Query Text Classification for E-commerce Search Accuracy
Solution Overview
Problem
Existing e-commerce platforms face challenges in providing users with efficient and relevant search results due to ambiguous query keywords and the need to wade through numerous products, leading to wasted computing resources and increased user interaction time.
Innovation Solution
The method involves classifying products into classes, associating search query text with products based on click log data, generating training data, and using an N-gram Convolutional Neural Network (CNN) to rank classes for query results, thereby improving the accuracy and efficiency of search results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If users manually browse through all products to find what they are looking for, then users can view product details, but computing resources are wasted and user interaction time increases
Solution Approach 1:
The system performs preliminary classification of products into categories and pre-processes user interaction data (click logs, search queries) before actual search operations. This allows the system to quickly retrieve pre-organized information rather than processing everything in real-time, reducing user interaction time while maintaining accuracy.
Solution Approach 2:
The patent introduces an intermediary classification layer between the user query and the product database. Instead of directly searching through all products, the system first classifies the query into relevant categories, then retrieves products from those categories. This intermediary step filters out irrelevant products before they reach the user, saving time and computing resources.
2Reliability
If the system renders graphics and information for all products, then users can see complete product information, but unnecessary processing occurs for products users are not interested in
Solution Approach 1:
The system applies local quality by providing different levels of information rendering based on relevance. Highly relevant products (those matching user intent based on classification and click log analysis) receive full rendering with detailed graphics and information, while less relevant products receive minimal or no rendering. This selective approach maintains reliability for important results while conserving computing resources.
Solution Approach 2:
The system performs partial action by rendering complete information only for the most relevant products rather than all products. Based on the classification and ranking derived from user interaction data, the system fully processes and renders only the top-ranked relevant products, while providing summaries or omitting less relevant ones, thus reducing unnecessary processing.
3Productivity
If query keywords are used to search products, then users can find relevant products, but ambiguous keywords lead to inaccurate results
Solution Approach 1:
The system incorporates feedback from user interaction data (click logs, search queries, purchase behavior) to continuously improve query interpretation. When users search with ambiguous keywords, the system analyzes historical feedback to determine which products the user actually clicked on or purchased, then uses this feedback to refine the classification and ranking of results, improving accuracy over time while maintaining search efficiency.
Solution Approach 2:
The system performs preliminary analysis of query keywords by comparing them against a database of historical search queries and click logs before executing the search. This pre-processing step classifies the query intent and identifies the most likely relevant product categories, allowing the system to handle ambiguous keywords more accurately by leveraging pre-computed relationships between queries and products.
Data Source
AI summary
A service may provide an e-commerce platform through which users may search for, view, and purchase products. To ensure users receive relevant results when querying for particular products or product classes, a list of ranked classes may be generated, from which the query results are based. For example, products may be classified into classes and particular text from search data may be associated with one or more of the products based on click log data. For instance, input text from previously performed queries may be associated with products interacted with following those queries (e.g., products clicked on to view/purchase). Training data that associates the particular text with the respective one or more classes into which the clicked on products are classified may be generated, and in response to receiving a query including the particular text as input, the list of ranked classes may be generated based on the training data.