Complementary Item Retrieval Using Latent Candidate Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation engines struggle to provide compatible suggestions to users based on recent purchase and pre-selected items or most likely context based matching items.
Innovation Solution
A system and method to recommend complementary items through candidate target item generation is provided.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing recommendation models use annotated compatible data for model fitting, then they can provide complementary item recommendations, but the recommendations are biased towards annotators preference and cannot be generalized
Solution Approach 1:
The patent uses generative AI models to create synthetic complementary item recommendations that replicate the structure and logic of valid recommendations without copying specific annotator biases. The model learns the underlying patterns of compatibility from training data and generates new recommendations that follow these patterns without being constrained by annotator subjectivity.
Solution Approach 2:
The system transforms the recommendation generation process from discrete annotator-based labeling to continuous generative modeling. By using diffusion models and transformer architectures, the system transitions from binary compatible/incompatible annotations to probabilistic generation of complementary items based on learned latent representations and attention mechanisms.
2Stability of the object's composition
If existing methods use deterministic approaches for complementary item retrieval, then they provide consistent results, but they cannot adapt to diverse user preferences and contradictory requirements
Solution Approach 1:
The patent introduces dynamic adaptability through transformer-based attention mechanisms that can adjust their focus based on input context. The model dynamically weights different features and relationships based on the specific user query and item characteristics, allowing it to adapt to diverse preferences while maintaining coherent recommendations through controlled generative processes.
Solution Approach 2:
The system uses learnable parameters in the transformer model that can be adjusted during inference to accommodate different user preferences. The attention weights and generation probabilities are dynamically computed based on the input, allowing the same model to adapt to various user contexts without requiring deterministic rules for each scenario.
3Productivity
If the system retrieves complementary items from a retrieval gallery, then it can provide recommendations, but it cannot generate new candidate target items that match user preferences
Solution Approach 1:
The patent performs preliminary generation of candidate target items using generative models before the actual recommendation retrieval. The system pre-generates a diverse set of potential complementary items based on user preferences and item characteristics, then uses these generated candidates as input for the retrieval gallery search, ensuring both creativity and efficiency in the recommendation process.
Solution Approach 2:
The system merges two previously separate functions: generative item creation and retrieval gallery search. By combining the generative capabilities of diffusion models with the retrieval efficiency of gallery search, the system achieves both novel recommendation generation and efficient filtering of suitable items from the retrieval gallery.
Data Source
AI summary
This disclosure relates generally to method and system to recommend complementary items by generating candidate items. Complementary recommendation is an important problem in e-commerce platforms that gives compatible suggestions to the users based on recent purchase and pre-selected items. The method receives a mixed query as input from a user to obtain complementary target candidate image items. The mixed query includes a set of product category images along with product category label preselected by the user. Further, for the mixed query a target latent representation for the combined latent representation is generated. Then, a set of compatible complementary target candidate image items is retrieved for the one or more target candidate images from a retrieval gallery. Finally, the set of compatible complementary target candidate image items are displayed on electronic device of the user.