Basket Recommendation System Using Clustering and Preference Updating
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommender systems in the restaurant and hotel industries are slow to adapt to changing offerings and customer preferences due to the need for extensive historical data analysis and lack of consideration for demographic information, making them inefficient in providing real-time recommendations.
Innovation Solution
A computer system that analyzes the entire basket of goods or services purchased by a customer to recommend additional items, using clustering and preference updating algorithms to minimize historical data requirements and adapt quickly to changes, incorporating demographic attributes and characteristics of the basket.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If frequent examination of historical data is performed to improve recommendation accuracy, then recommendation quality improves, but system speed and adaptability deteriorate due to the large amount of data processing required
Solution Approach 1:
The patent segments the historical data processing into distinct phases: an offline training phase where the system learns from historical basket data to build preference models, and an online recommendation phase where it quickly applies these models to new baskets. This segmentation allows comprehensive data analysis without slowing down real-time recommendations.
Solution Approach 2:
The system performs preliminary actions by pre-processing historical data to create preference models and item embeddings before they are needed for actual recommendations. The offline training phase pre-computes item similarities and customer preferences, so that when a new basket arrives, the system can immediately generate recommendations without extensive real-time computation.
2Measurement precision
If extensive historical data is used to improve recommendation accuracy, then recommendation quality improves, but the system complexity and computational cost increase
Solution Approach 1:
The patent extracts only the essential features and preferences from extensive historical data rather than processing all raw data during real-time operations. It extracts item embeddings, preference vectors, and clustering information during offline training, then uses only these extracted representations during online recommendations, significantly reducing computational complexity.
Solution Approach 2:
The system creates simplified copies of the complex historical data in the form of pre-computed item embeddings and preference models. Instead of repeatedly analyzing the full historical dataset, it works with these compact representations that capture the essential patterns, reducing both computational load and system complexity.
3Measurement precision
If traditional recommender systems are used that rely on extensive historical data analysis, then comprehensive recommendations can be generated, but the system cannot adapt quickly to changing offerings and customer preferences
Solution Approach 1:
The patent implements a dynamic system where preference models are continuously updated as new basket data becomes available. The offline training phase can be re-run periodically with new data to update item embeddings and preference vectors, allowing the system to adapt to changing customer preferences and offerings while maintaining comprehensive recommendation capabilities.
Solution Approach 2:
The system incorporates feedback mechanisms where actual customer purchases and basket compositions are fed back into the training process. By continuously learning from new data and updating preference models, the system adapts to changing offerings and preferences while maintaining comprehensive and accurate recommendations.
Data Source
AI summary
Embodiments provide a recommendation for an additional item in response to receiving a basket of goods determine a type for the basket of goods from a set of basket types, receive a set of additional targeted items as target recommendations and receive a history of received types of baskets of goods. Embodiments iteratively perform a clustering into a plurality of clusters of each of the basket types based on the history of received types of baskets of goods, and preference updating for each of the targeted items into each of the plurality of clusters. The iteratively performing, after a plurality of iterations, outputs a sequence of mappings and a sequence of preference parameters. Embodiments generate a frequency of tabulation of mappings from the sequence of mappings and then generate the recommendation based on the sequence of mappings, the sequence of preference parameters and the frequency of tabulation of mappings.