Deep Sub-Linear Hashing Network for Time Series Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in efficiently retrieving similar multivariate time series segments from large datasets, as existing methods face issues with memory usage and response time due to the need for extensive disk access, and current binary code representation methods do not effectively capture global decision boundaries, leading to inefficient similarity searches.
Innovation Solution
A Deep Sub-Linear Hashing Network (DSLHN) is employed, which uses Long Short-Term Memory (LSTM) units to extract features and generates both long and short binary codes, allowing for sub-linear searches by leveraging triplet loss and cross-entropy loss to preserve relative similarity relations and capture global decision boundaries, thereby reducing memory usage and improving search efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If binary code representation methods are used for time series segments, then memory usage is reduced and search efficiency is improved, but global decision boundaries are not effectively captured leading to inaccurate similarity searches
Solution Approach 1:
The patent segments the binary code representation into two distinct components: long binary codes that capture global decision boundaries and preserve semantic meaning, and short binary codes that enable efficient sub-linear search operations. This segmentation allows the system to maintain both search efficiency and similarity search accuracy by using each code type for its strengths.
Solution Approach 2:
The patent transitions from traditional single-dimension binary hashing to a multi-dimensional representation system where long and short binary codes operate in different dimensional spaces. The long codes provide semantic accuracy while short codes provide search efficiency, and the system integrates both dimensions to achieve superior performance.
2Reliability
If extensive disk access is used for retrieving time series segments from large datasets, then retrieval completeness is improved, but response time and memory usage increase significantly
Solution Approach 1:
The patent performs preliminary action by pre-computing and storing both long and short binary codes for all time series segments in the database during an offline phase. This allows the online retrieval process to work with pre-processed binary representations rather than raw data, significantly reducing response time while maintaining retrieval completeness through the use of long codes for accurate similarity measurement.
Solution Approach 2:
The patent extracts essential features from time series segments and represents them in compact binary form, separating the critical similarity-determining characteristics from the full raw data. This extraction enables efficient storage and retrieval operations while preserving the ability to perform accurate similarity searches.
Data Source
AI summary
Systems and methods for retrieving similar multivariate time series segments are provided. The systems and methods include extracting a long feature vector and a short feature vector from a time series segment, converting the long feature vector into a long binary code, and converting the short feature vector into a short binary code. The systems and methods further include obtaining a subset of long binary codes from a binary dictionary storing dictionary long codes based on the short binary codes, and calculating similarity measure for each pair of the long feature vector with each dictionary long code. The systems and methods further include identifying a predetermined number of dictionary long codes having the similarity measures indicting a closest relationship between the long binary codes and dictionary long codes, and retrieving a predetermined number of time series segments associated with the predetermined number of dictionary long codes.