Webpage Classification Framework Using Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing targeted content delivery systems face challenges in accurately classifying webpages for contextual targeting due to high computational and financial resource requirements, especially in a cookie-less world where traditional tracking methods are obsolete, and require efficient mechanisms for scalable webpage classification.
Innovation Solution
A novel taxonomic webpage classification framework using transfer learning, knowledge distillation, and class-based re-weighting techniques to train multi-label classification models, allowing for efficient classification of webpages without full-page content access, thereby reducing inference latency and computational costs while maintaining high accuracy across multiple languages.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional webpage classification approaches are used to accurately classify webpages for contextual targeting, then classification accuracy is improved, but computational resources and financial costs increase significantly
Solution Approach 1:
The system performs preliminary classification of webpages into broad categories using lightweight models before applying more resource-intensive classification only when necessary. This pre-filtering approach reduces the overall computational burden while maintaining accuracy for the most important classification cases.
Solution Approach 2:
The classification process is divided into multiple stages: first categorizing webpages into broad top-level categories using efficient methods, then progressively refining to more specific categories only for pages that require higher precision. This segmentation allows the system to allocate computational resources efficiently across different classification granularities.
2Measurement precision
If full page content is accessed to enable prediction at request time, then classification accuracy is improved, but performance requirements of service-level agreements are not met due to increased latency
Solution Approach 1:
The system extracts and utilizes only the most informative features from page content for classification, rather than processing the entire page content. By identifying and extracting key discriminatory features, the system achieves high classification accuracy while significantly reducing the time required for feature extraction and processing.
Solution Approach 2:
The system pre-processes and extracts relevant features from page content in advance, storing them for rapid access during request-time prediction. This preliminary feature extraction eliminates the need to crawl and process full page content at request time, thereby reducing inference latency while maintaining classification accuracy.
3Adaptability or versatility
If category-based targeting uses broad categories to apply to diverse web pages, then versatility is improved, but ability to capture meaningful user interests is reduced
Solution Approach 1:
The taxonomy is organized into hierarchical segments where broad categories provide versatility for diverse webpages, while narrower sub-categories capture specific user interests. The system can select the appropriate granularity level based on the specific application context, allowing both broad applicability and precise interest capture as needed.
Solution Approach 2:
The system adds a dimensional aspect to category classification by incorporating multiple classification dimensions (e.g., topic category, content type, audience demographic) rather than relying on a single categorical dimension. This multi-dimensional approach allows broad categories to provide versatility while specific dimensions capture meaningful user interests.
Data Source
AI summary
In some aspects, the techniques described herein relate to a method including: receiving, at a processor, an uncrawled URL corresponding to a webpage; applying, by the processor, a webpage classification model to the uncrawled URL to determine a probability for a plurality of webpage classifications; assigning, by the processor, one or more labels to the uncrawled URL corresponding to one or more classifications of the plurality of webpage classifications that meet a threshold; and providing, by the processor, a content item to be displayed on the webpage based on the one or more labels.