Multi-Level Taxonomy Classification With Logit Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for hierarchical taxonomy classification in retail sector are not directly applicable to few-shot datasets, and they are not directly applicable to few-shot datasets, given the existing training strategy.
Innovation Solution
A progressive multi-level training framework with a logit-masking strategy for retail taxonomy classification, which includes extracting a first set of visual information, and passing it to a set of fully connected layers of a convolution neural network, wherein the weights of the convolutional layers are frozen for each hierarchical level information, and the logit masking strategy addresses label inconsistency by hiding irrelevant classes based on the first level hierarchical information to obtain the second level hierarchical information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If level-based local classifier trains separate models for each level, then classification accuracy at each level is improved, but label inconsistency problem crops up where outputs of different levels create an impossible combination
Solution Approach 1:
The patent introduces a mediator mechanism that coordinates between level-based classifiers. A global classifier is trained to predict parent nodes, and this prediction is used to mask irrelevant classes in the child node classification. This mediator approach allows each level to be classified independently (maintaining accuracy) while ensuring consistency across levels through the masking mechanism (resolving label inconsistency).
Solution Approach 2:
The patent changes the parameter space by introducing a masking mechanism that dynamically adjusts the classification problem at each level. By masking out impossible classes based on parent node predictions, the system transforms the classification task to ensure only consistent labels are considered, thereby resolving the label inconsistency issue while maintaining level-specific accuracy.
2Reliability
If parent node-based local classifier is trained where model in one level is selected based on decision by predecessor, then label consistency is improved, but computational cost increases significantly in retail scenario having large number of classes
Solution Approach 1:
The patent extracts the computational burden by separating the classification process into two independent parts: a global classifier that predicts parent nodes, and level-specific classifiers that handle child nodes. By extracting the parent node prediction task from the child node classification, the system reduces the computational complexity at each level while maintaining consistency through the extracted parent predictions.
Solution Approach 2:
The patent segments the classification system into independent level-based classifiers that process different hierarchical levels separately. This segmentation allows each classifier to operate independently with its own training data and parameters, significantly reducing the computational cost compared to training a single comprehensive model, while the segmentation itself ensures label consistency through the masking mechanism.
3Quantity of substance
If existing hierarchical taxonomy classification frameworks are used, then large-scale datasets are handled effectively, but they are not directly applicable to few-shot datasets
Solution Approach 1:
The patent creates a universal classification framework that can handle both large-scale and few-shot datasets effectively. The level-based architecture with global classifier and masking mechanism is designed to be adaptable to different data quantities. The system uses transfer learning capabilities and can function with limited training data at each level, making it versatile for various dataset sizes including few-shot scenarios.
Solution Approach 2:
The patent applies preliminary action through pre-training the global classifier and convolutional layers on large-scale data before fine-tuning on specific retail datasets. This preliminary training establishes a robust feature extractor and classification framework that can be adapted to few-shot scenarios, as the foundational models are already trained on comprehensive data, allowing efficient fine-tuning with limited data.
Data Source
AI summary
The embodiments of present disclosure address unresolved problems of label inconsistency, where outputs of different levels create impossible combinations, and error propagation from previous level outputs can significantly impact its performance. Embodiments provide a method and system for a Progressive Multi-level Training framework with a Logit-masking strategy (PMTL) for a retail taxonomy classification. PMTL enables neural network models to be trained separately for each level to reduce error propagation problems. To further enhance the model's performance at each level and get the label-wise constraint from the previous level, the global representation from model of previous level is augmented. Further, a logit masking strategy is used to restrict model(s) to learning only relevant classes through part of final classification layer, thereby addressing label inconsistency issue, and incorporating benefit of parent node-based local classifier. This framework is generalized irrespective of dataset size and is configured for attaching to any hierarchical classification network.