Sliding Window LSTM for Shifting Pattern Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification methods face challenges in accurately identifying discriminative patterns in multi-variate sequential data due to shifting patterns and noise, especially when classifying land cover types or affective states, as they often treat all time steps equally and fail to detect informative periods effectively.
Innovation Solution
A method integrating a sliding time window with Long Short-Term Memory (LSTM) to capture temporal dependencies and a Multi-Instance Learning (MIL) approach to aggregate latent outputs from multiple time windows, identifying discriminative patterns and mitigating noise, while incorporating context information to improve classification performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all time steps are treated equally in classification, then the method is simple and fast, but classification accuracy deteriorates due to inability to detect shifting patterns and informative periods
Solution Approach 1:
The patent divides the temporal sequence into multiple time windows using a sliding window approach. Each time window is processed independently to generate latent outputs, which are then aggregated to produce the final classification. This segmentation allows the system to focus on specific informative periods rather than treating all time steps uniformly, thereby improving classification accuracy while maintaining manageable computational complexity through efficient aggregation mechanisms.
2Reliability
If the entire time series is processed uniformly, then computation is efficient, but noise and irrelevant information reduce classification reliability
Solution Approach 1:
The patent extracts and processes only the most informative time windows through the sliding window approach. By generating latent outputs for each time window and aggregating them, the system effectively filters out noise and irrelevant information from the entire time series. This extraction mechanism improves classification reliability by focusing computational resources on discriminative periods rather than processing all data uniformly, thus reducing unnecessary computational energy expenditure.
3Measurement precision
If shifting patterns are not accounted for, then the classification method is straightforward, but performance deteriorates when discriminative patterns appear at different time periods
Solution Approach 1:
The patent implements a dynamic sliding window approach that moves through the time series sequentially, allowing the system to adapt to shifting patterns automatically. As the window slides through different time periods, it captures discriminative patterns wherever they occur, making the system versatile and adaptable to temporal variations in data patterns. This dynamic processing enables accurate detection of shifting patterns without requiring manual configuration or complex adaptive algorithms.
Data Source
AI summary
A method includes receiving data for an entity for each of a plurality of time points. For each of a plurality of time windows that each comprises a respective plurality of time points, a confidence value is determined. The confidence value provides an indication of the degree to which the time window contains data that is useful in discriminating between classes. The confidence values are used to determine a probability of a class and the probability of the class is used to set a predicted class for the entity.