Aggregated Feature Importance for Reliable Customer Data Analytics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional customer data analytics methods face challenges with imperfect data, such as small sample sizes and multicollinearity, leading to unreliable predictive results, especially when dealing with customer satisfaction surveys.
Innovation Solution
A system and method that uses recursive feature elimination and aggregation of feature importance to identify influential features in customer data, minimizing multicollinearity and improving reliability by generating machine learning models with different random seeds and dataset combinations, and determining the most influential feature based on frequency analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional survey methods are used to collect customer data, then data collection is simple and quick, but the sample size is small and data quality is poor leading to unreliable analytics results
Solution Approach 1:
The patent segments the limited survey data into multiple bootstrap samples through random resampling with replacement. Each bootstrap sample serves as a separate dataset for training machine learning models, effectively multiplying the usable data quantity from a small original survey dataset.
Solution Approach 2:
The patent performs preliminary data preprocessing and feature engineering before model training, including handling missing values, encoding categorical variables, and extracting features. This preliminary preparation ensures that even small datasets are optimized for analysis before being subjected to bootstrap resampling and model training.
2Reliability
If machine learning models are trained on small survey datasets, then model generation is fast, but multicollinearity occurs and predictive reliability decreases
Solution Approach 1:
The patent implements a feedback mechanism where multiple machine learning models are trained on different bootstrap samples, and their predictions are aggregated. The feature importance is determined by aggregating results across all models, providing feedback that reduces multicollinearity effects and improves predictive reliability through ensemble consensus.
Solution Approach 2:
The patent changes the parameter of random seed initialization for each machine learning model training iteration. By varying the random seed and performing bootstrap resampling, the system explores different data configurations and model initializations, reducing sensitivity to initial conditions and mitigating multicollinearity issues.
3Reliability
If multiple machine learning models are generated with different random seeds and dataset combinations, then sensitivity to initial conditions is reduced and result reliability increases, but computational complexity and processing time increase
Solution Approach 1:
The patent performs a predetermined number of iterations (e.g., 10-100 bootstrap samples) to achieve sufficient statistical reliability without excessive computation. This partial action approach balances the need for reliable results with computational constraints, avoiding the need to exhaustively test all possible model configurations.
Solution Approach 2:
The patent merges the predictions and feature importance results from multiple machine learning models through aggregation. By combining results from models trained on different bootstrap samples with different random seeds, the system achieves more reliable and stable analytics results while distributing the computational workload across parallel model training iterations.
Data Source
AI summary
Systems, methods, and apparatuses for determining feature importance of analytics data in predicting a response value include receiving data records, each data record including a response value and values of features associated with the response value; splitting the data records into datasets, each dataset including a part of the data records; generating a machine learning model using each of the datasets, the machine learning model outputting one or more predicting features having influence in predicting the response value for each of the datasets; determining an important feature based on the one or more predicting features; and generating report data indicating that a business metric associated with the important feature has the highest predicted influence among the features on predicting the response value.