Visual Item Recommendation Using CNN Features Under Sparse Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing e-commerce systems face challenges in recommending similar items due to sparse purchase data, lack of manufacturer IDs, and insufficient behavioral data, especially for newly released or low-selling products, leading to inefficiencies in collaborative filtering methods.
Innovation Solution
An image analysis engine uses a Convolutional Neural Network (CNN) to extract feature vectors from item images, and a k-nearest neighbors (KNN) process to identify visually similar items, leveraging a custom or pre-trained model like GoogLeNet for enhanced recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If collaborative filtering methods are used for product recommendations, then recommendations can be generated based on user behavior data, but the method performs poorly when there is insufficient purchase data, manufacturer IDs, or behavioral data
Solution Approach 1:
The patent replaces the traditional collaborative filtering mechanism (which relies on user behavior data and purchase history) with an image-based recommendation system. Instead of using user interaction patterns, the system uses computer vision to analyze product images and extract visual features, thereby substituting a data-intensive mechanical system with one that operates on visual information
Solution Approach 2:
The patent introduces image features as an intermediary between products and recommendations. Rather than directly comparing user behaviors or product metadata, the system extracts visual features from product images and uses these features as the basis for finding similar products, creating a new information pathway that bypasses data scarcity issues
2Adaptability or versatility
If collaborative filtering is used, then product recommendations can be generated, but the method performs poorly with cross-category recommendations
Solution Approach 1:
The patent creates a universal recommendation system based on visual features that can operate across different product categories. The image-based approach extracts fundamental visual characteristics (colors, shapes, textures, patterns) that are category-agnostic, allowing the same system to effectively recommend similar items whether they are shoes, electronics, clothing, or other products
3Reliability
If traditional recommendation systems are used, then recommendations can be provided for well-documented products, but they fail for newly released or low-selling products with sparse data
Solution Approach 1:
The system replaces the purchase-data-dependent recommendation mechanism with an image-analysis-based approach. For newly released or low-selling products that lack purchase history, the system extracts visual features from product images to generate recommendations, substituting a system that requires substantial user data with one that operates effectively on visual information alone
Data Source
AI summary
A computer-implemented method includes determining a set of target listings, retrieving a seed image associated with the seed listing, the seed listing is categorized within a first item category, and generating a seed item feature vector for the seed image using a convolutional neural network (CNN) trained with images of items. The method also includes identifying a plurality of feature vectors associated with the first item category, comparing the seed item feature vector to the plurality of feature vectors using a k-nearest neighbors (kNN) algorithm, and generating a set of nearest neighbor listings to the seed listing. The method further includes storing the set of nearest neighbor listings as associated with the seed listing, selecting one or more nearest neighbor listings from the set of nearest neighbors, and presenting the one or more nearest neighbor listings as a recommendation to a user of the online e-commerce system.