Pairwise Interaction Tensor Decomposition for Tag Recommendation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing tensor decomposition methods for tag recommendation, such as Tucker decomposition and Pairwise Interaction Tensor Factorization, face high time complexity and accuracy issues due to their inability to efficiently handle large datasets, leading to inefficient training and prediction processes.
Innovation Solution
The proposed Pairwise Interaction Tensor Decomposition (PITD) method decomposes a target tensor into a core tensor and factor matrices, with the core tensor split into subtensors, reducing redundant calculations and improving efficiency by only involving target subtensors in the equation relationship, thereby decreasing time complexity and enhancing tag recommendation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Tucker decomposition model is used for tag recommendation, then the model can capture interaction relationships among user, resource and tag, but the time complexity of training and prediction is relatively high
Solution Approach 1:
The patent segments the three-dimensional core tensor obtained from Tucker decomposition into multiple two-dimensional matrices. This segmentation transforms the complex three-dimensional interaction relationships into multiple simpler two-dimensional relationships, reducing the time complexity while preserving the essential interaction patterns among users, resources, and tags.
Solution Approach 2:
The patent extracts the core tensor from the Tucker decomposition model and separates it from the factor matrices. By taking out only the necessary interaction information from the core tensor and representing it through multiple two-dimensional matrices, the model reduces computational burden while maintaining recommendation accuracy.
2Device complexity
If traditional tensor decomposition methods are used, then the model structure is simple, but the method cannot adapt to large volume datasets
Solution Approach 1:
The patent segments the large-scale three-dimensional tensor data into multiple manageable two-dimensional matrices. This segmentation allows the model to process large volume datasets by breaking them down into smaller computational units, thereby improving productivity without significantly increasing model structure complexity.
Solution Approach 2:
The patent transforms the three-dimensional tensor decomposition problem into multiple two-dimensional matrix decomposition problems. This dimensionality change from 3D to 2D reduces the computational complexity and enables the model to efficiently handle large volume datasets while maintaining a relatively simple model structure.
3Reliability
If three-dimensional core tensor is used in TD model, then the interaction relationships are preserved, but the time complexity and duration are relatively high
Solution Approach 1:
The patent segments the three-dimensional core tensor into multiple two-dimensional matrices, each capturing specific interaction patterns. This segmentation preserves the essential interaction relationships among users, resources, and tags while reducing the computational time complexity and improving training and prediction speed.
Solution Approach 2:
The patent extracts the interaction information from the three-dimensional core tensor and represents it through multiple two-dimensional matrices. This extraction process maintains the critical interaction relationships while reducing the computational burden, thereby improving productivity without sacrificing reliability.
Data Source
AI summary
A tag recommending method is performed at a computer device, the method including: obtaining a target account and a target resource for which tag recommendation is to be performed; determining, by using a tag recommending model obtained after tensor decomposition, recommendation values of tags in tag data respectively corresponding to the target account and the target resource, the tag recommending model defining correspondences among account data including the target account, resource data including the target resource and the tag data; and determining n tags with highest recommendation values in the tag data as tags to be recommended to the target account, the n tags being used for marking the target resource, and n being a positive integer.