Product Data Categorization via Text and Attribute Factorization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing product categorization techniques in e-commerce struggle to accurately categorize products due to the complexity of semantics and the lack of effective utilization of structured data for downstream NLP tasks.
Innovation Solution
A system and method for product data categorization based on text and attribute factorization, utilizing a pretrained Word2Vec model to extract multi-level contextual data and assign weights to product data for accurate categorization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If word embedding technique is used for product categorization, then semantic understanding is improved, but categorization accuracy deteriorates due to continuous bag of words approach
Solution Approach 1:
The patent segments the product categorization process into distinct components: structured attribute processing and unstructured text processing. This segmentation allows each component to be optimized independently, with structured attributes providing precise categorical information and unstructured text providing semantic context, thereby resolving the contradiction between semantic understanding and categorization accuracy
Solution Approach 2:
The patent introduces an intermediary mechanism that combines structured attribute data with unstructured text data. This intermediary approach allows the system to leverage both the precision of structured attributes and the semantic richness of unstructured text, overcoming the limitations of using word embedding alone
2Adaptability or versatility
If all available product attributes are considered for modeling, then categorization completeness is improved, but device complexity deteriorates
Solution Approach 1:
The patent segments attributes into structured and unstructured categories, processing them through different pathways. Structured attributes are processed systematically while unstructured text is processed through NLP techniques. This segmentation reduces modeling complexity by handling different attribute types appropriately while maintaining comprehensive categorization coverage
Solution Approach 2:
The patent applies different processing qualities to different attribute types: structured attributes receive systematic categorical processing while unstructured text receives semantic processing. This local quality approach optimizes the modeling process for each attribute type's characteristics, reducing overall complexity while maintaining completeness
3Measurement precision
If deep neural networks are used for product categorization, then categorization accuracy is improved, but computational efficiency deteriorates
Solution Approach 1:
The patent performs preliminary processing of structured attributes and unstructured text separately before combining them for final categorization. This preliminary action allows simpler processing steps to be executed independently, reducing the computational burden on the final classification model while maintaining high accuracy through the combined information
Data Source
Figure 1
Figure 2
AI summary
This disclosure relates generally to method and system for product data categorization based on text and attribute factorization. The method includes acquiring an input describing a set of product data from an application data store for categorization. The set of product data by removing extraneous text based on a predefined template. Further a dictionary for the set of product data based on a set of attributes comprising a product key with its corresponding product value. Further, a multi-level contextual data for the set of product data are extracted by assigning a weight to each product data based on likelihood and creating a set of datapoints for each product data. The set of product data are categorized by feeding the set of data points to a set of predefined parameters to compute a minimum count, a total size, total number of epochs, a skip gram value and a hierarchical softmax.