods by capturing temporal information, enables efficient search and retrieval in large datasets, and reduces the need for extensive manual labeling through SSL. SSRH is evaluated on a publicly available time series dataset (UCR archive), demonstrating its effectiveness in retrieving similar time series compared to baseline methods. This research contributes to the field of time series analysis by offering a novel and efficient framework for similarity search using self-supervised learning and recurrent hashing techniques.