Propensity Modeling With Similarity Scoring and Space Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning models lack efficiency in identifying propensities and generating accurate predictions due to their inherent strengths and weaknesses across different categories and types, particularly when analyzing user data from distinct groups like credit and debit card users.
Innovation Solution
A method involving a combination of supervised and unsupervised machine-learning models, specifically using XGBoost classifier and K-D Tree algorithms, to generate similarity scores and space-partitioning data structures for users, enabling targeted offer messaging based on historical transaction data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single machine-learning model is used to analyze user data, then the model structure is simple, but the accuracy of identifying user propensities is insufficient
Solution Approach 1:
The patent combines multiple machine-learning models (supervised model for similarity scoring and unsupervised model for space-partitioning) into an integrated propensity identification system. The supervised model generates similarity scores between target users and reference users, while the unsupervised model creates space-partitioning data structures to efficiently query and compare user representations. This merging of models resolves the contradiction by achieving high identification accuracy through model collaboration while maintaining a structured, modular architecture that manages complexity.
Solution Approach 2:
The patent segments the propensity identification process into distinct functional components: (1) data preparation and user representation generation, (2) similarity score calculation using supervised models, (3) space-partitioning using unsupervised models, and (4) propensity prediction. This segmentation allows each component to be optimized independently, improving overall accuracy while making the complex system more manageable and interpretable.
2Measurement precision
If multiple machine-learning models are combined to improve prediction accuracy, then the identification precision increases, but the system complexity increases
Solution Approach 1:
The patent implements a universal framework where the combination of supervised and unsupervised models serves multiple functions: the supervised model not only generates similarity scores but also learns from labeled data to improve over time, while the unsupervised model creates space-partitioning structures that enable efficient querying and can be applied to various user segmentation tasks. This multi-functionality justifies the increased system complexity by delivering comprehensive accuracy improvements across multiple operational dimensions.
Solution Approach 2:
The patent introduces an intermediary layer of user representations and similarity scores that bridge the supervised and unsupervised models. The supervised model outputs similarity scores that serve as intermediaries, which are then processed by the unsupervised space-partitioning model to generate final propensity predictions. This intermediary structure manages system complexity by creating clear interfaces between model components while maintaining high prediction accuracy through coordinated processing.
3Measurement precision
If comprehensive user data is analyzed using complex models, then the propensity identification accuracy improves, but the computational resources required increase
Solution Approach 1:
The patent applies preliminary action by pre-computing user representations and organizing them into space-partitioning data structures using the unsupervised model. These pre-processed structures enable efficient querying and comparison during propensity prediction, reducing the computational resources required at inference time. The supervised model also performs preliminary similarity scoring that narrows down the search space, further reducing computational demands while maintaining high identification accuracy.
Data Source
AI summary
The method may include inputting a first set of data into a first model; for each user in the second group, generating a first similarity score; generating a relevance score for each parameter; determining a subset of parameters based on relevance; inputting the subset of parameters, a second set of data, and a third set of data into a second model; generating a space-partitioning data structure based on the second set of data; for each user in the first group, determining a feature distance between a representation of the user in the first group and a representation of a user in the second group based on the third set of data and the space-partitioning data structure; for each user in the second group, generating a second similarity score; and for each user in the second group, generating an overall similarity score.