Progressive Retail Taxonomy Training with Logit Masking for Consistent Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing hierarchical taxonomy classification frameworks face challenges with label inconsistency and error propagation, particularly in retail scenarios with large numbers of classes, and are not effectively applicable to few-shot datasets.
Innovation Solution
A progressive multi-level training framework with a logit-masking strategy that trains models separately for each level, using logit masking to hide irrelevant classes and incorporate global representation from previous levels, thereby reducing error propagation and label inconsistency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If parent node-based local classifier is trained to mitigate label inconsistency, then label consistency is improved, but computational cost increases significantly
Solution Approach 1:
The classification task is segmented into multiple hierarchical levels where each level handles a specific granularity of classification. Instead of training separate parent node-based local classifiers for each parent node (which would be computationally expensive), the system divides the problem into L levels with progressively finer granularity, processing only the necessary subset of classes at each level.
Solution Approach 2:
The system applies partial action by processing only the relevant subset of classes at each hierarchical level rather than evaluating all classes globally. At level l, only classes belonging to the selected parent node from level l-1 are processed, reducing the computational burden while maintaining classification accuracy.
2Productivity
If level-based local classifier trains separate models for each level, then classification speed is improved, but label inconsistency problem occurs
Solution Approach 1:
The system performs preliminary classification at each hierarchical level before proceeding to the next level. By making decisions at level l-1 first and using those results to guide classification at level l, the system ensures label consistency while maintaining the speed benefits of separate level-based processing. The logit masking is applied preliminarily to restrict subsequent classification to valid child classes.
Solution Approach 2:
The hierarchical structure provides feedback from higher levels to lower levels. The classification results from level l-1 feed into level l, ensuring that labels are consistent across levels. The system uses the parent node selection to feedback and constrain the child node classification, preventing label inconsistency while maintaining processing efficiency.
3Device complexity
If global classifier is used for hierarchical training, then model simplicity is maintained, but error propagation from previous levels impacts performance
Solution Approach 1:
The global classifier is segmented into multiple level-specific classifiers arranged hierarchically. Each level l has its own classifier that processes only the relevant subset of classes at that level. This segmentation prevents error propagation because each level's classifier is independent and makes decisions based on its own features and the constrained input from the previous level, rather than propagating errors from a single global model.
Solution Approach 2:
Instead of using a single global classifier with uniform structure, the system employs local quality by having each hierarchical level have its own specialized classifier tailored to that level's classification task. Each classifier is optimized for its specific level's characteristics, improving overall accuracy while maintaining reasonable model complexity through the hierarchical organization.
Data Source
Figure 1
Figure 2
Figure 3
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 Multilevel 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.