Multi-Task Learning with Dynamic Imputation for Missing Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-task learning methods face challenges in handling missing data, often leading to biased results due to separate handling or imputation of missing values, which can distort the variance and correlation, and do not effectively capture the relatedness between tasks.
Innovation Solution
The system computes an updated weight matrix using an augmented observation matrix, plug-in autocovariance matrix, and plug-in covariance vector, incorporating a predefined penalization parameter and noise value, to predict characteristics in a scoring dataset, while handling missing features within the learning process through graph regularization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If missing features are deleted entirely, then the model training becomes simpler, but the dataset becomes statistically non-representative or biased
Solution Approach 1:
The patent introduces an intermediary imputation mechanism that fills missing features with estimated values (e.g., mean, median, or model-predicted values) rather than deleting observations or using crude constants. This intermediary step preserves the statistical properties of the dataset while enabling complete data processing, thus maintaining dataset representativeness without overwhelming complexity.
Solution Approach 2:
The patent transforms the handling approach by changing the parameter treatment of missing values - instead of treating them as absolute absences (deletion) or fixed constants (zero imputation), it introduces dynamic imputed parameters that can be updated during training. This allows the system to adapt missing value estimates as model knowledge improves, balancing complexity and reliability.
2Device complexity
If constant values are used to provide missing feature values, then the processing becomes simpler, but the data variability is underrated and bias is introduced
Solution Approach 1:
The patent makes the imputation values dynamic rather than static. Instead of using fixed constant values for all missing features, the system employs imputation values that can change during training - initially using simple constants but progressively refining them using model predictions or statistical estimates from the training data. This dynamic approach preserves data variability while maintaining processing simplicity.
Solution Approach 2:
The patent implements a feedback mechanism where the model's predictions are used to refine the imputation of missing values. As the model learns from complete observations, its predictions feed back into improving the quality of imputed values for missing features. This feedback loop gradually reduces bias and better captures data variability without significantly increasing processing complexity.
3Quantity of substance
If imputation methods are used to provide missing feature values, then the complete data can be utilized, but the uncertainty of missing values is neglected and bias is induced
Solution Approach 1:
The patent performs preliminary actions by implementing a two-phase approach: first, complete observations are used to train the model and establish reliable patterns; second, these learned patterns are applied to impute missing values in incomplete observations. This preliminary training on complete data ensures that imputation is based on validated model knowledge rather than naive assumptions, reducing bias while maximizing data utilization.
Solution Approach 2:
The patent applies partial action by selectively processing different types of observations differently - complete observations are used for primary training, while incomplete observations are processed with imputed values for supplementary learning. This partial approach allows the system to fully utilize available data while maintaining high reliability by not over-relying on potentially biased imputed values.
4Device complexity
If observations with missing features are deleted, then the training process is simplified, but considerable error is introduced in the trained model
Solution Approach 1:
The patent creates a universal training framework that can handle both complete and incomplete observations through a unified imputation mechanism. Instead of requiring separate processing pipelines for complete and incomplete data, the system uses a single multi-functional approach where imputation is applied consistently across all incomplete observations, simplifying the training process while preserving model accuracy through consistent handling of all data types.
Data Source
AI summary
A computing device computes a weight matrix to compute a predicted value. For each of a plurality of related tasks, an augmented observation matrix, a plug-in autocovariance matrix, and a plug-in covariance vector are computed. A weight matrix used to predict the characteristic for each of a plurality of variables and each of a plurality of related tasks is computed. (a) and (b) are repeated with the computed updated weight matrix as the computed weight matrix until a convergence criterion is satisfied: (a) a gradient descent matrix is computed using the computed plug-in autocovariance matrix, the computed plug-in covariance vector, the computed weight matrix, and a predefined relationship matrix, wherein the predefined relationship matrix defines a relationship between the plurality of related tasks, and (b) an updated weight matrix is computed using the computed gradient descent matrix.