Evasive Listing Detection via Word Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Online transactional fraud, particularly price gouging, evolves rapidly, making it difficult for retailers to detect using heuristic-based filtering systems as fraudsters modify listing terms to evade detection.
Innovation Solution
A machine learning-based approach that generates n-dimensional word embeddings for text corpora to identify sensitive terms related to potential fraud, using Levenshtein distance and pre-trained language models to determine evasive terms and flag listings that violate price guardrails.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If heuristic-based filtering systems are used to detect fraudulent listings, then the system is easy to implement and understand, but fraudsters can easily evade detection by modifying listing terms
Solution Approach 1:
The patent replaces heuristic-based filtering (mechanical rule-based system) with machine learning models that use n-dimensional word embeddings and semantic analysis. This substitution enables the system to detect evasive terminology through semantic similarity rather than exact keyword matching, significantly improving detection effectiveness while maintaining automated operation.
Solution Approach 2:
The patent transforms the detection approach by changing from discrete keyword matching to continuous n-dimensional semantic space representation. By representing terms as vectors in high-dimensional space and measuring semantic similarity through distance metrics, the system can detect subtle variations and evasive modifications that heuristic filters miss.
2Reliability
If machine learning-based detection with n-dimensional word embeddings is used, then detection effectiveness against evasive terms improves, but system complexity increases
Solution Approach 1:
The patent pre-trains language models and generates word embeddings offline before deployment. These pre-computed embeddings and trained models are then deployed as fixed components, eliminating the need for complex real-time training and reducing operational complexity while maintaining high detection effectiveness.
Solution Approach 2:
The patent introduces n-dimensional word embeddings as an intermediary layer between the input text and the detection logic. This embedding layer transforms complex linguistic variations into a standardized semantic representation space, simplifying the subsequent detection process and enabling efficient similarity comparisons.
3Adaptability or versatility
If semantic similarity analysis using n-dimensional embeddings is performed, then the ability to detect evasive terminology across multiple languages improves, but computational resources required increase
Solution Approach 1:
The patent employs multi-lingual language models that generate word embeddings in a shared semantic space across multiple languages. This universal embedding space allows the same detection mechanism to operate effectively across different languages without requiring language-specific models or additional computational infrastructure.
Solution Approach 2:
The patent uses pre-trained language models that have already learned linguistic patterns and semantic relationships during offline training. By copying these pre-learned representations into the detection system, the patent avoids the computational expense of training models from scratch while maintaining multi-lingual detection capabilities.
4Measurement precision
If the Levenshtein distance and pre-trained language models are used to identify evasive terms, then precision in detecting modified terms improves, but processing time increases
Solution Approach 1:
The patent applies a threshold-based filtering approach where only terms with semantic similarity scores above a certain threshold are flagged for further investigation. This partial action approach focuses computational resources on high-probability cases, reducing overall processing time while maintaining high precision in detecting evasive terminology.
Data Source
AI summary
Systems and techniques are generally described for detecting evasive terms in item listings. In some examples, a first list of terms is determined. Terms of the first list of terms may be associated with evasive listing practices. In various examples, a first embedding representing a first term in an embedding space may be determined. In some examples, the top k closest embeddings in the embedding space to the first embedding may be determined. A term associated with one of the top k embeddings may be determined. A first Levenshtein distance between the first term and the second term may be determined. A list of items associated with the second term may be determined based on the Levenshtein distance being less than a threshold distance. An alert related to the list of items may be generated.