Neural Tensor Completion via Influence-Based Data Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional tensor completion methods often lead to inaccurate predictions and high computational time and space complexities, making them impractical for large, sparse arrays, especially in applications like movie recommendation systems where missing values need to be predicted efficiently.
Innovation Solution
The use of neural tensor-completion models combined with influence functions and data augmentation techniques to estimate cell and entity importance, allowing for weighted sampling and improved prediction accuracy, thereby reducing time and space complexities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional tensor completion methods are used to predict missing values in large sparse arrays, then prediction accuracy can be maintained, but computational time and space complexities become prohibitively high
Solution Approach 1:
The patent segments the tensor completion problem by identifying and focusing on a small subset of critical training samples that have the most significant influence on prediction accuracy. Instead of processing all training samples uniformly, the method divides them into influential and non-influential groups, processing only the influential subset to reduce computational time while maintaining prediction accuracy.
Solution Approach 2:
The patent changes the parameter of sample selection from uniform random sampling to influence-function-based sampling. By computing influence functions, the method identifies training samples with high influence on test predictions and adjusts the sampling distribution to prioritize these samples, thereby improving computational efficiency without sacrificing accuracy.
2Measurement precision
If conventional tensor completion methods are used to predict missing values in large sparse arrays, then prediction accuracy can be maintained, but space complexity becomes prohibitively high
Solution Approach 1:
The patent extracts only the most influential training samples from the complete training set, discarding samples that have minimal impact on prediction accuracy. This extraction is achieved through influence function computation, which identifies and isolates the critical subset of training data needed for effective tensor completion, thereby reducing space complexity.
Solution Approach 2:
The training set is segmented into influential and non-influential samples based on influence function measurements. Only the influential segment is retained and used for training the tensor completion model, while the non-influential segment is discarded, significantly reducing the space required to store training data.
3Ease of manufacture
If uniform random sampling is used to select training samples for tensor completion, then implementation is simple, but prediction accuracy deteriorates due to inclusion of unimportant samples
Solution Approach 1:
The patent changes the sampling parameter from uniform probability to influence-based probability. By computing influence functions for each training sample, the method assigns different sampling probabilities proportional to each sample's influence on test predictions, ensuring that more important samples are selected more frequently while maintaining a probabilistic sampling framework.
Solution Approach 2:
The patent introduces feedback through influence function computation, where the impact of each training sample on test predictions is measured and used to adjust the sampling strategy. This feedback loop ensures that samples with higher influence are prioritized in the sampling process, improving prediction accuracy without excessive complexity.
Data Source
AI summary
A data augmentation framework enhances the prediction accuracy of tensor completion methods. An array having a set of cells associated with a set of entities is received. Influence metrics of cells from the array are determined based on an influence of the cells on minimizing loss while training a machine learning model. An entity-importance metric is generated for each entity of the set of entities based on the influence metrics. A cell from the array for which to augment the array with a predicted value is identified. The cell is identified based on a sampling of the set of entities that is weighted by the entity-importance metric for each entity of the set of entities.