Machine Learning Prediction Accuracy via Data Binning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face significant predictive accuracy issues due to data imbalance, where a skewed distribution of observations biases the models towards generating false positives or false negatives, particularly in classification and regression tasks, leading to unreliable predictions.
Innovation Solution
The system employs a multi-class data structure by binning observation values into distinct ranges, allowing feature vectors to be classified and assigned to buckets, and then determines regression values through an aggregation of simulation feature vectors, enabling the model to predict regression values and mitigate the effects of data imbalance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning models are trained on imbalanced data distributions, then the model can be trained faster and with simpler data processing, but the predictive accuracy deteriorates due to bias towards false positives or false negatives
Solution Approach 1:
The patent segments the continuous target variable into discrete buckets, transforming the regression problem into a classification problem. This segmentation allows the model to handle imbalanced data more effectively by treating each bucket as a distinct class, thereby improving predictive accuracy without requiring complex resampling techniques that would slow down training.
Solution Approach 2:
The patent changes the parameter representation by converting continuous observation values into discrete bucket assignments. This parameter transformation enables the use of classification algorithms that are more robust to data imbalance, improving predictive accuracy while maintaining training efficiency through standard classification loss functions.
2Device complexity
If machine learning models attempt to handle regression tasks directly with imbalanced data, then the model structure remains simple, but the reliability of predictions deteriorates due to skewed distribution bias
Solution Approach 1:
By segmenting the continuous target space into discrete buckets, the patent transforms the regression task into a classification task. This approach improves prediction reliability by enabling the model to learn distinct decision boundaries for each bucket, reducing the impact of skewed data distributions while maintaining a relatively simple model architecture.
Solution Approach 2:
Instead of trying to solve the regression problem directly with imbalanced data, the patent inverts the approach by first discretizing the target variable and then applying classification techniques. This inversion allows the use of well-established classification algorithms that are more robust to data imbalance, thereby improving prediction reliability.
3Reliability
If manual hyperparameter tuning and alternative model execution are used to improve predictive accuracy, then the prediction reliability improves, but the hardware resource usage and time consumption increase
Solution Approach 1:
The patent enables the model to self-adjust by learning bucket assignments and decision boundaries directly from the imbalanced data through classification training. This self-service approach eliminates the need for manual hyperparameter tuning and multiple alternative model executions, improving prediction reliability while reducing hardware resource consumption and training time.
Data Source
AI summary
Training a machine learning model includes electronically retrieving feature vectors that comprise a electronic representations of multidimensional observations, each observation uniquely associated with a predetermined observation value. A multi-class data structure comprising a plurality of buckets is generated by binning the observation values associated with the multidimensional observations. Each bucket corresponds to a range of values and contains observations whose associated observation values lie within the range. A machine learning model is trained using the feature vectors to classify feature vector inputs, assigning each feature vector input to a bucket. Simulated execution of the machine learning model classifies simulation feature vectors by assigning each simulation feature vector to a bucket based on the feature. For each bucket, a regression value is determined based on an aggregation of simulation feature vectors assigned to the bucket, thereby enabling the machine learning model to predict regression values corresponding to subsequent feature vector inputs.