Sliding Information Distance for Robust Boundary Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing boundary detection methods in datasets are inefficient and produce noisy signals due to high computational costs and lack of robustness to small changes in information content, making it difficult to accurately identify boundary locations in a variety of data sources.
Innovation Solution
The Sliding Information Distance (SLID) method approximates the normalized information distance using a dictionary-based approach, optimized for sliding boundary detection, which employs a modified smoothed z-score algorithm to locate peaks in the distance metric, indicating boundary locations, and can be adapted for different window sizes and dimensions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard compression algorithms are used to approximate normalized information distance, then accuracy of boundary detection is improved, but computational cost increases making it too costly for sliding boundary detection applications
Solution Approach 1:
The patent segments the data sequence into sliding windows of fixed or adaptive size, computing information distance only between adjacent windows rather than all pairs. This segmentation reduces the computational burden from O(n²) to O(n) while maintaining boundary detection accuracy through localized comparison.
Solution Approach 2:
The patent pre-computes and stores the information distance metric for all positions in the sequence before performing boundary detection. This preliminary computation allows the actual boundary detection to proceed efficiently by simply querying pre-computed values, separating the expensive computation phase from the detection phase.
2Productivity
If conventional boundary detection methods are used, then computational speed is improved, but the signal becomes noisy and robustness to small changes in information content deteriorates
Solution Approach 1:
The patent combines multiple detection signals by summing the information distance values across different sliding window positions to create a composite boundary signal. This merging of multiple measurements smooths out noise and small fluctuations, making the boundary detection more robust while maintaining computational efficiency through vectorized operations.
Solution Approach 2:
The patent employs adaptive window sizes that change based on the local characteristics of the data sequence. By adjusting the window size parameter dynamically, the method maintains robustness to small changes in information content while adapting to different scales of boundaries, preventing false detections from minor fluctuations.
3Device complexity
If fixed window size is used for sliding windows, then computational simplicity is improved, but adaptability to different boundary scales and dimensions deteriorates
Solution Approach 1:
The patent transitions from fixed to adaptive window sizes that dynamically adjust based on the local data characteristics and detected boundary scales. The window size becomes a dynamic parameter that grows or shrinks according to the information content density, enabling the algorithm to adapt to different boundary scales while maintaining relatively simple implementation through iterative adjustment.
Solution Approach 2:
The patent designs the sliding window framework to be universally applicable across different data dimensions and types by implementing both fixed and adaptive modes. The same core algorithm can handle 1D sequences, 2D images, and higher-dimensional data structures, providing multi-functionality without requiring separate specialized methods for each case.
Data Source
AI summary
Fast, efficient, and robust compression-based methods for detecting boundaries in arbitrary datasets, including sequences (1D datasets), are desired. The methods, each employing three simple algorithms, approximate the information distance between two adjacent sliding windows within a dataset. One of the algorithms calculates an initial ordered list of subsequences; while a second algorithm updates the ordered list of subsequences by dropping a first entry and appending a last entry rather than calculating completely new ordered lists with each iteration. Large values in the distance metric are indicative of boundary locations. A smoothed z-score or a wavelet-based algorithm may then be used to locate peaks in the distance metric, thereby identifying boundary locations. An adaptive version of the method employs a collection of window sizes and corresponding weighting functions, making it more amenable to real datasets with unknown, complex, and changing structures.