ML Recommendation Framework for Resource Allocation Products
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional approaches to recommend resource allocation and storage products have distinct transactional characteristics that make them unique from many consumable products such as clothing or automobiles. These characteristics include extended consumption tenure, infrequent purchase, sparse product features, and transitory nature of feedback, which challenge existing recommendation systems like item-based recommendation, content-based filtering, and matrix factorization due to the absence of a rating capture mechanism, cold-start issues, and statistically insignificant product feature differences.
Innovation Solution
A training and deployment framework for a machine learning-based recommendation system, comprising a Resource Recommender (RR) subsystem and a Runtime Consumption Unit (RCU), which includes data distillation, product rating derivation, model evaluation, and enrichment stages to address the unique transactional characteristics and cold-start issues of resource allocation and storage products, using machine learning models like Item-Based Collaborative Filtering (IBCF) and Matrix Factorization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional recommendation systems (item-based, content-based, matrix factorization) are used for resource allocation and storage products, then the systems can provide recommendation functionality, but they fail to capture ratings due to the absence of a rating capture mechanism and suffer from cold-start issues
Solution Approach 1:
The patent derives product ratings in advance from transactional data and product features before the recommendation process. This preliminary action creates a rating foundation that eliminates the cold-start problem and enables conventional recommendation algorithms to function effectively for resource allocation products that lack inherent rating mechanisms.
Solution Approach 2:
The patent introduces product ratings as an intermediary construct that bridges the gap between transactional data and recommendation algorithms. By deriving ratings from transaction patterns, product features, and user behavior, the system creates a mediator that enables standard recommendation systems to work with resource allocation products.
2Productivity
If recommendation systems use sparse product features for resource allocation products, then the systems can process available data, but the statistical significance of product feature differences is insufficient for accurate recommendations
Solution Approach 1:
The patent transforms sparse product features into meaningful rating parameters by deriving product ratings from multiple data sources including transactional relationships, product attributes, and user behavior patterns. This parameter transformation converts insufficient feature data into statistically significant rating values that enable accurate recommendations.
Solution Approach 2:
The patent creates composite rating values by combining multiple data sources (transactional data, product features, user interactions) into a unified product rating metric. This composite approach compensates for the sparsity of individual features by aggregating information from diverse sources to achieve statistical significance.
3Reliability
If the system derives product ratings from transactional data and product features, then the system can overcome cold-start issues, but the complexity of the data processing framework increases
Solution Approach 1:
The patent segments the rating derivation process into distinct components: transactional relationship analysis, product feature evaluation, and rating synthesis. This segmentation allows each component to be processed independently and systematically, managing the overall complexity through modular decomposition of the data processing framework.
Data Source
AI summary
Disclosed herein is a training and deployment framework for a machine learning based recommendation system. In one aspect, a computer-implement method is provided for synthesizing training and testing data and using the training and testing data to generate a machine learning model. The computer-implement method includes deriving product ratings for products, the product ratings being stored in a user product score table in association with identifiers for users and the products, sampling the user product score table to generate a training and testing data set, training, using the training data set, machine learning models for a task of predicting product ratings, evaluating, using the testing data set, performance of the machine learning models, selecting one of the machine learning models for production use in predicting product ratings based on the evaluating, and predicting, using the selected machine learning model, new product ratings for the products.