Machine Learning Model Change Quantification With Coefficient Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle with determining the appropriate number of model input dimensions and normalizing data with long-tailed distributions in machine-generated data, leading to overfitting and ineffective model construction for complex software systems with heterogeneous workloads.
Innovation Solution
The method selects an optimal number of model input dimensions using a logarithmic kernel function for normalization and constructs multiple candidate models, evaluating them based on error measures to identify the best model, while quantifying changes over time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of model input dimensions is increased to capture more product usage data, then the model's predictive accuracy is improved, but the risk of overfitting increases and the model becomes more complex
Solution Approach 1:
The patent transforms the raw product usage data through logarithmic normalization, changing the parameter scale from linear to logarithmic. This transformation compresses the range of values and redistributes them, allowing the model to capture patterns across multiple orders of magnitude without requiring excessive dimensions, thus improving predictive accuracy while controlling model complexity
Solution Approach 2:
The patent introduces a temporal dimension by training multiple models at different points in time and measuring the drift between them. This additional temporal dimension allows the system to detect changes in product usage patterns over time without increasing the complexity of individual cross-sectional models, as the drift measurement operates at a higher level of abstraction
2Adaptability or versatility
If raw data is stored and analyzed later to maintain flexibility, then the ability to analyze all generated data is improved, but the challenge of searching and analyzing vast amounts of diverse data increases
Solution Approach 1:
The patent performs preliminary transformation of the raw data by applying logarithmic normalization and selecting key dimensions before model training. This preliminary action structures the data in a way that makes it more amenable to analysis, reducing the difficulty of detecting and measuring patterns while preserving the flexibility to analyze all generated data later
Solution Approach 2:
The patent extracts specific dimensions and features from the vast amount of raw product usage data that are most relevant to predicting product usage. By taking out only the essential features needed for the predictive models, the system maintains analysis flexibility while reducing the complexity of data search and analysis
Data Source
AI summary
One embodiment disclosed herein relates to a computerized method for quantifying a change between two machine learning models. Operations of the computerized method include training a first machine learning (ML) model and a second ML model and determining a first vector of coefficients of the first ML model and a second vector of coefficients of the second ML model. A fractional change is then determined between the first ML model and the second ML model based on a Euclidean distance between the first vector of coefficients and the second vector of coefficients, wherein the Euclidean distance is correlated with a Euclidean magnitude of the first vector of coefficients. A graphical user interface is then generated illustrating the fractional change between the first ML model and the second ML model, wherein the graphical user interface is configured to be rendered on a display screen of a network device.