Product Information Encoding for Cold-Start Co-Purchase Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
New stores or products with little to no usable sales history face challenges in providing relevant product recommendations due to the lack of a common product identification system across stores and the introduction of new products with different identifiers, making it difficult to map recommendations and utilize existing sales history.
Innovation Solution
A machine-learning model is developed using product information encoding to generate numeric vectors, pairing them based on co-purchase history, and training the model to predict likely product pairs, applicable across different stores and products, regardless of available sales history.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a common product identification system is implemented across all stores, then product recommendation mapping becomes feasible, but system complexity and implementation cost increase
Solution Approach 1:
The patent introduces a centralized product identification service that acts as an intermediary between stores and recommendation systems. This service assigns universal product identifiers and maintains a mapping database, enabling recommendation mapping across stores without requiring direct integration between them. The intermediary handles the complexity of identifier harmonization centrally.
Solution Approach 2:
The patent implements a universal product identifier system that serves multiple functions: identifying products across different stores, enabling recommendation mapping, and supporting transfer learning. This single identifier system replaces the need for store-specific identifiers and multiple mapping mechanisms.
2Productivity
If sales history from existing stores is used to inform recommendations for new stores, then recommendation capability is provided for new stores, but accuracy decreases due to lack of store-specific purchasing patterns
Solution Approach 1:
The patent pre-trains a machine learning model using sales history from existing stores before the new store begins operations. This preliminary training establishes baseline recommendation capabilities immediately, which are then refined as the new store accumulates its own sales data. The model is updated periodically to incorporate new store-specific patterns.
Solution Approach 2:
The recommendation system dynamically adapts between using aggregated multi-store data for new stores and transitioning to store-specific data as it accumulates. The system smoothly transitions from relying on transfer learning from existing stores to relying on the new store's own sales history, adjusting the weight of each data source over time.
3Adaptability or versatility
If product information encoding is used instead of sales history, then recommendations can be provided for products without sales history, but the model requires extensive training data to achieve accurate predictions
Solution Approach 1:
The patent encodes product information (descriptions, attributes, categories) into numerical representations that capture product characteristics. This encoded information serves as a substitute for actual sales history, allowing the model to make recommendations based on product similarity and inferred relationships rather than requiring extensive historical purchase data for each product.
Data Source
AI summary
Although sales history may be used to determine which products are likely to be purchased together, new products or stores may have little to no usable sales history. It is challenging to use the sales history of other products or stores because there is no system for uniquely identifying products that is common to all stores. Aspects of the present disclosure provide systems and methods for processing product information using a machine-learning model to predict the likelihood of products being purchased together. According to some aspects of the present disclosure, product information may be encoded to obtain numerical vectors for input to a machine-learning model to transform product information into a format in which it can be leveraged by the machine-learning model to identify co-purchasing trends that may be common to different stores and/or products.