FEATS Model Temporal Feature Impact Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing predictive data analysis solutions lack interpretability and effectively preserve the multi-variate temporal structure of data, leading to difficulties in handling time-dependent and cross-variable relationships in predictive models.
Innovation Solution
The FEATS model employs feature attention heads to generate per-temporal feature time impact scores and attention head scores, preserving the multi-variate temporal structure by processing each temporal feature set over a customized time window without concatenation, and considers both time-dependent and time-independent features through transformation functions, allowing for parallel processing and reduced computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional predictive models are used, then predictive accuracy can be achieved, but interpretability and preservation of multi-variate temporal structure are lost
Solution Approach 1:
The patent segments the temporal data into multiple time windows and processes each window separately through dedicated attention heads. This segmentation allows the model to preserve multi-variate temporal structure by analyzing specific time periods independently while maintaining overall predictive accuracy through aggregation of window-level predictions.
Solution Approach 2:
The patent introduces a temporal dimension by processing data through multiple time windows and assigning temporal weights to different time points. This dimensional approach enables the model to capture time-dependent patterns and cross-variable relationships while maintaining interpretability through temporal feature importance scores.
2Loss of information
If complex temporal processing is applied to preserve multi-variate temporal structure, then interpretability improves, but computational complexity increases
Solution Approach 1:
By dividing the temporal data into discrete time windows and processing each through separate attention heads, the model reduces computational complexity compared to processing the entire temporal sequence simultaneously. Each attention head operates on a specific time window, enabling parallel processing and reducing overall computational burden.
Solution Approach 2:
The patent applies attention mechanisms selectively to specific time windows and variables rather than processing all data uniformly. This partial action approach allows the model to focus computational resources on the most relevant temporal patterns and variables, reducing overall computational complexity while maintaining temporal structure preservation.
3Productivity
If all temporal features are processed equally, then computational efficiency is maintained, but time-dependent relationships and cross-variable interactions are not captured
Solution Approach 1:
The patent assigns different weights and processing intensities to different time windows and variables based on their local importance. Attention heads dynamically adjust the processing focus for each time window, allowing the model to capture time-dependent relationships and cross-variable interactions in regions where they are most significant while maintaining computational efficiency through selective processing.
Solution Approach 2:
The model dynamically adjusts the temporal weightings and attention distributions based on the input data characteristics and time window content. This dynamic adaptation enables the system to capture evolving time-dependent relationships and cross-variable interactions while maintaining computational efficiency through adaptive resource allocation rather than uniform processing.
Data Source
AI summary
Systems, apparatuses, methods, and computer program products are disclosed for generating a predictive temporal feature impact report using a feature engineering machine with attention for time series (FEATS model). An example method includes receiving an entity input data object. The method further includes determining one or more attention head scores for each feature attention head included in the FEATS model based at least in part on one or more per-temporal feature time impact scores over each time window for each temporal feature set. The method further includes generating a predictive temporal feature impact report based at least in part on at least one of the one or more attention head scores for each attention head or the one or more per-temporal feature time impact scores for each temporal feature time point as determined in each attention head.