Complementary Item Selection via Category-Based Subspace-Attention Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems lack an efficient method for selecting complementary consumer items that are visually and functionally compatible with a given set, especially in online retail, where brute force approaches are impractical due to the need for scalable and accurate compatibility assessment.
Innovation Solution
A category-based subspace-attention network is employed to generate feature embeddings for consumer items, using a neural network that processes images and category information to produce embeddings that reflect compatibility, allowing for efficient indexing and selection of compatible items based on a chosen distance metric.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If brute force approaches are used for complementary item selection, then comprehensive compatibility assessment can be achieved, but computational complexity and processing time become impractical
Solution Approach 1:
The patent segments the compatibility assessment task by dividing it into visual feature extraction, category-based filtering, and compatibility scoring components. The visual feature extraction uses pre-trained CNN models to extract features from item images, category-based filtering narrows down candidate items using category information, and compatibility scoring computes final compatibility metrics. This segmentation reduces computational complexity from exhaustive pairwise comparison to a multi-stage filtering process.
Solution Approach 2:
The patent performs preliminary actions by pre-computing visual feature embeddings for all items in the database and organizing them by category. When a user requests complementary items, the system retrieves pre-computed features and performs only the final compatibility calculation, rather than computing all features on-demand. This preliminary action significantly reduces processing time while maintaining assessment accuracy.
2Measurement precision
If exhaustive pairwise comparison is performed, then all compatibility relationships can be evaluated, but processing time becomes prohibitively long
Solution Approach 1:
The patent segments the evaluation process into three phases: (1) retrieve pre-computed visual features for query item and candidate items, (2) filter candidates using category compatibility rules, and (3) compute compatibility scores only for filtered candidates. This segmentation reduces processing time by avoiding unnecessary computations for incompatible category combinations while maintaining complete evaluation for all potentially compatible pairs.
Solution Approach 2:
The patent performs preliminary action by pre-computing and storing visual feature embeddings for all items in the database during an offline phase. During online query processing, these pre-computed features are retrieved and used for rapid compatibility assessment, eliminating the need for real-time feature extraction and significantly reducing processing time.
3Productivity
If traditional recommendation systems are used, then item suggestions can be provided, but visual and functional compatibility cannot be effectively assessed
Solution Approach 1:
The patent introduces visual feature embeddings as an intermediary representation that bridges traditional recommendation systems and compatibility assessment. These embeddings capture visual characteristics of items and enable computation of visual compatibility through distance metrics in the embedding space. The category information serves as another intermediary that encodes functional compatibility rules. Together, these intermediaries enable both fast processing and accurate compatibility assessment.
Solution Approach 2:
The patent transforms the compatibility assessment problem into a parameter-based computation by representing items as vectors of visual features and categories. Compatibility is computed as a function of these parameters (e.g., Euclidean distance between visual features, category matching). This parameter transformation enables efficient computation while capturing complex visual and functional compatibility relationships that traditional rule-based systems cannot handle.
Data Source
AI summary
An example method of complementary consumer item selection includes: receiving, by a computer system, an image representing a first consumer item of a reference set of consumer items; identifying a target category of a complementary consumer item to be associated with the reference set of consumer items; generating, by a neural network processing the set of images, a feature embeddings representing the first consumer item in relation to the target category; selecting, using the feature embedding, from a set of available consumer items, a plurality of candidate consumer items associated with the target category; and selecting, among the plurality of candidate consumer items, the complementary consumer item to be associated with the reference set of consumer items.